Csaba Ringhofer has posted comments on this change. ( http://gerrit.cloudera.org:8080/17869 )
Change subject: IMPALA-10956: datasketches UDFs: memory leak and merge overhead ...................................................................... Patch Set 3: (2 comments) lgtm in general, I will to go through the code again to check all possible leaks http://gerrit.cloudera.org:8080/#/c/17869/3//COMMIT_MSG Commit Message: http://gerrit.cloudera.org:8080/#/c/17869/3//COMMIT_MSG@9 PS3, Line 9: - call destructors of sketch and union objects Thanks a lot for fixing this! A few notes here: - We should have some utilities in Impala that help with memory management of complex types in aggregate states. I am thinking about creating an std::allocator like class that would use the same allocation as FunctionContext through some thread local state. This seems a valid thing to do as it is guaranteed that the aggregate state will be used only within a single thread. datasketch classes already can get an allocator template, so this would allow us to allocate all memory the "proper" way. - We don't test this area too well - there are no GROUP BYs in https://github.com/apache/impala/blob/master/testdata/workloads/functional-query/queries/QueryTest/datasketches-hll.test and without GROUP BY, we'll only create a single sketch per fragment instance during aggregation, so we won't leak too much memory. Adding a test with a GROUP BY with e.g. 1000 groups would increase the chance of catching such issues. http://gerrit.cloudera.org:8080/#/c/17869/3/be/src/exprs/aggregate-functions-ir.cc File be/src/exprs/aggregate-functions-ir.cc: http://gerrit.cloudera.org:8080/#/c/17869/3/be/src/exprs/aggregate-functions-ir.cc@1689 PS3, Line 1689: agg_state_ptr->second = new (ctx->Allocate<datasketches::hll_sketch>()) : datasketches::hll_sketch(DS_SKETCH_CONFIG, DS_HLL_TYPE); DsHllInit always initializes a hll_sketch while we'll only modify it if DsHllUpdate is called later. My idea would be to skip creating a hll_sketch in DsHllInit (e.g. by creating an uninitialized state) and then initialize a hll_sketch/hll_union during Update/Merge. -- To view, visit http://gerrit.cloudera.org:8080/17869 To unsubscribe, visit http://gerrit.cloudera.org:8080/settings Gerrit-Project: Impala-ASF Gerrit-Branch: master Gerrit-MessageType: comment Gerrit-Change-Id: I8dd0e6736f4266f74f5f265f58d40a4e4707287f Gerrit-Change-Number: 17869 Gerrit-PatchSet: 3 Gerrit-Owner: Alexander Saydakov <al...@apache.org> Gerrit-Reviewer: Csaba Ringhofer <csringho...@cloudera.com> Gerrit-Reviewer: Fucun Chu <chufu...@hotmail.com> Gerrit-Reviewer: Impala Public Jenkins <impala-public-jenk...@cloudera.com> Gerrit-Comment-Date: Fri, 29 Oct 2021 09:57:23 +0000 Gerrit-HasComments: Yes
