Repair performance regression in information_schema.triggers view

Enterprise / PostgreSQL - Tom Lane [] - 24 April 2020 16:02 EDT

Commit 32ff26911 introduced use of rank() into the triggers view to calculate the spec-mandated action_order column. As written, this prevents query constraints on the table-name column from being pushed below the window aggregate step. That's bad for performance of this typical usage pattern, since the view now has to be evaluated for all tables not just the one(s) the user wants to see. It's also the cause of some recent buildfarm failures, in which trying to evaluate the view rows for triggers in process of being dropped resulted in "cache lookup failed for function NNN" errors. Those rows aren't of interest to the test script doing the query, but the filter that would eliminate them is being applied too late. None of this happened before the rank() call was there, so it's a regression compared to v10 and before.

We can improve matters by changing the rank() call so that instead of partitioning by OIDs, it partitions by nspname and relname, casting those to sql_identifier so that they match the respective view output columns exactly. The planner has enough intelligence to know that constraints on partitioning columns are safe to push down, so this eliminates the performance problem and the regression test failure risk. We could make the other partitioning columns match view outputs as well, but it'd be more complicated and the performance benefits are questionable.

Side note: as this stands, the planner will push down constraints on event_object_table and trigger_schema, but not on event_object_schema, because it checks for ressortgroupref matches not expression equivalence. That might be worth improving someday, but it's not necessary to fix the immediate concern.

Back-patch to v11 where the rank() call was added. Ordinarily we'd not change information_schema in released branches, but the test failure has been seen in v12 and presumably could happen in v11 as well, so we need to do this to keep the buildfarm happy. The change is harmless so far as users are concerned. Some might wish to apply it to existing installations if performance of this type of query is of concern, but those who don't are no worse off.

I bumped catversion in HEAD as a pro forma matter (there's no catalog incompatibility that would really require a re-initdb). Obviously that can't be done in the back branches.


baf17ad9df Repair performance regression in information_schema.triggers view.
src/backend/catalog/information_schema.sql | 11 +++++++++--
src/include/catalog/catversion.h | 2 +-
2 files changed, 10 insertions(+), 3 deletions(-)


  • Share