Hi Maryann,
I created https://issues.apache.org/jira/browse/PHOENIX-3354 for this issue. I
could not assign to you.
Best regards,Sumit
From: Maryann Xue <[email protected]>
To: "[email protected]" <[email protected]>; Sumit Nigam
<[email protected]>
Sent: Wednesday, October 5, 2016 11:27 AM
Subject: Re: Hash join confusion
Not sure if it's related, coz your DDL does not have DESC columns, but we do
have a sort-merge-join bug fix in 4.8.0:
https://issues.apache.org/jira/browse/PHOENIX-2894.
Otherwise could you please just file a JIRA and assign to me? Thanks a lot!
Thanks,Maryann
On Tue, Oct 4, 2016 at 8:24 PM, Sumit Nigam <[email protected]> wrote:
Hi Maryann,
Here are the 2 DDLs (for data and index tables):
CREATE TABLE IF NOT EXISTS "ldmns:exDocStoreb" (CURRENT_TIMESTAMP BIGINT NOT
NULL, ID VARCHAR(96), BINARY_CURR_EXDOC VARBINARY, CURR_CHECKSUM VARCHAR(32),
BINARY_PREV_EXDOC VARBINARY, PREV_CHECKSUM VARCHAR(32), PREV_TIMESTAMP BIGINT,
SUMMARY VARCHAR, OBJ_SUMMARY VARCHAR, PARAM_SAMPLES VARCHAR, BULK_PUBLISH_UUID
VARCHAR, TOTAL_FACTS INTEGER, CURR_EXDOC VARCHAR, PREV_EXDOC VARCHAR CONSTRAINT
PK PRIMARY KEY(CURRENT_TIMESTAMP, ID)) COMPRESSION = 'SNAPPY', BLOCKCACHE =
false, SALT_BUCKETS = 36
CREATE INDEX IF NOT EXISTS "ldmns:indx_exdocb" ON "ldmns:exDocStoreb"(ID)
INCLUDE (SUMMARY, OBJ_SUMMARY, PARAM_SAMPLES, BULK_PUBLISH_UUID)
Here is the upsert query for this table:
UPSERT INTO "ldmns:exDocStoreb" (CURRENT_TIMESTAMP, BULK_PUBLISH_UUID, ID,
CURR_CHECKSUM, CURR_EXDOC, SUMMARY, PREV_EXDOC, PREV_CHECKSUM, PREV_TIMESTAMP,
OBJ_SUMMARY, PARAM_SAMPLES, TOTAL_FACTS, BINARY_CURR_EXDOC, BINARY_PREV_EXDOC)
VALUES (TO_NUMBER(CURRENT_TIME()), ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
Here is explain plan of a SELECT with merge sort:
explain select /* +USE_SORT_MERGE_JOIN*/ ID, CURR_EXDOC, BINARY_CURR_EXDOC,
CURRENT_TIMESTAMP, CURR_CHECKSUM, PREV_EXDOC, BINARY_PREV_EXDOC, PREV_CHECKSUM,
PREV_TIMESTAMP from "ldmns:exDocStoreb" as a inner join (select
max(CURRENT_TIMESTAMP) as mct, ID as tid from "ldmns:exDocStoreb" where ID like
' 006389a6b10667f39bdbbdafdc4611 e03cc04418cbc2619ddc01f54d88d7 c3bf%' group by
ID) as tmp on a.ID=tmp.tid and a.CURRENT_TIMESTAMP=tmp.mct where id like '
006389a6b10667f39bdbbdafdc4611 e03cc04418cbc2619ddc01f54d88d7 c3bf%'
;+----------------------------- -------------+| PLAN
|+----------------------------- -------------+| SORT-MERGE-JOIN
(INNER) TABLES || CLIENT 36-CHUNK PARALLEL 36-WAY RANGE SCAN OVER
ldmns:exDocStoreb [0] || SERVER FILTER BY ID LIKE '
006389a6b10667f39bdbbdafdc4611 e03cc04418cbc2619ddc01f54d88d7 c3bf%' ||
SERVER SORTED BY [A.ID, A.CURRENT_TIMESTAMP] || CLIENT MERGE SORT
|| AND (SKIP MERGE) || CLIENT 36-CHUNK
PARALLEL 36-WAY RANGE SCAN OVER ldmns:indx_exdocb [0,'
006389a6b10667f39bdbbdafdc4611 e03cc04418cbc2619ddc01f54d88d7 c3bf'] -
[0,'006389 || SERVER FILTER BY FIRST KEY ONLY || SERVER
AGGREGATE INTO ORDERED DISTINCT ROWS BY ["ID"] || CLIENT MERGE SORT
|| CLIENT SORTED BY ["ID", MAX("CURRENT_TIMESTAMP")]
|+----------------------------- -------------+11 rows selected (0.025 seconds)
Here is explain plan with default join:
explain SELECT ID, CURR_EXDOC, BINARY_CURR_EXDOC, CURRENT_TIMESTAMP,
CURR_CHECKSUM, PREV_EXDOC, BINARY_PREV_EXDOC, PREV_CHECKSUM, PREV_TIMESTAMP
from "ldmns:exDocStoreb" as a inner join (select max(CURRENT_TIMESTAMP) as mct,
ID as tid from "ldmns:exDocStoreb" where ID like '
42ecf4abd4bd7e7606025dc8eee3de 6a3cc04418cbc2619ddc01f54d88d7 c3bf%' group by
ID) as tmp on a.ID=tmp.tid and a.CURRENT_TIMESTAMP=tmp.mct where ID like '
42ecf4abd4bd7e7606025dc8eee3de 6a3cc04418cbc2619ddc01f54d88d7 c3bf%'
;+----------------------------- -------------+| PLAN
|+----------------------------- -------------+| CLIENT 3-CHUNK
PARALLEL 3-WAY RANGE SCAN OVER ldmns:exDocStoreb [0] || SERVER FILTER BY ID
LIKE ' 42ecf4abd4bd7e7606025dc8eee3de 6a3cc04418cbc2619ddc01f54d88d7 c3bf%' ||
PARALLEL INNER-JOIN TABLE 0 (SKIP MERGE) || CLIENT 3-CHUNK PARALLEL
3-WAY RANGE SCAN OVER ldmns:indx_exdocb [0,' 42ecf4abd4bd7e7606025dc8eee3de
6a3cc04418cbc2619ddc01f54d88d7 c3bf'] - [0,' 42ecf4abd4bd7e7606025dc8eee3de
6a3cc0 || SERVER FILTER BY FIRST KEY ONLY || SERVER
AGGREGATE INTO ORDERED DISTINCT ROWS BY ["ID"] || CLIENT MERGE SORT
|| DYNAMIC SERVER FILTER BY (A.CURRENT_TIMESTAMP, A.ID) IN
((TMP.MCT, TMP.TID)) |+----------------------------- -------------+8 rows
selected (0.033 seconds)
Looking forward to hearing from you.
Best regards,Sumit
From: Sumit Nigam <[email protected]>
To: "[email protected]" <[email protected]>
Sent: Wednesday, October 5, 2016 12:13 AM
Subject: Re: Hash join confusion
Thanks Maryann.
I will share the details in a few hours.
Under heavy load scenario, the default hash join failed with time-out (and
memory issue), so I switched to sort-merge. But sort-merge is missing data
randomly. So, as of now I am not sure what is the issue with sort-merge join.
Hash join does not miss any data but has the issue of not fitting in memory
(the actual issue with which I started this thread).
Thanks again!Sumit
From: Maryann Xue <[email protected]>
To: Sumit Nigam <[email protected]>; "[email protected]"
<[email protected]>
Sent: Tuesday, October 4, 2016 10:04 PM
Subject: Re: Hash join confusion
Hi Sumit,
Thank you for the update! Would you mind sharing the queries and their plans,
as well as the DDL for both the data tables and the index?
And just to confirm, you are saying hash joins are working, is it with changes
to the config or without?
Thanks,
Maryann
On Tue, Oct 4, 2016 at 9:17 AM Sumit Nigam <[email protected]> wrote:
Thank you Maryann.
>From the time I have moved to sort-merge join, my use cases have stopped
>working. However, if I remove the hint (and fall back to hash), then they all
>work. I am on phoenix 4.6/ hbase 1.1
I thought just changing the join algorithm would be enough. I would assume that
changing the hash join to sort-merge join would not alter the query results,
right? Do I need to re-write my query?
I am using global index.
Thanks,Sumit
From: Maryann Xue <[email protected]>
To: Sumit Nigam <[email protected]>
Cc: "[email protected]" <[email protected]>
Sent: Sunday, October 2, 2016 5:30 AM
Subject: Re: Hash join confusion
So if either or both sides of a sort-merge-join will have to be sorted simply
depends on whether this side is already ordered on the join key.
So far we don't have any documentation specifically for explain plan yet, but
the Phoenix website does have some examples for different types of queries or
functionalities, including join queries.
Thanks,Maryann
On Wed, Sep 28, 2016 at 10:52 PM, Sumit Nigam <[email protected]> wrote:
Thanks Maryann.
Yes let me switch to merge sort join because the other query uses lots more
columns. Also, if I just change the hint to use merge sort would that be enough
or I need to sort both the driving query and subquery with same order by for
merge sort?
As an aside, is there a document to interpret explain plan?
Thanks,Sumit
From: Maryann Xue <[email protected]>
To: Sumit Nigam <[email protected]>
Cc: "[email protected]" <[email protected]>
Sent: Thursday, September 29, 2016 11:03 AM
Subject: Re: Hash join confusion
Thank you Sumit, for trying this out! So right now it's very clear that the
table to be cached IS too big so there should be no point of using hash join in
this case. Is the other table much smaller, or it is about the same size or
even bigger? If it's considerably smaller you can probably rewrite your query
to do the join the other way, otherwise let's just stick to sort-merge join.
Thanks,Maryann
On Wed, Sep 28, 2016 at 10:29 PM, Sumit Nigam <[email protected]> wrote:
Thank you Maryann.
I am not using multi-tenancy for these tables. Increasing phoenix.
coprocessor.maxServerCacheTime ToLiveMs and the corresponding cache size config
just delayed the error.
I have also started seeing some memory problem -
Caused by: org.apache.phoenix.memory.Insu fficientMemoryException: Requested
memory of 22871932 bytes could not be allocated from remaining memory of
776776654 bytes from global pool of 778469376 bytes after waiting for 10000ms.
at org.apache.phoenix.memory. GlobalMemoryManager. allocateBytes(
GlobalMemoryManager.java:78)
at org.apache.phoenix.memory. GlobalMemoryManager.access$
300(GlobalMemoryManager.java: 30)
at org.apache.phoenix.memory. GlobalMemoryManager$
GlobalMemoryChunk.resize( GlobalMemoryManager.java:139)
What I am having trouble with is, that the total size of csv produced by
sub-query is only ~7Mb. I have 12 region servers with 5GB heap each. So, when
this result gets sent across to all region servers to perform the server side
join, not sure why a memory issue should show up (or a time out occur). Any
insights?
These tables are salted. Not sure if it is https://issues.apache.org/
jira/browse/PHOENIX-2900 issue.
Switching to sort merge join helped. But not sure if that is the right solution
going forward.
Thanks again!Sumit
From: Maryann Xue <[email protected]>
To: "[email protected]" <[email protected]>; Sumit Nigam
<[email protected]>
Sent: Wednesday, September 28, 2016 11:36 PM
Subject: Re: Hash join confusion
Yes, Sumit, the sub-query will get cached in hash join. Are you using
multi-tenancy for these tables? If yes, you might want to checkout Phoenix 4.7
or 4.8, since a related bug fix got in the 4.7 release. https://issues.
apache.org/jira/browse/ PHOENIX-2381?jql=project%20%
3D%20PHOENIX%20AND%20text%20~% 20%22hash%20cache%20id%22
Otherwise I think it's the hash cache timeout issue, in which case changing
phoenix.coprocessor.m axServerCacheTimeToLiveMs migh t be helpful.
Thanks,Maryann
On Tue, Sep 27, 2016 at 10:02 PM, Sumit Nigam <[email protected]> wrote:
Hi,
Is there any document which can help me understand explain plan output in
detail? Or, which piece of code should I look at, to get an idea?
Here is explain plan for inner join query below. Can anyone help in explaining
it to me? Like, as per the plan which table is being cached, etc.?Here,
indx_exdocb is index table (on ID) and exDocStoreb is main table with rowkey as
(current_timestamp, ID).
+----------------------------- -------------+| PLAN
|+----------------------------- -------------+| CLIENT 36-CHUNK
PARALLEL 36-WAY FULL SCAN OVER exDocStoreb || PARALLEL INNER-JOIN TABLE 0
(SKIP MERGE) || CLIENT 36-CHUNK PARALLEL 36-WAY RANGE SCAN OVER
indx_exdocb [0,' 42ecf4abd4bd7e7606025dc8eee3de 6a3cc04418cbc2619ddc01f54d88d7
c3bf'] - [0,' 42ecf4abd4bd7e7606025dc8eee3de 6a3cc04418cbc2619ddc01f54d88d7
c3bg' || SERVER FILTER BY FIRST KEY ONLY || SERVER
AGGREGATE INTO ORDERED DISTINCT ROWS BY ["ID"] || CLIENT MERGE SORT
|| DYNAMIC SERVER FILTER BY (A.CURRENT_TIMESTAMP, A.ID) IN
((TMP.MCT, TMP.TID)) |+----------------------------- -------------+
Also, is there a way to turn ON more verbose explain plan? Like, seeing number
of bytes, rows that each step results in?
Thanks,Sumit
From: Sumit Nigam <[email protected]>
To: Users Mail List Phoenix <[email protected]>
Sent: Tuesday, September 27, 2016 9:17 PM
Subject: Hash join confusion
Hi,
I am using hbase 1.1 with phoenix 4.6.
I have a table with row key as (current_timestamp, id) which is salted and
index on (id). This table has ~3 million records.
I have a query like given below.
SELECT ID, CURRENT_TIMESTAMP, <SOME OTHER COLUMNS> from TBL
as a inner join ( select
max(CURRENT_TIMESTAMP) as mct, ID as tid from TBL where ID like '
42ecf4abd4bd7e7606025dc8eee3de 6a3cc04418cbc2619ddc01f54d88d7 c3bf%' group by
ID) as tmp on a.ID=tmp.tid and
a.CURRENT_TIMESTAMP=tmp.mct
The query hangs for long and finally fails with a timeout. I have 12 region
servers each with 5GB heap and also the total records satisfying the above
query is 62K whose CSV dump is ~10MB only.
DoNotRetryIOException: Could not find hash cache for join Id: Ӧ�8�D�. The cache
might have expired and have been removed
and -
Caused by: java.sql.SQLException: Encountered exception in sub plan [0]
execution.
at org.apache.phoenix.execute. HashJoinPlan.iterator(
HashJoinPlan.java:175)
at com.infa.products.ldm. ingestion.server.java.hadoop. impl.
FixPhoenixIngestInputFormat. getQueryPlan( FixPhoenixIngestInputFormat. java:94)
... 22 more
and
Caused by: java.sql.SQLException:
java.util.concurrent.Timeo utException at org.apache.phoenix.cache.
ServerCacheClient. addServerCache( ServerCacheClient.java:264)
I can try playing around with parameters such as phoenix.coprocessor.
maxServerCacheTimeToLiveMs and switching to sort_merge_join actually helped.
But my question is as per Joins | Apache Phoenix in a case such as lhs INNER
JOIN rhs, it is rhs which will be built as hash table in server cache. So, in
the above query I assume this gets cached?
select max(CURRENT_TIMESTAMP) as mct, ID as tid from TBL where ID like '
42ecf4abd4bd7e7606025dc8eee3de 6a3cc04418cbc2619ddc01f54d88d7 c3bf%' group by
ID) as tmp
Thanks,
Sumit
