[Bug] Parallel dedup semi-join planning takes an exclusive lock on gp_segment_configuration
Maintainer antworten meist innerhalb von 2 Tagen
Dieses Issue hat noch niemand übernommen.
Bewertung
- Schwierigkeit
- 4/5
- Geschätzter Aufwand
- 3-5 Tage
- Anfängerfreundlichkeit
- 53/100
Rechercherichtung
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.
Vom Indexierungsmodell aus dem Issue-Text verfasst.
Beschreibung
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
- I agree to follow this project's Code of Conduct.
- Vorherrschende Sprache
- C
- Sterne
- 1.4k
- Forks
- 273
- Ø Merge
- 4 T. 6 Std.
- Gemergte PRs (30 T.)
- 105
Entwicklungsumgebung
- Kein Dockerfile und keine Docker-Compose-Datei
- Hat eine Pull-Request-Vorlage
- Beitragsleitfaden lesen
Erste Schritte
- Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
- Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
- Forken Sie das Repository und arbeiten Sie in einem Branch.
- Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.
Mehr aus apache/cloudberry
-
isolation2 lockmodes fails on cassert builds: extra AccessShareLock on partition rootEvtl. vergeben @tuhaihe hat das heute übernommen. Offen
Schwierigkeit 2/5 1-3 Stunden Anfängerfreundlichkeit 78/100
apache/cloudberry#2356 ·
Maintainer antworten meist innerhalb von 2 Tagen
-
psql tab completion broken for storage servers, tags, profiles and resource groups (pre-PG15 query format)Evtl. vergeben @allenvox hat das vor 5 Tagen übernommen. Offen
Schwierigkeit 2/5 1-3 Stunden Anfängerfreundlichkeit 72/100
apache/cloudberry#2316 ·
Maintainer antworten meist innerhalb von 2 Tagen
-
[Bug] Expand fails to recognize user-defined tablespacesEvtl. vergeben @nix-oss hat das vor 92 Tagen übernommen. Offentype: Bug
Schwierigkeit 2/5 1-3 Stunden Anfängerfreundlichkeit 76/100
apache/cloudberry#1885 · 2 Reaktionen ·
Maintainer antworten meist innerhalb von 2 Tagen
-
gpexpand first stage fails: 'SyncPackages' object has no attribute 'ret' (extension sync never runs; regression)Evtl. vergeben @talmacschen-arch hat das vor 153 Tagen übernommen. Offen
Schwierigkeit 2/5 1-3 Stunden Anfängerfreundlichkeit 86/100
apache/cloudberry#1825 ·
Maintainer antworten meist innerhalb von 2 Tagen
-
Schwierigkeit 5/5 Über eine Woche Anfängerfreundlichkeit 15/100
apache/cloudberry#2348 · 1 Kommentar ·
Maintainer antworten meist innerhalb von 2 Tagen
Alle Issues in apache/cloudberry
Ähnliche Issues
-
Schwierigkeit 2/5 1-3 Stunden Anfängerfreundlichkeit 78/100
Maintainer antworten meist innerhalb von 1 Tag
-
windows-artifacts: the win-arm64 job uses a Visual Studio generator missing from windows-11-armOffen
Schwierigkeit 1/5 Unter einer Stunde Anfängerfreundlichkeit 86/100
-
Schwierigkeit 1/5 1-3 Stunden Anfängerfreundlichkeit 78/100
crosspoint-reader/crosspoint-reader#4535 ·
Maintainer antworten meist innerhalb von 1 Tag
-
[core] Clients never advertise RNS_UD_CS_SUPPORT_ERRINFO_PDU, so a conforming server can't send them Set Error InfoEvtl. vergeben @akallabeth hat das heute übernommen. Offen
Schwierigkeit 2/5 1-3 Stunden Anfängerfreundlichkeit 75/100
FreeRDP/FreeRDP#14230 · 1 Kommentar ·
Maintainer antworten meist innerhalb von 1 Tag
-
Wrong release linkedOffen
Schwierigkeit 1/5 Unter einer Stunde Anfängerfreundlichkeit 88/100
Genymobile/scrcpy#7178 ·
Maintainer antworten meist innerhalb von 3 Tagen