[
https://issues.apache.org/jira/browse/IMPALA-12800?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17852918#comment-17852918
]
Michael Smith edited comment on IMPALA-12800 at 6/6/24 7:24 PM:
----------------------------------------------------------------
Difference from null slots cache:
{code}
# With caching
Query Compilation: 4s678ms
- Metadata of all 1 tables cached: 26.276ms (26.276ms)
- Analysis finished: 3s466ms (3s440ms)
- Authorization finished (noop): 3s467ms (130.395us)
- Value transfer graph computed: 3s486ms (19.860ms)
- Single node plan created: 4s402ms (915.149ms)
- Runtime filters computed: 4s453ms (51.628ms)
- Distributed plan created: 4s486ms (33.064ms)
- Planning finished: 4s678ms (191.281ms)
# Without caching via 'set use_null_slots_cache=false'
Query Compilation: 14s845ms
- Metadata of all 1 tables cached: 7.608ms (7.608ms)
- Analysis finished: 3s207ms (3s199ms)
- Authorization finished (noop): 3s207ms (120.606us)
- Value transfer graph computed: 3s221ms (14.231ms)
- Single node plan created: 14s610ms (11s389ms)
- Runtime filters computed: 14s661ms (51.286ms)
- Distributed plan created: 14s662ms (246.301us)
- Planning finished: 14s845ms (183.164ms)
{code}
So speeds up single node planning, adds some overhead to distributed planning.
I'll look into disabling it for distributed planning.
Update: the time to produce cache logging was actually being lumped into
"Distributed plan created", so that extra 30ms is from debug logging in
logCacheStats.
was (Author: JIRAUSER288956):
Difference from null slots cache:
{code}
# With caching
Query Compilation: 4s678ms
- Metadata of all 1 tables cached: 26.276ms (26.276ms)
- Analysis finished: 3s466ms (3s440ms)
- Authorization finished (noop): 3s467ms (130.395us)
- Value transfer graph computed: 3s486ms (19.860ms)
- Single node plan created: 4s402ms (915.149ms)
- Runtime filters computed: 4s453ms (51.628ms)
- Distributed plan created: 4s486ms (33.064ms)
- Planning finished: 4s678ms (191.281ms)
# Without caching via 'set use_null_slots_cache=false'
Query Compilation: 14s845ms
- Metadata of all 1 tables cached: 7.608ms (7.608ms)
- Analysis finished: 3s207ms (3s199ms)
- Authorization finished (noop): 3s207ms (120.606us)
- Value transfer graph computed: 3s221ms (14.231ms)
- Single node plan created: 14s610ms (11s389ms)
- Runtime filters computed: 14s661ms (51.286ms)
- Distributed plan created: 14s662ms (246.301us)
- Planning finished: 14s845ms (183.164ms)
{code}
So speeds up single node planning, adds some overhead to distributed planning.
I'll look into disabling it for distributed planning.
Update: the time to produce cache logging was actually being lumped into
"Distributed plan created", so that extra 20s is from debug logging in
logCacheStats.
> Queries with many nested inline views see performance issues with
> ExprSubstitutionMap
> -------------------------------------------------------------------------------------
>
> Key: IMPALA-12800
> URL: https://issues.apache.org/jira/browse/IMPALA-12800
> Project: IMPALA
> Issue Type: Improvement
> Components: Frontend
> Affects Versions: Impala 4.3.0
> Reporter: Joe McDonnell
> Assignee: Michael Smith
> Priority: Critical
> Attachments: impala12800repro.sql, impala12800schema.sql,
> long_query_jstacks.tar.gz
>
>
> A user running a query with many layers of inline views saw a large amount of
> time spent in analysis.
>
> {noformat}
> - Authorization finished (ranger): 7s518ms (13.134ms)
> - Value transfer graph computed: 7s760ms (241.953ms)
> - Single node plan created: 2m47s (2m39s)
> - Distributed plan created: 2m47s (7.430ms)
> - Lineage info computed: 2m47s (39.017ms)
> - Planning finished: 2m47s (672.518ms){noformat}
> In reproducing it locally, we found that most of the stacks end up in
> ExprSubstitutionMap.
>
> Here are the main stacks seen while running jstack every 3 seconds during a
> 75 second execution:
> Location 1: (ExprSubstitutionMap::compose -> contains -> indexOf -> Expr
> equals) (4 samples)
> {noformat}
> java.lang.Thread.State: RUNNABLE
> at org.apache.impala.analysis.Expr.equals(Expr.java:1008)
> at java.util.ArrayList.indexOf(ArrayList.java:323)
> at java.util.ArrayList.contains(ArrayList.java:306)
> at
> org.apache.impala.analysis.ExprSubstitutionMap.compose(ExprSubstitutionMap.java:120){noformat}
> Location 2: (ExprSubstitutionMap::compose -> verify -> Expr equals) (9
> samples)
> {noformat}
> java.lang.Thread.State: RUNNABLE
> at org.apache.impala.analysis.Expr.equals(Expr.java:1008)
> at
> org.apache.impala.analysis.ExprSubstitutionMap.verify(ExprSubstitutionMap.java:173)
> at
> org.apache.impala.analysis.ExprSubstitutionMap.compose(ExprSubstitutionMap.java:126){noformat}
> Location 3: (ExprSubstitutionMap::combine -> verify -> Expr equals) (5
> samples)
> {noformat}
> java.lang.Thread.State: RUNNABLE
> at org.apache.impala.analysis.Expr.equals(Expr.java:1008)
> at
> org.apache.impala.analysis.ExprSubstitutionMap.verify(ExprSubstitutionMap.java:173)
> at
> org.apache.impala.analysis.ExprSubstitutionMap.combine(ExprSubstitutionMap.java:143){noformat}
> Location 4: (TupleIsNullPredicate.wrapExprs -> Analyzer.isTrueWithNullSlots
> -> FeSupport.EvalPredicate -> Thrift serialization) (4 samples)
> {noformat}
> java.lang.Thread.State: RUNNABLE
> at java.lang.StringCoding.encode(StringCoding.java:364)
> at java.lang.String.getBytes(String.java:941)
> at
> org.apache.thrift.protocol.TBinaryProtocol.writeString(TBinaryProtocol.java:227)
> at
> org.apache.impala.thrift.TClientRequest$TClientRequestStandardScheme.write(TClientRequest.java:532)
> at
> org.apache.impala.thrift.TClientRequest$TClientRequestStandardScheme.write(TClientRequest.java:467)
> at org.apache.impala.thrift.TClientRequest.write(TClientRequest.java:394)
> at
> org.apache.impala.thrift.TQueryCtx$TQueryCtxStandardScheme.write(TQueryCtx.java:3034)
> at
> org.apache.impala.thrift.TQueryCtx$TQueryCtxStandardScheme.write(TQueryCtx.java:2709)
> at org.apache.impala.thrift.TQueryCtx.write(TQueryCtx.java:2400)
> at org.apache.thrift.TSerializer.serialize(TSerializer.java:84)
> at
> org.apache.impala.service.FeSupport.EvalExprWithoutRowBounded(FeSupport.java:206)
> at
> org.apache.impala.service.FeSupport.EvalExprWithoutRow(FeSupport.java:194)
> at org.apache.impala.service.FeSupport.EvalPredicate(FeSupport.java:275)
> at
> org.apache.impala.analysis.Analyzer.isTrueWithNullSlots(Analyzer.java:2888)
> at
> org.apache.impala.analysis.TupleIsNullPredicate.requiresNullWrapping(TupleIsNullPredicate.java:181)
> at
> org.apache.impala.analysis.TupleIsNullPredicate.wrapExpr(TupleIsNullPredicate.java:147)
> at
> org.apache.impala.analysis.TupleIsNullPredicate.wrapExprs(TupleIsNullPredicate.java:136){noformat}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
