Horton

Re: generic plans and "initial" pruning

Ilmar Y <tanswis42@gmail.com>

View thread View on postgr.es
From: Ilmar Yunusov <tanswis42@gmail.com>
To: pgsql-hackers@lists.postgresql.org
Cc: Amit Langote <amitlangote09@gmail.com>
Date: 2026-07-06T08:14:08Z
Lists: pgsql-hackers 
The following review has been posted through the commitfest application:
make installcheck-world:  not tested
Implements feature:       tested, failed
Spec compliant:           not tested
Documentation:            not tested

Hi,

Thanks for clarifying the prepared UPDATE behavior.  My DML check now makes
sense to me: in v13 the reduced locking applies to the SELECT portal path, and
actual UPDATE/DELETE execution is still outside that path.

I looked separately at the current CFBot Windows MinGW failure for v13.

The failure is in test_plan_advice/001_replan_regress, while running the core
graph_table_rls test under pg_plan_advice.  The diff shows:

ERROR:  plancache reference 000002ef2f2d77b0 is not owned by resource owner Portal

I could not reproduce that locally on Linux.  I used the same v13 series on
origin/master at e994f956e4864f424320f5243b9af11e173ad398.  The series applied
cleanly with git am, and git diff --check reported no issues.

This passed:

./configure --prefix="$PWD/pg-install" --without-readline --without-zlib --without-icu --enable-tap-tests
make -s -j$(nproc)
make -s install
make -C src/test/regress check
make -C src/test/modules/test_plan_advice check

I also repeated test_plan_advice with a cassert/debug build:

./configure --prefix="$PWD/pg-install" --without-readline --without-zlib --without-icu --enable-tap-tests --enable-cassert --enable-debug CFLAGS="-O0 -g"
make -s clean
make -s -j$(nproc)
make -s install
make -C src/test/modules/test_plan_advice check

That passed too.

Even though I cannot reproduce the MinGW failure locally, the error looks like
it may point at a real ownership mismatch in the new portal replan path.

For a SQL EXECUTE portal, ExecuteQuery() gets the plan with:

GetCachedPlan(entry->plansource, paramLI, NULL, NULL)

That matches the portal ownership model: PortalDefineQuery() stores the cplan,
and PortalReleaseCachedPlan() later releases the portal reference with:

ReleaseCachedPlan(portal->cplan, NULL)

exec_bind_message() appears to follow the same pattern.

But the invalidation/retry path in PortalLockCachedPlan() does:

ReleaseCachedPlan(portal->cplan, portal->resowner)

and then reacquires the replacement plan with:

GetCachedPlan(portal->plansource,
              portal->portalParams,
              portal->resowner,
              portal->queryEnv)

That looks inconsistent with the way the portal cached-plan reference was
originally acquired. If the plan was acquired with owner = NULL, then releasing
it with portal->resowner seems like it would explain the CFBot error: the
Portal resource owner does not own that plancache reference.

Am I reading that correctly?  If so, should PortalLockCachedPlan() keep the
stored portal->cplan reference as a portal-owned reference in the retry path
too, by releasing and reacquiring it with owner = NULL, while still using
portal->resowner for the locks acquired during
ExecutorPrepAndLock()/AcquireExecutorLocks()?

Regards,
Ilmar Yunusov

Commits

  1. Stamp 19beta1.

  2. Revert "Don't lock partitions pruned by initial pruning"

  3. Ensure first ModifyTable rel initialized if all are pruned

  4. Fix bug in cbc127917 to handle nested Append correctly

  5. Remove unstable test suite added by 525392d57

  6. Don't lock partitions pruned by initial pruning

  7. Fix an oversight in cbc127917 to handle MERGE correctly

  8. Track unpruned relids to avoid processing pruned relations

  9. Perform runtime initial pruning outside ExecInitNode()

  10. Move PartitionPruneInfo out of plan nodes into PlannedStmt

  11. Fix setrefs.c's failure to do expression processing on prune steps.

  12. Remove obsolete executor cleanup code

  13. Revert "Move PartitionPruneInfo out of plan nodes into PlannedStmt"

  14. Move PartitioPruneInfo out of plan nodes into PlannedStmt

  15. Refactor and cleanup runtime partition prune code a little

  16. Remove some unnecessary fields from Plan trees.

  17. Remove more redundant relation locking during executor startup.

  18. Shut down Gather's children before shutting down Gather itself.