[Bug] Parallel dedup semi-join planning takes an exclusive lock on gp_segment_configuration

Open
#2,350 0 comments 1 reaction 0 assignees View on GitHub

Maintainers usually reply within 2 days

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
53/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
c, postgresql
Domain
backend, databases

Research direction

Start with the parallel semi-join regression fixture in src/test/regress/sql/cbdb_parallel.sql and trace the two planner paths in src/backend/cdb/cdbpath.c to cdbcomponent_get_maxdbid() in src/backend/cdb/cdbutil.c. Check how the catalog is opened and closed, and identify a read-only lookup that preserves locking for topology-changing callers. Verify with EXPLAIN and the fixture that planning no longer holds an AccessExclusiveLock while topology changes retain their required locking.

Written by the indexing model from the issue text.

Description

type: Bug
Apache Cloudberry version

Apache Cloudberry 2.2.0-incubating, build 13d3c58eaa0 (PostgreSQL 14.10)

What happened

While diagnosing intermittent connection delays on a multi-segment cluster, we observed ordinary read-only SELECTs holding a granted AccessExclusiveLock on pg_catalog.gp_segment_configuration. New connections were waiting for AccessShareLock on the same shared catalog. pg_blocking_pids() identified the SELECT backend as the blocker, including in a different database on the same coordinator.

On a separate three-segment development cluster running the same build, EXPLAIN without ANALYZE alone acquired the lock. In the on/off comparison below, the final plan was Parallel Hash Semi Join in both cases. The lock can therefore come from evaluating a candidate that is never executed.

What you think should happen instead

Planning a read-only query should not hold a transaction-duration AccessExclusiveLock on the shared topology catalog. The planner only needs to read the maximum segment DBID. Please preserve the locking semantics required by topology-changing callers.

How to reproduce

On the affected build in a development environment, use the parallel semi-join regression fixture as a starting point. We compared EXPLAIN (without ANALYZE) on an equivalent EXISTS query joining two distributed tables, with optimizer=off, enable_parallel=on, and two workers per gather.

Planner settings Current backend’s lock on gp_segment_configuration after EXPLAIN
Both parallel dedup semi-join options enabled Granted AccessExclusiveLock
enable_parallel_dedup_semi_join=off and enable_parallel_dedup_semi_reverse_join=off No lock observed

We inspected pg_locks in each read-only test transaction and rolled back immediately after taking the sample.

Operating System

Enterprise Linux 9, x86_64

Anything else

The two parallel join paths in cdbpath.c call cdbcomponent_get_maxdbid() when the deduplicated input uses multiple parallel workers. The catalog implementation opens the relation with AccessExclusiveLock and closes it with NoLock, keeping the lock until transaction end.

A possible fix is to give the planner a separate read-only lookup with appropriately scoped locking. Turning off both dedup options at session scope avoided this path on the tested build while leaving ordinary parallel semi-joins available. We verified the behavior on 2.2.0 only.

Are you willing to submit PR?
  • Yes, I am willing to submit a PR!
Code of Conduct
Dominant language
C
Stars
1.4k
Forks
273
Avg merge
4d 6h
Merged PRs (30d)
105

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from apache/cloudberry

All issues in apache/cloudberry

Similar issues

More C issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.