[
https://issues.apache.org/jira/browse/IMPALA-13075?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17846065#comment-17846065
]
Riza Suminto commented on IMPALA-13075:
---------------------------------------
Comparing between [^Mem_Limit_1G_Failed.txt] and [^Batch_size_0_Success.txt], I
see that both has "Per-Host Resurce Estimates" beyond 1GB MEM_LIMIT, and
"Per-Host Resource Estimates" that is barely under MEM_LIMIT for some nodes.
That being said, I also see abnormally high "PeakMemoryUsage" for
HASH_JOIN_NODE (id=8) at
cdp-datahub-prod-worker4.cdp-prod.c8g3-pfxs.cloudera.site:27000:
{code:java}
Fragment F07
- InactiveTotalTime: 0ns (0)
- TotalTime: 0ns (0)
Instance eb4b5cd4eb0fa41e:80d577e000000053
(host=cdp-datahub-prod-worker4.cdp-prod.c8g3-pfxs.cloudera.site:27000)
...
HASH_JOIN_NODE (id=8)
ExecOption: Join Build-Side Prepared Asynchronously
- InactiveTotalTime: 0ns (0)
- LocalTime: 2.7m (164763538054)
- PeakMemoryUsage: 7.6 GiB (8132571136)
- ProbeRows: 65,536 (65536)
- ProbeRowsPartitioned: 0 (0)
- ProbeTime: 2.7m (164759538041)
- RowsReturned: 0 (0)
- RowsReturnedRate: 0 per second (0)
- TotalTime: 2.8m (166909545061) {code}
This high memory usage seems linked with IMPALA-3286 where
HashTableCtx::ExprValuesCache::capacity_ is pushed up by BATCH_SIZE number.
[https://github.com/apache/impala/blob/0d215da8d4e3f93ad3c1cd72aa801fbcb9464fb0/be/src/exec/hash-table.cc#L369-L373]
Note also that the failed query does not have "Probe Side Codegen Enabled" as
in the success query.
> Setting very high BATCH_SIZE can blow up memory usage of fragments
> ------------------------------------------------------------------
>
> Key: IMPALA-13075
> URL: https://issues.apache.org/jira/browse/IMPALA-13075
> Project: IMPALA
> Issue Type: Improvement
> Components: Backend
> Affects Versions: Impala 4.0.0
> Reporter: Ezra Zerihun
> Priority: Major
> Attachments: Batch_size_0_Success.txt, Failed (1).txt,
> Failed_Cognos_pool.txt, Mem_Limit_1G_Failed.txt, Success (1).txt,
> Success_Tableau_Pool.txt
>
>
> In Impala 4.0, setting a very high BATCH_SIZE or near max limit of 65536 can
> cause some fragment's memory usage to spike way past the query's defined
> MEM_LIMIT or pool's Maximum Query Memory Limit with Clamp on. So even though
> MEM_LIMIT is set reasonable, the query can still fail with out of memory and
> a huge amount of memory used on fragment. Reducing BATCH_SIZE to a reasonable
> amount or back to default will allow the query to run without issue and use
> reasonable amount of memory within query's MEM_LIMIT or pool's Maximum Query
> Memory Limit.
>
> 1) set BATCH_SIZE=65536; set MEM_LIMIT=1g;
>
> {code:java}
> Query State: EXCEPTION
> Impala Query State: ERROR
> Query Status: Memory limit exceeded: Error occurred on backend ...:27000
> by fragment ... Memory left in process limit: 145.53 GB Memory left in query
> limit: -6.80 GB Query(...): memory limit exceeded. Limit=1.00 GB
> Reservation=86.44 MB ReservationLimit=819.20 MB OtherMemory=7.71 GB
> Total=7.80 GB Peak=7.84 GB Unclaimed reservations: Reservation=8.50 MB
> OtherMemory=0 Total=8.50 MB Peak=56.44 MB Runtime Filter Bank:
> Reservation=4.00 MB ReservationLimit=4.00 MB OtherMemory=0 Total=4.00 MB
> Peak=4.00 MB Fragment ...: Reservation=1.94 MB OtherMemory=7.59 GB
> Total=7.59 GB Peak=7.63 GB HASH_JOIN_NODE (id=8): Reservation=1.94 MB
> OtherMemory=7.57 GB Total=7.57 GB Peak=7.57 GB Exprs: Total=7.57 GB
> Peak=7.57 GB Hash Join Builder (join_node_id=8): Total=0 Peak=1.95 MB
> ...
> Query Options (set by configuration):
> BATCH_SIZE=65536,MEM_LIMIT=1073741824,CLIENT_IDENTIFIER=Impala Shell
> v4.0.0.7.2.16.0-287 (5ae3917) built on Mon Jan 9 21:23:59 UTC
> 2023,DEFAULT_FILE_FORMAT=PARQUET,...
> ...
> ExecSummary:
> ...
> 09:AGGREGATE 32 32 0.000ns 0.000ns 0
> 4.83M 36.31 MB 212.78 MB STREAMING
> 08:HASH JOIN 32 32 5s149ms 2m44s 0
> 194.95M 7.57 GB 1.94 MB RIGHT OUTER JOIN, PARTITIONED
> |--18:EXCHANGE 32 32 93.750us 1.000ms 10.46K
> 1.55K 1.65 MB 2.56 MB HASH(...
> {code}
>
>
> 2) set BATCH_SIZE=0; set MEM_LIMIT=1g;
>
> {code:java}
> Query State: FINISHED
> Impala Query State: FINISHED
> ...
> Query Options (set by configuration and planner):
> MEM_LIMIT=1073741824,CLIENT_IDENTIFIER=Impala Shell v4.0.0.7.2.16.0-287
> (5ae3917) built on Mon Jan 9 21:23:59 UTC
> 2023,DEFAULT_FILE_FORMAT=PARQUET,...
> ...
> ExecSummary:
> ...
> 09:AGGREGATE 32 32 593.748us 18.999ms 45
> 4.83M 34.06 MB 212.78 MB STREAMING
> 08:HASH JOIN 32 32 10s873ms 5m47s 10.47K
> 194.95M 123.48 MB 1.94 MB RIGHT OUTER JOIN, PARTITIONED
> |--18:EXCHANGE 32 32 0.000ns 0.000ns 10.46K
> 1.55K 344.00 KB 1.69 MB HASH(...
> {code}
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
