2010YOUY01 commented on PR #21775: URL: https://github.com/apache/datafusion/pull/21775#issuecomment-4362636504
Thank you—this is a great observation and optimization. - I suggest adding a few micro-benchmarks to https://github.com/apache/datafusion/blob/main/benchmarks/src/hj.rs, as it’s hard to see the impact in complex workloads like TPC-H. - Also, there is a minor `explain analyze` metric regression, it seems `probe_hit_rate` and `avg_fanout` is not tracked in the fast path (Result from PR branch in `datafusion-cli`, `main` got a different expected value for those two metrics) ``` > explain analyze select * from generate_series(10000000) as t1(v1) right semi join generate_series(10000000) as t2(v1) on t1.v1 = t2.v1; +-------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | plan_type | plan | +-------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | Plan with Metrics | HashJoinExec: mode=Partitioned, join_type=RightSemi, on=[(v1@0, v1@0)], metrics=[output_rows=10.00 M, elapsed_compute=715.88ms, output_bytes=77.0 MB, output_batches=1.23 K, array_map_created_count=0, build_input_batches=1.23 K, build_input_rows=10.00 M, input_batches=1.23 K, input_rows=10.00 M, build_mem_used=330.3 M, build_time=286.08ms, join_time=389.82ms, avg_fanout=N/A (0/0), probe_hit_rate=0% (0/10.00 M)] | ``` - Additionally, to further optimize this in the future, probably we can create a new hash map implementation that avoids inserting duplicates at the building phase, this way both build time and probe time can be optimized, also the solution seem to be simpler since we won't need a new probe method for only finding the first match. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
