Repository navigation
fix: publish a CollectLeft join's dynamic filter on the first build report - #26847
lakshmie6x wants to merge 1 commit into
Conversation
|
run benchmark hj |
|
Hi @lakshmie6x, thanks for the request (#26847 (comment)). Only whitelisted users can trigger benchmarks. Allowed users: 2010YOUY01, Dandandan, Fokko, HippoBaro, Jefffrey, Omega359, Rachelint, Rich-T-kid, adriangb, alamb, asubiotto, avantgardnerio, brunal, buraksenn, cetra3, codephage2020, coderfender, comphead, erenavsarogullari, etseidl, friendlymatthew, gabotechs, geoffreyclaude, grtlr, haohuaijin, jayzhan211, jonathanc-n, kevinjqliu, klion26, kosiew, kumarUjjawal, kunalsinghdadhwal, liamzwbao, mbutrovich, mkleen, mzabaluev, neilconway, rluvaton, sdf-jkl, timsaucer, xudong963, zhuqi-lucas. File an issue against this benchmark runner |
|
run benchmark tpch |
|
Hi @lakshmie6x, thanks for the request (#26847 (comment)). Only whitelisted users can trigger benchmarks. Allowed users: 2010YOUY01, Dandandan, Fokko, HippoBaro, Jefffrey, Omega359, Rachelint, Rich-T-kid, adriangb, alamb, asubiotto, avantgardnerio, brunal, buraksenn, cetra3, codephage2020, coderfender, comphead, erenavsarogullari, etseidl, friendlymatthew, gabotechs, geoffreyclaude, grtlr, haohuaijin, jayzhan211, jonathanc-n, kevinjqliu, klion26, kosiew, kumarUjjawal, kunalsinghdadhwal, liamzwbao, mbutrovich, mkleen, mzabaluev, neilconway, rluvaton, sdf-jkl, timsaucer, xudong963, zhuqi-lucas. File an issue against this benchmark runner |
…eport Every probe partition of a CollectLeft join reports the one shared build, yet each waited for all of them before probing. A scheduler that runs only some of a stage's partitions at a time then deadlocks: the running partitions wait for reports from partitions that never get a slot, and hold every slot. The first report already has the whole build, so it now finalizes the filter. Partitioned joins still wait for every build partition.
3c47013 to
c668e2a
Compare
Every probe partition of a CollectLeft hash join reports the same shared build to
SharedBuildAccumulator, yet the filter is only finalized after all of them have reported, so each partition waits for every other one before probing. That build is already complete at the first report. A scheduler that runs only some of a plan's partitions at a time deadlocks on this wait: the running partitions wait for reports from partitions that never get to run.The first CollectLeft report now finalizes the filter. No probe runs with a partial filter, unlike #21641, because the first report already has the whole build. Partitioned joins still wait for every build partition. The new test runs one probe partition of two; it hangs without the change.