[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17265616#comment-17265616
]
Joel Bernstein edited comment on SOLR-15036 at 1/15/21, 1:22 AM:
-
The cardinality in this case is the combination of all the fields. But having
one huge field would also cause performance problems.
When you get into the high cardinality use case almost all the time is going be
spent looking up the BytesRefs. You'll have to do that with either JSON facet
or export. JSON facet still needs to lookup the ByteRef's after the aggregation
to send them out. So it's going to be slow in either case. But it's worth
trying JSON facet in streaming mode to see how it compares to drill.
was (Author: joel.bernstein):
The cardinality in this case is the combination of all the fields. But having
one huge field would also cause performance problems.
When you get into the high cardinality use case almost all the time is going be
spent looking up the BytesRefs. You'll have to do that with either JSON facet
or export. JSON facet still needs to lookup the ByteRef's after the aggregation
to send them out. So it's going to be slow in either case. But it's worth
trying JSON facet in streaming mode to see how it compares to drill.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Fix For: 8.8, master (9.0)
>
> Attachments: relay-approach.patch
>
> Time Spent: 2h
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg),
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17265147#comment-17265147
]
Joel Bernstein edited comment on SOLR-15036 at 1/14/21, 7:37 PM:
-
It was a multi-dimensional. Here are the two expressions:
The facet expression:
{code}
facet(testex, buckets="stock_s,day_i,num_s,test_s", bucketSorts="index desc",
rows="-1")
{code}
The rollup expression. I used rollup -> search rather than drill because I was
actually testing export performance.
{code}
rollup(search(testex, q="*:*", fl="stock_s,day_i,num_s,test_s", sort="stock_s
desc,day_i desc,num_s desc,test_s desc", qt="/export", queueSize="15"),
over="stock_s,day_i,num_s,test_s", count(*))
{code}
was (Author: joel.bernstein):
It was a multi-dimensional. Here are the two expressions:
The facet expressions:
{code}
facet(testex, buckets="stock_s,day_i,num_s,test_s", bucketSorts="index desc",
rows="-1")
{code}
The rollup expression. I used rollup -> search rather than drill because I was
actually testing export performance.
{code}
rollup(search(testex, q="*:*", fl="stock_s,day_i,num_s,test_s", sort="stock_s
desc,day_i desc,num_s desc,test_s desc", qt="/export", queueSize="15"),
over="stock_s,day_i,num_s,test_s", count(*))
{code}
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Fix For: 8.8, master (9.0)
>
> Attachments: relay-approach.patch
>
> Time Spent: 2h
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg)
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17265151#comment-17265151
]
Joel Bernstein edited comment on SOLR-15036 at 1/14/21, 7:26 PM:
-
The facet expression actually did not complete with this particular query. Not
only was the rollup able to complete but I was able to run 3 concurrently and
they all completed. So, at this level of cardinality (3 plus million) facet
just is not an option. But for lower cardinality facet is 10x faster then the
rollup.
was (Author: joel.bernstein):
The facet expression actually did not complete. Not only was the rollup able to
complete but I was able to run 3 concurrently and they all completed. So, at
this level of cardinality (3 plus million) facet just is not an option. But for
lower cardinality facet is 10x faster then the rollup.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Fix For: 8.8, master (9.0)
>
> Attachments: relay-approach.patch
>
> Time Spent: 2h
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It needs to be a *weighted avg.* when rolling
> up the avg. from each facet expression in the plist. However, we have the
>
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17265147#comment-17265147
]
Joel Bernstein edited comment on SOLR-15036 at 1/14/21, 7:25 PM:
-
It was a multi-dimensional. Here are the two expressions:
The facet expressions:
{code}
facet(testex, buckets="stock_s,day_i,num_s,test_s", bucketSorts="index desc",
rows="-1")
{code}
The rollup expression. I used rollup -> search rather than drill because I was
actually testing export performance.
{code}
rollup(search(testex, q="*:*", fl="stock_s,day_i,num_s,test_s", sort="stock_s
desc,day_i desc,num_s desc,test_s desc", qt="/export", queueSize="15"),
over="stock_s,day_i,num_s,test_s", count(*))
{code}
was (Author: joel.bernstein):
It was a multi-dimensional. Here are the two expressions:
The facet expressions:
{code}
facet(testex, buckets="stock_s,day_i,num_s,test_s", bucketSorts="index desc",
rows="-1")
{code}
The rollup expression. I used I rollup -> search rather drill because I was
actually testing export performance.
{code}
rollup(search(testex, q="*:*", fl="stock_s,day_i,num_s,test_s", sort="stock_s
desc,day_i desc,num_s desc,test_s desc", qt="/export", queueSize="15"),
over="stock_s,day_i,num_s,test_s", count(*))
{code}
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Fix For: 8.8, master (9.0)
>
> Attachments: relay-approach.patch
>
> Time Spent: 2h
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg),
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17265111#comment-17265111
]
Michael Gibney edited comment on SOLR-15036 at 1/14/21, 6:28 PM:
-
Thanks for following up on this, [~jbernste]. Is this test case
one-dimensional, or multi-dimensional? Would you be able to share examples of
the streaming expressions you're comparing to arrive at this rough threshold?
was (Author: mgibney):
Thanks for following up on this, [~jbernste]. Is this test case
one-dimensional, or multi-dimensional? Would you be able to share examples of
the streaming expressions your comparing to arrive at this rough threshold?
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Fix For: 8.8, master (9.0)
>
> Attachments: relay-approach.patch
>
> Time Spent: 2h
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It needs to be a *weighted avg.* when rolling
> up the avg. from each facet expression in the plist. However, we have the
> count per collection, so this is doable but will require some changes to the
> rollup expression to support weighted average.
> While this plist approach is doable, it’s a pain for use
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17265047#comment-17265047
]
Joel Bernstein edited comment on SOLR-15036 at 1/14/21, 5:28 PM:
-
In some of the earlier comments on this ticket there was discussion about facet
VS drill. I mentioned that drill is a high cardinality tool. The question came
up about what high cardinality actually means. I've been testing SOLR-14608 and
as part of that testing I've been testing drill VS facet performance. What I've
found is that at around 3 million unique buckets drill starts to perform
better. Facet still worked at 3 million unique buckets but eventually it will
break.
was (Author: joel.bernstein):
In some of the earlier comments on this ticket there was discussion about facet
VS drill. I mentioned that drill is a high cardinality tool. The question came
up about what high cardinality actually means. I've been testing SOLR-14608 and
as part of that testing I've been testing drill VS facet performance. What I've
found is that at around 3 million unique buckets drill starts to perform
better. Facet still worked at 3 million but eventually it will break.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Fix For: 8.8, master (9.0)
>
> Attachments: relay-approach.patch
>
> Time Spent: 2h
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg)
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261727#comment-17261727
]
Joel Bernstein edited comment on SOLR-15036 at 1/9/21, 2:34 AM:
A little more about the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the /stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr so you don't have to worry about zk disconnects
etc... And you don't have to expose zk to the client.
was (Author: joel.bernstein):
A little more abouth the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the /stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr so you don't have to worry about zk disconnects
etc... And you don't have to expose zk to the client.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> mi
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261727#comment-17261727
]
Joel Bernstein edited comment on SOLR-15036 at 1/9/21, 2:34 AM:
A little more about the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the /stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless stream source as it holds no Zk state. It is also loosely
coupled with Solr so you don't have to worry about zk disconnects etc... And
you don't have to expose zk to the client.
was (Author: joel.bernstein):
A little more about the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the /stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr so you don't have to worry about zk disconnects
etc... And you don't have to expose zk to the client.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d),
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261727#comment-17261727
]
Joel Bernstein edited comment on SOLR-15036 at 1/9/21, 2:19 AM:
A little more abouth the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the /stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr so you don't have to worry about zk disconnects
etc... And you don't have to expose zk to the client.
was (Author: joel.bernstein):
A little more abouth the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the /stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr you don't have to worry about zk disconnects etc...
And you don't have expose zk to the client.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261727#comment-17261727
]
Joel Bernstein edited comment on SOLR-15036 at 1/9/21, 2:18 AM:
A little more abouth the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr you don't have to worry about zk disconnects etc...
And you don't have expose zk to the client.
was (Author: joel.bernstein):
A little more abouth the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the export handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr you don't have to worry about zk disconnects etc...
And you don't have expose zk to the client.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261727#comment-17261727
]
Joel Bernstein edited comment on SOLR-15036 at 1/9/21, 2:18 AM:
A little more abouth the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the /stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr you don't have to worry about zk disconnects etc...
And you don't have expose zk to the client.
was (Author: joel.bernstein):
A little more abouth the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the stream handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr you don't have to worry about zk disconnects etc...
And you don't have expose zk to the client.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261727#comment-17261727
]
Joel Bernstein edited comment on SOLR-15036 at 1/9/21, 2:17 AM:
A little more abouth the SolrStream. The SolrStream is meant to be the only
external Streaming Expression at this point. You use it to send a string
expression to the export handler and iterate the results. All other Streaming
Expressions are meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr you don't have to worry about zk disconnects etc...
And you don't have expose zk to the client.
was (Author: joel.bernstein):
A little more the SolrStream. The SolrStream is meant to be the only external
Streaming Expression at this point. You use it to send a string expression to
the export handler and iterate the results. All other Streaming Expressions are
meant to be executed by the /stream handler.
One of the reasons for this is that SolrStream is the only stream source that
uses an HttpSolrClient rather then a CloudSolrClient. So SolrStream is the only
truly stateless Streaming Expression as it holds no Zk state. It is also
loosely coupled with Solr you don't have to worry about zk disconnects etc...
And you don't have expose zk to the client.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), co
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261708#comment-17261708
]
Joel Bernstein edited comment on SOLR-15036 at 1/9/21, 1:39 AM:
So, basic auth happens automatically under the covers. Due to how
CloudSolrClient behaves, Streaming Expressions got basic auth for free. I was
skeptical about this at first but there are test cases now that support this
and after code reviewing and lot's of manual testing I feel confident that this
is correct.
SolrStream has explicit support for basic auth params because it's used as an
external Solrj client for sending streaming expressions to the stream handler.
So it needs to initiate the basic auth request. But once the params make it to
the stream handler they are propagated automatically.
was (Author: joel.bernstein):
So, basic auth happens automatically under the covers. This is due to how
CloudSolrClient behaves. Streaming Expressions got basic auth for free. I was
skeptical about this at first but there are test cases now that support this
and after code reviewing and lot's of manual testing I feel confident that this
is correct.
SolrStream has explicit support for basic auth params because it's used as an
external Solrj client for sending streaming expressions to the stream handler.
So it needs to initiate the basic auth request. But once the params make it to
the stream handler they are propagated automatically.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> mi
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261708#comment-17261708
]
Joel Bernstein edited comment on SOLR-15036 at 1/9/21, 1:39 AM:
So, basic auth happens automatically under the covers. Due to how
CloudSolrClient behaves Streaming Expressions got basic auth for free. I was
skeptical about this at first but there are test cases now that support this
and after code reviewing and lot's of manual testing I feel confident that this
is correct.
SolrStream has explicit support for basic auth params because it's used as an
external Solrj client for sending streaming expressions to the stream handler.
So it needs to initiate the basic auth request. But once the params make it to
the stream handler they are propagated automatically.
was (Author: joel.bernstein):
So, basic auth happens automatically under the covers. Due to how
CloudSolrClient behaves, Streaming Expressions got basic auth for free. I was
skeptical about this at first but there are test cases now that support this
and after code reviewing and lot's of manual testing I feel confident that this
is correct.
SolrStream has explicit support for basic auth params because it's used as an
external Solrj client for sending streaming expressions to the stream handler.
So it needs to initiate the basic auth request. But once the params make it to
the stream handler they are propagated automatically.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261568#comment-17261568
]
Joel Bernstein edited comment on SOLR-15036 at 1/8/21, 8:25 PM:
Ok, let's go with an explicit opt-in for 8.8 if you want to get it in quickly.
Then let's make this the default behavior in 9.0.
The reason I like "tiered" is that the plist sets up a middle tier of
aggregator nodes, one per aliased collection. This happens because of how the
facet expression works, which is to setup a CloudSolrClient and push the facet
call to the collection. Calling it "parallel" makes it sound like it's just
threaded on the same node, which would not scale nearly as well. But, since the
parameter is going away anyway I'm not sure it matters that much.
About adding more intelligence to the functions... One approach to do this
would be to build a top-level optimizer that rewrites expressions. That way we
could separate the optimization logic from the functions.
was (Author: joel.bernstein):
Ok, let's go with an explicit opt-in for 8.8 if you want to get it in quickly.
Then let's make this the default behavior in 9.0.
The reason I like "tiered" is that the plist sets up a middle tier of
aggregator nodes, one per aliased collection. This happens because of how the
facet expression works, which is to setup a CloudSolrClient and push the facet
call to the collection. Calling it "parallel" makes it sound like it's just
threaded on the same node, which would not scale as nearly as well. But, since
the parameter is going away anyway I'm not sure it matters that much.
About adding more intelligence to the functions... One approach to do this
would be to build a top-level optimizer that rewrites expressions. That way we
could separate the optimization logic from the functions.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261550#comment-17261550
]
Timothy Potter edited comment on SOLR-15036 at 1/8/21, 7:36 PM:
Thanks for looking [~jbernste]
I don't see any real risk to existing applications with this functionality (I
know ~ famous last words right ;-). The auto-wrapping only happens if the facet
expression only requests metrics that are safe to rollup, namely count, sum,
min, max, or avg. If while attempting to auto-wrap, we encounter a metric that
isn't safe to rollup, the impl. falls back to non-parallel.
In terms of being simpler, I'd counter that the current impl. is not simple for
users with collection alias and that the additions here are not very
complicated. The diff of {{FacetStream}} looks big because of code reformatting
to match our standard as there were many improperly formatted sections in this
file. I'm happy to undo that and just add my changes if that makes a review
easier, I think you'll see the changes are quite simple.
I'm ok with this being an explicit opt-in (vs. opt-out as I have it) your
{{tiered="true"}} idea. However, I prefer the parameter be named {{plist}} or
{{parallel}} instead of {{tiered}} if you're ok with that? The reason it is
opt-out now is because I didn't want users to have to explicitly specify this
when using a collection alias, i.e. opt-in requires changes to existing apps to
make use of this. However, I appreciate your concern about existing apps and
feel it's OK to require a change to enable an optimization, esp. if we keep the
System property to allow enabling this globally vs. in each expression
specifically.
was (Author: thelabdude):
Thanks for looking [~jbernste]
I don't see any real risk to existing applications with this functionality (I
know ~ famous last words right ;-). The auto-wrapping only happens if the facet
expression only requests metrics that are safe to rollup, namely count, sum,
min, max, or avg. If while attempting to auto-wrap, we encounter a metric that
isn't safe to rollup, the impl. falls back to non-parallel.
In terms of being simpler, I'd counter that the current impl. is not simple for
users with collection alias and that the additions here are not very
complicated. The diff of {{FacetStream}} looks big because of code reformatting
to match our standard as there were many improperly formatted sections in this
file. I'm happy to undo that and just add my changes if that makes a review
easier, I think you'll see the changes are quite simple.
I'm ok with this being an explicit opt-in (vs. opt-out as I have it) your
{{tiered="true"}} idea. However, I prefer the parameter be named {{plist}} or
{{parallel} instead of {{tiered}} if you're ok with that? The reason it is
opt-out now is because I didn't want users to have to explicitly specify this
when using a collection alias, i.e. opt-in requires changes to existing apps to
make use of this. However, I appreciate your concern about existing apps and
feel it's OK to require a change to enable an optimization, esp. if we keep the
System property to allow enabling this globally vs. in each expression
specifically.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}}
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261539#comment-17261539
]
Joel Bernstein edited comment on SOLR-15036 at 1/8/21, 7:25 PM:
One thing we could do to remove any risks and move faster with this would be to
make the default behavior be the standard non-plist approach. Then we could
turn it on with a named param or a system param.
{code}
facet(alias, tiered="true")
{code}
The tiered parameter would turn on the tiered aggregations.
was (Author: joel.bernstein):
One thing we could do to remove any risks and move faster with this would be to
make the default behavior be the standard non-plist approach. Then we could
turn it on with a named param or a system param.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It needs to be a *weighted avg.* when rolling
> up the avg. from each facet expression in the plist. However, we have the
> count per collection, so this is doable but will require some changes to the
> rollup expression to support weighted average.
> While this pli
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261539#comment-17261539
]
Joel Bernstein edited comment on SOLR-15036 at 1/8/21, 7:24 PM:
One thing we could do to remove any risks and move faster with this would be to
make the default behavior be the standard non-plist approach. Then we could
turn it on with a named param or a system param.
was (Author: joel.bernstein):
One thing we could do to remove any risks and move faster with this would be
make the default behavior by the standard, non-plist, approach. Then we could
turn it on with named param or a system param.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It needs to be a *weighted avg.* when rolling
> up the avg. from each facet expression in the plist. However, we have the
> count per collection, so this is doable but will require some changes to the
> rollup expression to support weighted average.
> While this plist approach is doable, it’s a pain for users to have to create
> the rollup / sort over plist expression f
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17261532#comment-17261532
]
Joel Bernstein edited comment on SOLR-15036 at 1/8/21, 7:15 PM:
A couple of thoughts based mainly on the description of the PR.
It sounds like a change was made directly to the facet expression so that it
behaves differently. In general with streaming expressions I like to create
wrapper classes, or new implementations, where possible, rather than changing
the behavior of existing functions. I like to do this for several reasons:
1) Lower risk. There are applications in the field that use the facet
expression heavily. Changing the behavior of the facet expression adds more
risk that a bug or behavioral change impacts existing apps.
2) Using composition keeps the class implementations simpler.
3) The design of streaming expressions is really all about composable
functions. So in keeping with that design if possible I usually choose to add a
new function rather than changing behavior of existing functions.
4) Speed of development. When adding new functions you can move fast and break
things, without breaking any existing applications. The amount of review and
testing I would do for a change to the facet expression is 10x the amount of
review and testing I would do for a new function.
That all being said, if there is some important reason to change the facet
expression directly then of course it can be done.
was (Author: joel.bernstein):
A couple of thoughts based mainly on the description of the PR.
It sounds like a change was made directly to the facet expression so that it
behaves differently. In general with streaming expressions I like to create
wrapper classes, or new implementations, where possible, rather than changing
the behavior of existing functions. I like to do this for several reasons:
1) Lower risk. There are applications in the field that use the facet
expression heavily. Changing the behavior of the facet expression adds more
risk that a bug or behavioral changes impacts existing apps.
2) Using composition keeps the class implementations simpler.
3) The design of streaming expressions is really all about composable
functions. So in keeping with that design if possible I usually choose to add a
new function rather than changing behavior of existing functions.
4) Speed of development. When adding new functions you can move fast and break
things, without breaking any existing applications. The amount of review and
testing I would do for a change to the facet expression is 10x the amount of
review and testing I would do for a new function.
That all being said, if there is some important reason to change the facet
expression directly then of course it can be done.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controll
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17249692#comment-17249692
]
Joel Bernstein edited comment on SOLR-15036 at 12/15/20, 4:32 PM:
--
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields. The fl for drill specifies the
exported fields.
Maybe we should change the syntax of drill so that the input() function takes
field names as parameters and drill selects the export fl from this field list.
This is quite clean and ties together all the fields needed for export with the
expression wrapping the input function.
was (Author: joel.bernstein):
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields. The fl for drill specifies the
exported fields.
Maybe we should change the syntax of drill so that the input() function takes
field names as parameters and drill selects the export fl from this field list.
This is quit clean and ties together all the fields needed for export with the
expression wrapping the input function.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 0.5h
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17249692#comment-17249692
]
Joel Bernstein edited comment on SOLR-15036 at 12/15/20, 1:48 PM:
--
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields. The fl for drill specifies the
exported fields.
Maybe we should change the syntax of drill so that the input() function takes
field names as parameters and drill selects the export fl from this field list.
This is quit clean and ties together all the fields needed for export with the
expression wrapping the input function.
was (Author: joel.bernstein):
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields. The fl for drill specifies the
exported fields.
Maybe we should change the syntax of drill so that the input() function takes
field names as parameters and drill selects the export fl from this field list.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17249692#comment-17249692
]
Joel Bernstein edited comment on SOLR-15036 at 12/15/20, 1:46 PM:
--
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields. The fl for drill specifies the
exported fields.
Maybe we should change the syntax of drill so that the input() function takes
field names as parameters and drill selects the export fl from this field list.
was (Author: joel.bernstein):
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields.
Maybe we should change the syntax of drill so that the input() function takes
field names as parameters and drill selects the export fl from this field list.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It needs to be a *weighted avg.* when rolling
> up the avg. from each facet expression in the plist. However, we have the
> count per collection, so this is doable but wi
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17249692#comment-17249692
]
Joel Bernstein edited comment on SOLR-15036 at 12/15/20, 1:33 PM:
--
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields.
Maybe we should change the syntax of drill so that the input() function takes
field names as parameters and drill selects the export fl from this field list.
was (Author: joel.bernstein):
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It needs to be a *weighted avg.* when rolling
> up the avg. from each facet expression in the plist. However, we have the
> count per collection, so this is doable but will require some changes to the
> rollup expression to support weighted average.
> While this plist approach is doable, it’s a pain for users to have to create
> the rollup / sort over plist expression for co
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17249692#comment-17249692
]
Joel Bernstein edited comment on SOLR-15036 at 12/15/20, 1:31 PM:
--
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields.
was (Author: joel.bernstein):
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It needs to be a *weighted avg.* when rolling
> up the avg. from each facet expression in the plist. However, we have the
> count per collection, so this is doable but will require some changes to the
> rollup expression to support weighted average.
> While this plist approach is doable, it’s a pain for users to have to create
> the rollup / sort over plist expression for collection aliases. After all,
> aliases are supposed to hide these types of complexities from client
> applications!
> The point of this ticket is to investigate
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17249692#comment-17249692
]
Joel Bernstein edited comment on SOLR-15036 at 12/15/20, 1:30 PM:
--
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported fields
was (Author: joel.bernstein):
The fl for drill needs to include the *a_d* field. Basically you're rolling up
and aggregating from the exported field.s
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> over="a_i",
> sum(the_sum), avg(the_avg), min(the_min), max(the_max), sum(cnt)
> ),
> a_i, sum(the_sum) as the_sum, avg(the_avg) as the_avg, min(the_min) as
> the_min, max(the_max) as the_max, sum(cnt) as cnt
> )
> {code}
> One thing to point out is that you can’t just avg. the averages back from
> each collection in the rollup. It needs to be a *weighted avg.* when rolling
> up the avg. from each facet expression in the plist. However, we have the
> count per collection, so this is doable but will require some changes to the
> rollup expression to support weighted average.
> While this plist approach is doable, it’s a pain for users to have to create
> the rollup / sort over plist expression for collection aliases. After all,
> aliases are supposed to hide these types of complexities from client
> applications!
> The point of this ticket is to investigate
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246569#comment-17246569
]
Joel Bernstein edited comment on SOLR-15036 at 12/9/20, 2:37 PM:
-
The other important thing about drill is accuracy. With high cardinality,
multi-dimension use cases you can ensure 100% accuracy with facets by using the
limit of -1, which means track everything in memory. Otherwise you're relying
on over-fetching and facet refinements to try and ensure accuracy.
Drill is always 100% accurate no matter the cardinality.
There are data warehousing use cases, where accuracy matters (billing etc...).
So in cases where accuracy matters more than speed, and limits of -1 for facets
don't work, then drill can solve the problem.
was (Author: joel.bernstein):
The other important thing about drill is accuracy. With high cardinality,
multi-dimension use cases you can ensure 100% accuracy with facets by using the
limit of -1, which means track everything in memory. Otherwise you're relying
on over-fetching and facet refinements to try and ensure accuracy.
Drill is always 100% accurate no matter the cardinality.
There are data warehousing use cases, where accuracy matters (billing etc...).
So in cases where accuracy matters more the speed, and limits of -1 for facets
don't work, then drill can solve the problem.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
>
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246561#comment-17246561
]
Joel Bernstein edited comment on SOLR-15036 at 12/9/20, 2:24 PM:
-
I haven't put an exact number on what high cardinality means, but it can get
ugly pretty fast with multiple dimensions. Let's say you have 3 dimensions each
with 100,000 unique values. The number of possible combinations that need to be
tracked is so high it cannot be done in memory.
Drill solves this problem by first sorting on the dimensions and rolling up one
combination at a time. So it will never run out of memory no-matter how many
combinations there are.
Drill is fundamentally a high cardinality, multi-dimension specific tool. It
just so happens that its a very important use case because data warehousing
applications often have these high cardinality, multi-dimension rollups to
perform.
was (Author: joel.bernstein):
I haven't put an exact number on what high cardinality means, but it can get
ugly pretty fast with multiple dimensions. Let's say you have 3 dimensions each
with 100,000 unique values. The number of possible combinations that need to be
tracked is so high it cannot be done in memory.
Drill solves this problem by first sorting on the dimensions and rolling up one
combination at a time. So it will never run out of memory no-matter how many
combinations there are.
Drill is fundamentally a high cardinality, multi-dimension specific tool. It
just so happens the it's a very important use case because data warehousing
applications often have these high cardinality, multi-dimension rollups to
perform.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*",
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246561#comment-17246561
]
Joel Bernstein edited comment on SOLR-15036 at 12/9/20, 2:22 PM:
-
I haven't put an exact number on what high cardinality means, but it can get
ugly pretty fast with multiple dimensions. Let's say you have 3 dimensions each
with 100,000 unique values. The number of possible combinations that need to be
tracked is so high it cannot be done in memory.
Drill solves this problem by first sorting on the dimensions and rolling up one
combination at a time. So it will never run out of memory no-matter how many
combinations there are.
Drill is fundamentally a high cardinality, multi-dimension specific tool. It
just so happens the it's a very important use case because data warehousing
applications often have these high cardinality, multi-dimension rollups to
perform.
was (Author: joel.bernstein):
I haven't put an exact number on what high cardinality means, but it can get
ugly pretty fast with multiple dimensions. Let's say you have 3 dimensions each
with 100,000 unique values. The number of possible combinations that need to be
tracked is so high it cannot be done in memory.
Drill solves this problem by first sorting on the dimensions and rolling up one
combination at a time. So it will never run out of memory no-matter how many
combinations there are.
Drill is fundamentally a high cardinality, multi-dimension specific tool. It
just so happens the it's a very important use case because data warehousing
application often have these high cardinality, multi-dimension rollups to
perform.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", f
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246561#comment-17246561
]
Joel Bernstein edited comment on SOLR-15036 at 12/9/20, 2:21 PM:
-
I haven't put an exact number on what high cardinality means, but it can get
ugly pretty fast with multiple dimensions. Let's say you have 3 dimensions each
with 100,000 unique values. The number of possible combinations that need to be
tracked is so high it cannot be done in memory.
Drill solves this problem by first sorting on the dimensions and rolling up one
combination at a time. So it will never run out of memory no-matter how many
combinations there are.
Drill is fundamentally a high cardinality, multi-dimension specific tool. It
just so happens the it's a very important use case because data warehousing
application often have these high cardinality, multi-dimension rollups to
perform.
was (Author: joel.bernstein):
I haven't put an exact number on what high cardinality means, but it can get
ugly pretty fast with multiple dimensions. Let's say you have 3 dimensions each
with 100,000 unique values. The number of possible combinations that needed to
be tracked is so high it cannot be done in memory.
Drill solves this problem by first sorting on the dimensions and rolling up one
combination at a time. So it will never run out of memory no-matter how many
combinations there are.
Drill is fundamentally a high cardinality, multi-dimension specific tool. It
just so happens the it's a very important use case because data warehousing
application often have these high cardinality, multi-dimension rollups to
perform.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*",
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246561#comment-17246561
]
Joel Bernstein edited comment on SOLR-15036 at 12/9/20, 2:21 PM:
-
I haven't put an exact number on what high cardinality means, but it can get
ugly pretty fast with multiple dimensions. Let's say you have 3 dimensions each
with 100,000 unique values. The number of possible combinations that needed to
be tracked is so high it cannot be done in memory.
Drill solves this problem by first sorting on the dimensions and rolling up one
combination at a time. So it will never run out of memory no-matter how many
combinations there are.
Drill is fundamentally a high cardinality, multi-dimension specific tool. It
just so happens the it's a very important use case because data warehousing
application often have these high cardinality, multi-dimension rollups to
perform.
was (Author: joel.bernstein):
I haven't put an exact number on what high cardinality means, but it can ugly
pretty fast with multiple dimensions. Let's say you have 3 dimensions each with
100,000 unique values. The number of possible combinations that needed to be
tracked is so high it cannot be done in memory.
Drill solves this problem by first sorting on the dimensions and rolling up one
combination at a time. So it will never run out of memory no-matter how many
combinations there are.
Drill is fundamentally a high cardinality, multi-dimension specific tool. It
just so happens the it's a very important use case because data warehousing
application often have these high cardinality, multi-dimension rollups to
perform.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246198#comment-17246198
]
Joel Bernstein edited comment on SOLR-15036 at 12/9/20, 12:22 AM:
--
In the high cardinality use case, faceting will eventually run into performance
and memory problems, so I don't really consider it a great high cardinality
solution. Not because it sends all tuples to aggregator nodes, but because it's
an in-memory aggregation.
I was comparing Streaming Expressions, prior to drill, when I mentioned sending
all tuples to the aggregator nodes.
Streaming Expressions, prior to drill, could use the export handler to send all
sorted tuples to the aggregator node and accomplish high cardinality
aggregation.
So, drill improves on previous implementations of Streaming Expressions by
first aggregating inside the export handler.
Just updated my prior comment to make this more clear.
was (Author: joel.bernstein):
In the high cardinality use case, faceting will eventually run into performance
and memory problems, so I don't really consider it a great high cardinality
solution. Not because it sends all tuples to aggregator nodes, but because it's
an in-memory aggregation.
I was comparing Streaming Expressions, prior to drill, when I mentioned sending
all tuples to the aggregator nodes.
Streaming Expressions, prior to drill, could use the export handler to send all
sorted tuples to the aggregator node and accomplish high cardinality
aggregation.
So, drill improves on previous implementations of Streaming Expressions by
first aggregating inside the export handler.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246157#comment-17246157
]
Joel Bernstein edited comment on SOLR-15036 at 12/9/20, 12:21 AM:
--
I can comment on the drill vs facet question. Facet will always be faster than
drill except in the high cardinality use case. Drill really shines in the high
cardinality use case though. Rather than sending all tuples to the aggregator
node, and using the rollup Stream, drill can first aggregate inside of the
export handler and compress the result significantly before hitting the
network. And drill never runs out of memory, where faceting will eventually run
out of memory.
More work is coming that improves the export handler performance by about 300%.
But even this improvement doesn't allow drill to match the speed of facet on
low cardinality aggregations.
was (Author: joel.bernstein):
I can comment on the drill vs facet question. Facet will always be faster than
drill except in the high cardinality use case. Drill really shines in the high
cardinality use case though. Rather than sending all tuples to the aggregator
node, drill can first aggregate inside of the export handler and compress the
result significantly before hitting the network. And drill never runs out of
memory.
More work is coming that improves the export handler performance by about 300%.
But even this improvement doesn't allow drill to match the speed of facet on
low cardinality aggregations.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_mi
[jira] [Comment Edited] (SOLR-15036) Use plist automatically for executing a facet expression against a collection alias backed by multiple collections
[
https://issues.apache.org/jira/browse/SOLR-15036?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17246157#comment-17246157
]
Joel Bernstein edited comment on SOLR-15036 at 12/8/20, 9:18 PM:
-
I can comment on the drill vs facet question. Facet will always be faster than
drill except in the high cardinality use case. Drill really shines in the high
cardinality use case though. Rather than sending all tuples to the aggregator
node, drill can first aggregate inside of the export handler and compress the
result significantly before hitting the network. And drill never runs out of
memory.
More work is coming that improves the export handler performance by about 300%.
But even this improvement doesn't allow drill to match the speed of facet on
low cardinality aggregations.
was (Author: joel.bernstein):
I can comment on the drill vs facet. Facet will always be faster than drill
except in the high cardinality use case. Drill really shines in the high
cardinality use case though. Rather than sending all tuples to the aggregator
node, drill can first aggregate inside of the export handler and compress the
result significantly before hitting the network. And drill never runs out of
memory.
More work is coming that improves the export handler performance by about 300%.
But even this improvement doesn't allow drill to match the speed of facet on
low cardinality aggregations.
> Use plist automatically for executing a facet expression against a collection
> alias backed by multiple collections
> --
>
> Key: SOLR-15036
> URL: https://issues.apache.org/jira/browse/SOLR-15036
> Project: Solr
> Issue Type: Improvement
> Security Level: Public(Default Security Level. Issues are Public)
> Components: streaming expressions
>Reporter: Timothy Potter
>Assignee: Timothy Potter
>Priority: Major
> Attachments: relay-approach.patch
>
>
> For analytics use cases, streaming expressions make it possible to compute
> basic aggregations (count, min, max, sum, and avg) over massive data sets.
> Moreover, with massive data sets, it is common to use collection aliases over
> many underlying collections, for instance time-partitioned aliases backed by
> a set of collections, each covering a specific time range. In some cases, we
> can end up with many collections (think 50-60) each with 100's of shards.
> Aliases help insulate client applications from complex collection topologies
> on the server side.
> Let's take a basic facet expression that computes some useful aggregation
> metrics:
> {code:java}
> facet(
> some_alias,
> q="*:*",
> fl="a_i",
> sort="a_i asc",
> buckets="a_i",
> bucketSorts="count(*) asc",
> bucketSizeLimit=1,
> sum(a_d), avg(a_d), min(a_d), max(a_d), count(*)
> )
> {code}
> Behind the scenes, the {{FacetStream}} sends a JSON facet request to Solr
> which then expands the alias to a list of collections. For each collection,
> the top-level distributed query controller gathers a candidate set of
> replicas to query and then scatters {{distrib=false}} queries to each replica
> in the list. For instance, if we have 60 collections with 200 shards each,
> then this results in 12,000 shard requests from the query controller node to
> the other nodes in the cluster. The requests are sent in an async manner (see
> {{SearchHandler}} and {{HttpShardHandler}}) In my testing, we’ve seen cases
> where we hit 18,000 replicas and these queries don’t always come back in a
> timely manner. Put simply, this also puts a lot of load on the top-level
> query controller node in terms of open connections and new object creation.
> Instead, we can use {{plist}} to send the JSON facet query to each collection
> in the alias in parallel, which reduces the overhead of each top-level
> distributed query from 12,000 to 200 in my example above. With this approach,
> you’ll then need to sort the tuples back from each collection and do a
> rollup, something like:
> {code:java}
> select(
> rollup(
> sort(
> plist(
> select(facet(coll1,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt),
> select(facet(coll2,q="*:*", fl="a_i", sort="a_i asc", buckets="a_i",
> bucketSorts="count(*) asc", bucketSizeLimit=1, sum(a_d), avg(a_d),
> min(a_d), max(a_d), count(*)),a_i,sum(a_d) as the_sum, avg(a_d) as the_avg,
> min(a_d) as the_min, max(a_d) as the_max, count(*) as cnt)
> ),
> by="a_i asc"
> ),
> ove
