[
https://issues.apache.org/jira/browse/SPARK-27296?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17083098#comment-17083098
]
Patrick Cording edited comment on SPARK-27296 at 4/14/20, 10:48 AM:
--------------------------------------------------------------------
I've been trying this out, and I have a couple of questions.
First, here's a snippet showing what I did.
{code:java}
class LongestRunUdaf extends UserDefinedAggregateFunction {
// ...
}
class LongestRunAggregator extends Aggregator[String, LongestRunBuffer,
Option[Run]] {
// ...
}
// This is to get a reference. It is slow.
val lrUdaf = new LongestRunUdaf
val result = df.select(lrUdaf(col("value")))
// This is many times faster, but `res` is now a `DataFrame`
val longestRunAggregator = udaf(new LongestRunAggregator, Encoders.STRING)
val res = dataset.select(longestRunAggregator(col("value")))
res.show()
// This creates a `Dataset[Option[Run]]` as needed, but now it is as slow as
the old UDAF
res.as(Encoders.kryo[Option[Run]])
// Furthermore, this is still as slow as before
val longestRunAggregator = (new LongestRunAggregator).toColumn
dataset.select(longestRunAggregator).first{code}
I am confused by the fact that the performance improvement only is visible when
using UDAFs, but in order to get there, I need to define an Aggregator, which
is still slow if I use it directly. Is it not possible to also achieve the
improvement for the type-safe cases where Aggregators are used directly?
Also, wrapping an Aggregator where the output is of a non-standard type leaves
you with a DataFrame with binary data. Converting it to a Dataset seems to
negate the performance improvement. Is this intended behaviour? In my example,
it should just be one row that should be deserialized.
Or am I just using this in the wrong way?
My full experiment is here: [https://github.com/patrickcording/udaf-benchmark]
was (Author: cording):
I've been trying this out, and I have a couple of questions.
First, here's a snippet showing what I did.
{code:java}
class LongestRunUdaf extends UserDefinedAggregateFunction {
// ...
}
class LongestRunAggregator extends Aggregator[String, LongestRunBuffer,
Option[Run]] {
// ...
}
// This is to get a reference. It is slow.
val lrUdaf = new LongestRunUdaf
val result = df.select(lrUdaf(col("value")))
// This is many times faster, but `res` is now a `DataFrame`
val longestRunAggregator = udaf(new LongestRunAggregator, Encoders.STRING)
val res = dataset.select(longestRunAggregator(col("value")))
res.show()
// This creates a `Dataset[Option[Run]]` as needed, but now it is as slow as
the old UDAF
res.as(Encoders.kryo[Option[Run]])
// Furthermore, this is still as slow as before
val longestRunAggregator = (new LongestRunAggregator).toColumn
dataset.select(longestRunAggregator).first{code}
I am confused by the fact that the performance improvement only is visible when
using UDAFs, but in order to get there, I need to define an Aggregator, which
is still slow if I use it directly. Is it not possible to also achieve the
improvement for the type-safe cases where Aggregators are used directly?
Also, wrapping the Aggregator where the output is of a non-standard type leaves
you with a DataFrame with binary data. Converting it to a Dataset seems to
negate the performance improvement. Is this intended behaviour? In my example,
it should just be one row that should be deserialized.
Or am I just using this in the wrong way?
My full experiment is here: [https://github.com/patrickcording/udaf-benchmark]
> Efficient User Defined Aggregators
> -----------------------------------
>
> Key: SPARK-27296
> URL: https://issues.apache.org/jira/browse/SPARK-27296
> Project: Spark
> Issue Type: Bug
> Components: Spark Core, SQL, Structured Streaming
> Affects Versions: 2.3.3, 2.4.0, 3.0.0
> Reporter: Erik Erlandson
> Assignee: Erik Erlandson
> Priority: Major
> Labels: performance, usability
> Fix For: 3.0.0
>
>
> Spark's UDAFs appear to be serializing and de-serializing to/from the
> MutableAggregationBuffer for each row. This gist shows a small reproducing
> UDAF and a spark shell session:
> [https://gist.github.com/erikerlandson/3c4d8c6345d1521d89e0d894a423046f]
> The UDAF and its compantion UDT are designed to count the number of times
> that ser/de is invoked for the aggregator. The spark shell session
> demonstrates that it is executing ser/de on every row of the data frame.
> Note, Spark's pre-defined aggregators do not have this problem, as they are
> based on an internal aggregating trait that does the correct thing and only
> calls ser/de at points such as partition boundaries, presenting final
> results, etc.
> This is a major problem for UDAFs, as it means that every UDAF is doing a
> massive amount of unnecessary work per row, including but not limited to Row
> object allocations. For a more realistic UDAF having its own non trivial
> internal structure it is obviously that much worse.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
