Skip to content

Support PostgreSQL 15.20, 16.16 and 17.12 in Citus 13.2 - #8977

Open
Onur Tirtir (onurctirtir) wants to merge 34 commits into
release-13.2from
ihalatci-citus-13-pg-update
Open

Onur Tirtir (onurctirtir) wants to merge 34 commits into
release-13.2from
ihalatci-citus-13-pg-update

Conversation

@onurctirtir

@onurctirtir Onur Tirtir (onurctirtir) commented Nov 16, 2026 •

Copy link
Copy Markdown
Member

DESCRIPTION: Support PostgreSQL 15.20, 16.16 and 17.12 in Citus 13.2

HOLD: stable-image testing is now authorized, but merge and publication of the separately validated q13/q70 planner correction remain blocked. The pushed PR still does not contain that correction.

Current paired stable images

Current head: 5d7f47cd1fc72adf58d7665b7da4d782a2383909. Its only change from 8b35613da1eb885d3c57605dbb288a6ba2e5c2c8 is .github/workflows/build_and_test.yml: -dev-8b259ad -> -v7148bd3. All previous compatibility and fixture corrections remain unchanged.

Paired with merged citusdata/the-process#271, merge 7148bd3cb9f8b40b168b2021a5cb4c1644b05b61. Exact-merge stable publication45002632140 completed successfully, all13jobs. The image owner independently verified all13 GHCR manifests: extbuilder/exttester/failtester for15.20/16.16/17.12, citusupgradetester for15.20/16.16, pgupgradetester15.20-16.16-17.12, and stylechecker0.8.18. The stable image source tree is identical to the tested development source8b259ad; no dependency upgrades or support-category expansion were introduced by promotion.

Automatically triggered exact-head Build & Test45003262705 and packaging45003262762 are in progress. No fresh-green or release-readiness claim. No duplicate dispatch was started.

Scope and changes

  • Update supported PostgreSQL minors to15.20/16.16/17.12, corresponding SQL snapshots and upgrade matrix; preserve historical upgrade Citus versions and release dependencies.
  • Required minor compatibility includes the PG15.20 have_createdb_privilege guard and release13 adaptations of planner normalization/catalog identity handling. Existing focused alias, dropped-column, whole-row, correlated-reference, CTE-side-effect and outer-join regressions remain.
  • Keep the published fixture repairs: actual follower topology/setup and fresh clusters per repetition, bounded metadata-daemon startup, complete statistics sort keys, COPY/ownership cleanup, coordinator registration, remote-only transaction pool cap, denied-output-plugin allowlist setup, two constraint sort keys, native metadata-sync prerequisites and nested EXPLAIN-label filtering. Queries, values, security assertions and repeat counts were not weakened.
  • Regression Pipfiles/locks remain identical to release baseline 37cd2d55548b84cb25950f82a80cd7a7a370ebaf; canonical/devcontainer differences remain. No history rewrite, runtime representative-order change, new dependency repair or merge.

Separately gated q13/q70 correction

The pre-existing query-local planner bitmap collision is tracked by #8838, with the proven correction in #8842. These are non-closing references: this release backport does not close the main issue. No duplicate issue or PR is needed.

The exact three-file #8842 correction is locally applied and validated but uncommitted and unpushed. The suffix-only commit excludes it, and its retained patch hash is unchanged. Both complete original250-query workloads match their same-backend local oracle on PG15.20/16.16/17.12: q13=6/0.5; q70=26/18.3333333333333333. Each side has250markers, each target occurs exactly once and is error-free; targets are not filtered. Only the repository's two unsupported-query error classes are excluded symmetrically. Focused existing regressions passed11/11per major (33total), with strict builds, pinned formatting and expected-output normalization passing. These results validate local unpublished code, not this PR's current pushed head.

Original immutable artifacts remain14558684680(q13) and14038580795(q70); later same-name artifact14213888555 is a different workload. Unchanged release13 on actual PG16.15 and PG16.16 also fails both originals; current pre-repair PG15.20 fails both. Untouched release14/main reproduce the defect, so it is not claimed release13-only. The untouched release13 PG15.20 build conflict is separate compatibility evidence, not a fabricated baseline oracle.

Historical CI and remaining blockers

Latest development-image head8b35613da had PR44794002638 required CI success,179successful/10failed/2skipped; full dispatch44793994712 required CI failure,177successful/12failed/2skipped. Both had30/32 advisory jobs pass; packaging44794002613 passed23/23. These remain historical development-image outcomes, not results for the new stable suffix.

The full dispatch's additional required PG15 background_rebalance_parallel_reference_tables failure is proven representative/debug-order nondeterminism:44dependency rows on each side, identical numeric dependency edges and logical colocation/reference-dependency multisets,12representative-label differences. User explicitly keeps #8789 separate. Its runtime representative-order policy was not ported and expectations were not masked.

Reference-DDL errors in issue5248/issue8243 and schema-worker warning-selection differences remain unclassified. Historical8Lib/Coordinator N-1 failures remain disclosed. The earlier six-hour node-health advisory hang did not recur in the latest development runs, but is not claimed resolved. Fresh randomized green cannot clear the known original q13/q70 fixtures or these independent blockers.

Detailed retained CI/fixture chronology and causal controls: ce607 audit, a020 audit, 8b356 audit, and separate8789 disposition.

Historical focused commands: ./configure PG_CONFIG=/usr/lib/postgresql/<major>/bin/pg_config CFLAGS='-O0 -g -Werror' and make -s -j2 && make -s install passed all3supported majors for the normalization backport; make -s -C src/test/regress check-custom-schedule SCHEDULE=citus13-validation.schedule WORKERCOUNT=2 passed8/8per major; make -s -C src/test/regress check-arbitrary-configs CONFIGS=AllSingleShardTableDefaultConfig parallel=1 seed=41255 passed all3majors. These are prior development-source validation, not a fresh stable-image runtime test. The suffix-only change passed git diff --check; the new CI matrix is still running.

Siblings: #8973(main), #8975(release14), #8974(release12).

ihalatci-msft and others added 16 commits November 13, 2026 12:20
Use PostgreSQL 15.20, 16.16, and 17.12 with the-process bdeab0996eb3670ec7b9e487ccb9b2d7b23f70fa dev images.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
PostgreSQL 15.20 exports have_createdb_privilege, so retain the local fallback only for older PostgreSQL 15 minors.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Retain exact alternative outputs for older supported PostgreSQL minors. New minors reject unsafe row-type defaults before Citus and block schema renames until transaction locks are released. Apply the pinned formatter's spacing to the PG15 helper guard.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Update pyasn1, certifi, Werkzeug and the compatible Tornado 6.5 patch line. Keep the existing mitmproxy commit explicit so resolution cannot follow its now-incompatible main branch. Regenerate targeted lock entries with Pipenv and keep the devcontainer mirror identical.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Decorate injected Vars with their outer joins' nulling relids on PostgreSQL 16 and later so join removal does not invalidate catalog filters.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Keep recursive planning consistent with PostgreSQL join removal without bypassing managed-local routing. Preserve dropped attribute positions and cover right joins, join aliases and materialized CTEs.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Citus 13 renamed the two-argument helper to foreach_declared_ptr; PostgreSQL 17 reserves foreach_ptr for a three-argument typed macro.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Cover LEFT and RIGHT joins plus retained nullable-side Vars against the managed-local fixture with dropped columns.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
PG17 retains inner self-joins; use nested outer self-joins to exercise actual elimination while retaining the referenced-Var case.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Carry outer-join nullability through nested joins before replacing eliminated relation RTEs. Cover the nested distributed LEFT JOIN on the nullable side of a managed-local RIGHT JOIN.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
(cherry picked from commit 005b02a)
Do not replace an original nullable relation merely because a planned subquery no longer appears in the transformed query. Preserve conversion of all matching transformed copies and add a deterministic wrong-result regression.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
(cherry picked from commit 3ccb168)
The sorted merge GUC is not available on release-13.2; retain the exact relational regression without that newer-branch setting.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Retain the query assertion in local_dist_join for regular-user and MX execution. Match shared fixture layout from 3e42d97 and 1419228 without newer-branch GUCs or query-time DDL.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
PostgreSQL 16.16 and 17.12 remove the nullable "distributed" side of
local LEFT JOIN distributed before Citus wraps "local", so the DEBUG
output for that query differs between PG15.20 and PG16.16/17.12. Only
assert the count for this query, as release-12.1 does.
Use the same Python-3.9-compatible test dependency set as the
release-12.1 update:

- Move mitmproxy to the citusdata fork commit bd8874ca, which allows
  the patched Brotli 1.2.1 on the pinned legacy runtime.
- Pin brotli 1.2.1, msgpack 1.0.8, filelock 3.19.1, tornado 6.5.10,
  pyasn1 0.6.5, certifi 2026.10.19 and black 25.11.0.
- Pin pyopenssl 24.3.0. The previous lock combined pyopenssl 24.0.0
  (cryptography<43) with cryptography 44.0.1, so pip could not install
  the exported requirements.
- Pin werkzeug 2.3.8. flask 2.0.3 (capped below 2.1 by the mitmproxy
  fork) imports werkzeug.urls.url_quote, which Werkzeug 3.x removed,
  so mitmdump crashed at startup with the previous werkzeug 3.x pin.

The lock was resolved with Pipenv 2025.0.4 on Python 3.9.25; only the
entries above changed. Both "pipenv requirements" and
"pipenv requirements --dev" install cleanly on Python 3.9.25 and pass
"pip check".
@codecov

codecov Bot commented Nov 16, 2026 •

Copy link
Copy Markdown

Codecov Report

❌ Patch coverage is 98.83721% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 88.76%. Comparing base (37cd2d5) to head (6699634).

Additional details and impacted files
@@               Coverage Diff                @@
##           release-13.2    #8977      +/-   ##
================================================
+ Coverage         88.74%   88.76%   +0.01%     
================================================
  Files               291      291              
  Lines             64969    64794     -175     
  Branches           8195     8176      -19     
================================================
- Hits              57658    57514     -144     
- Misses             4885     4946      +61     
+ Partials           2426     2334      -92     
🚀 New features to boost your workflow:
  • ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

@onurctirtir
Onur Tirtir (onurctirtir) marked this pull request as ready for review November 16, 2026 12:49
function_propagation: PostgreSQL 15.20/16.16/17.12 (CVE-2026-6460) reject
ALTER TABLE table_1_for_circ_dep ADD COLUMN ... because the default of
table_2_for_circ_dep uses table_1_for_circ_dep's row type, so PostgreSQL
errors before Citus detects the cycle. Create the column up front and set
its default with ALTER COLUMN ... SET DEFAULT instead. That creates the
same function/table cycle without changing the row type, so Citus raises
its circular-dependency error on old and new minors alike.

isolation_citus_schema_distribute_undistribute: PostgreSQL 15.20/16.16/17.12
(CVE-2026-6467) lock the schema during ALTER SCHEMA ... RENAME, so the
rename waits for s1 on new minors but not on old ones. Make s1 lock the
schema's pg_namespace row before s2 renames it, so the rename waits for
s1 on every minor.

Remove function_propagation_0.out and
isolation_citus_schema_distribute_undistribute_0.out.
Adding nulling relids to the shard-hiding and Citus-object-hiding filters
moved them above the outer join. As a result, a LEFT JOIN row that matched
only a hidden row was removed instead of being null-extended.

When the meta table is on the nullable side of an outer join, add the
filter to the relation's security quals so it runs while scanning the
relation. Elsewhere keep the old WHERE clause filter.
The query returns true whether or not the planner removes the outer join, and also returns true when the filter is placed above the join, so it does not test what its comment claims.
Keep dependency upgrades on main only. Restore both regression manifests and locks exactly to release-13.2, including existing lock differences, preserve all newer compatibility changes, and use the successfully published baseline-dependency images from 8b259add.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Backport 48be61f with PG15/16/17 callback and permission APIs. Normalize after PostgreSQL planning and recursive replanning, repair dropped-column references, and retain modifying CTE effects and surviving distributed subqueries. Include release-14 local-execution-disabled assertions without newer-major planner hooks.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Retain equivalent unaliased coverage on PG15, whose single-shard deparser rejects explicit column alias lists. Preserve the exact dropped-column collision regression on PG16 and PG17. Verified AllSingleShardTableDefaultConfig seed 41255 on all three supported majors.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Run follower repetitions on fresh replica clusters with their actual setup; fully order pg_stats rows and wait for metadata daemon startup.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
The setup_extension action already downloads the exact PostgreSQL build artifact. Do not fetch unrelated concurrently uploaded regression artifacts before it.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Reuse existing role/table setup dependencies and drop the fixture table after its permission checks. Preserve COPY results and denial assertions.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@ihalatci

Copy link
Copy Markdown
Contributor

Advisory followup at 7a9ef30, unchanged PostgreSQL versions, baseline dependencies and -dev-8b259ad images.

Both preceding exact-head runs completed with required CI passed: full44710853676 and PR44710859112, each149success/40failure/2skip, overall failure. Packaging44710859534 passed. All32 full advisory artifacts were inspected:29 fail only multi_multiuser_copy; one has that failure plus a single_node_enterprise prerequisite notice difference; two instead fail issue_8243/issue_5248 with relation table1 is not distributed.

The COPY fixture lacked its existing role-creation prerequisites and never dropped its table. Six-line correction reuses the established multi_create_table/multi_create_users dependencies and drops the COPY table after all permission assertions. Unchanged PG17 negative control reproduced missing roles and repeat cascades; corrected fixture passed8/8 on each PG15/16/17, and the complete PG17 enterprise schedule passed31/31. Expected successful COPY rows and permission-denial assertions are unchanged.

One-line CI correction removes the redundant outer download of all result artifacts: PR advisory9 failed before tests with ArtifactNotFound for same-run, unexpired, unrelated columnar-result artifacts14616803301/14617636077. setup_extension still downloads the exact build-${PG_MAJOR} artifact. No action/dependency upgrade or test skip.

Fresh full44714669912, PR44714678194, and packaging44714678196 are running. Previous runs were allowed to finish and retained, not canceled or silently rerolled.

The two reference-DDL errors occur in the parallel issue-test group containing issue_7477, which creates/drops the global reference table table1. Cross-test metadata-race attribution is suspected, not yet established; no runtime fix, schedule serialization or relaxed expectation has been included. The separate single-node extra BEGIN/ROLLBACK notice is also not claimed fixed merely by removing an unnecessary COPY prerequisite. q13/q70 remain unfixed and HOLD remains.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@ihalatci

Copy link
Copy Markdown
Contributor

Published two narrowly scoped fixture fixes at e21cf0f338c3cd62a91c3f215993715da1f24ac8:

  • 0f4ffc647: remove the four auxiliary schemas and test role left by non_super_user_object_metadata (11 SQL/expected lines).
  • e21cf0f33: register the coordinator before propagate_foreign_servers, which removes it before returning (7 SQL/expected lines). Unchanged PG15 passed its first execution and failed executions 2–8 because coordinator metadata was absent; worker-owner checks then cascaded.

Both fixtures passed eight consecutive executions on each of PG15.20/16.16/17.12 (48 passes), using gosu circleci src/test/regress/citus_tests/run_test.py <fixture> --repeat 8 --use-whole-schedule-line with each fixture name. gosu circleci make -s -C /workspace/src/test/regress check-multi-1 passed 208/208 on PG17 with both the original ownership fixture and the final fixed fixtures. All permission, ownership, and unmark-security assertions are unchanged.

An earlier expanded local run failed 207/208 at schema_based_sharding's final coordinator removal (tuple concurrently deleted). Subsequent initial controls had local archive-ownership/CRLF comparison defects; their logs and raw errors are retained, not treated as clean baseline attribution. The final canonical original and fixed full runs both passed; no runtime fix, expectation normalization, or test skipping was applied to that unclassified local failure. The unchanged parallel issue_8243 schedule also passed eight repeats on PG16 (70/70 including prerequisites); its earlier CI reference-DDL error remains unclassified.

Exact-head CI is in progress: full dispatch, PR, packaging. Full dispatch includes advisory/flakiness tests. Previous failed runs and artifacts remain visible; their green required CI is not a result for this new head.

HOLD remains: q13/q70 are unfixed and owned by the separate root-cause investigation. Dependency baselines, PostgreSQL versions and -dev-8b259ad images are unchanged; no merge or stable promotion.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@ihalatci

Copy link
Copy Markdown
Contributor

Current head: ce607f48372c142a9f3d04e23a854fa1de9a65b2. The prior e21 full and PR runs were allowed to finish before this feature push; no failing run was cancelled.

All 32 full-run advisory artifacts were inspected: 28 ran split_output_plugin_denied in the wrong, allow-Citus cluster configuration; two had unordered named/unnamed exclusion-constraint rows in pg17; one changed which of two unsupported view dependencies was named in a warning/hint; one issued could not find any shard placements for shardId 102020 during issue_5248's otherwise successful table distribution. That last case is a warning, not the older table1 is not distributed error.

Published minimal corrections:

  • 89eb88ced: canonical 69-dash expected header, fixing the introduced check-style failure.
  • 169928053: one-session pool limit only inside the final remote-only transaction. Controlled delayed worker reads reproduced extra BEGIN/ROLLBACK notices; the same control with the limit retained the exact trace. Original queries, count 53 and permission assertions are unchanged.
  • 1f565103f: route plugin-denial repetitions through an allowlist-disabled custom target and the dedicated schedule's preceding setup. The original denial, shard-count, slot and publication assertions are completely unchanged.
  • ce607f483: order the two exclusion-constraint catalog queries by conname; no expected rows changed.

run_test.py single_node_enterprise --repeat 8 --use-whole-schedule-line and run_test.py split_output_plugin_denied --repeat 8 --use-whole-schedule-line passed on each PG15.20/16.16/17.12 (48 passes). make -s -C /workspace/src/test/regress check-enterprise passed 31/31 on PG16. run_test.py pg17 --repeat 8 --use-whole-schedule-line passed 24/24 including eight pg17 and eight pg17_json executions. Existing PG16 follower routing, pinned-tool Black checks and expected-file normalization also passed.

Prior e21 outcomes: full 147 success/42 failure/2 skipped; PR 144 success/45 failure/2 skipped; required CI red on both; packaging 23/23 passed. The additional PR PG15 query-generator failure was a 20-minute Run Test timeout, not established wrong-result evidence. Artifact 14624885644 retains the generated workload and cluster logs, but no completed comparison outputs; timeout attribution remains open. The two diagnostic-order/metadata-warning cases are also not masked or declared fixed.

New exact-head runs are in progress: full with advisory tests, PR, packaging. HOLD remains for unfixed q13/q70 and unresolved evidence. Dependencies, PG versions and development images remain unchanged; no merge or stable promotion.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@ihalatci

Copy link
Copy Markdown
Contributor

Follow-up at a020482e4db5d46acea0189db14768f5ec5ede45: both preceding ce607 runs finished with required CI SUCCESS, but overall failure (149 successful jobs, 40 failed, two skipped each): eight PG17 N-1 failures and 32 advisory failures. Packaging passed 23/23. No failed run was cancelled or hidden.

All 32 full-run advisory artifacts were retained and inspected. The denial-environment correction now passes all eight actual CI repetitions. The next first failure is start_stop_metadata_sync in 31 artifacts; the remaining artifact has issue_8243 failing create_reference_table('ref2') with relation table1 is not distributed, followed by plan cascades.

The metadata fixture was incorrectly run with generic minimal setup instead of its native MX prerequisites. Negative control on matched PG17 reproduced all eight failures. Removing the coordinator alone removed the foreign-key errors but still failed the fixture's colocation-ID assertions; that insufficient control is retained too. The published correction is one DEPS entry, using existing multi_mx_create_table setup as the native schedule does. No SQL, expected result, foreign-key graph, synchronization assertion or runtime code changed. src/test/regress/citus_tests/run_test.py start_stop_metadata_sync --repeat 8 --use-whole-schedule-line passed 16/16 including setup on each PG15.20/16.16/17.12 (24 unchanged fixture executions).

The actual eight-test issue_8243 parallel schedule line was also replayed unchanged eight times on PG17: src/test/regress/citus_tests/run_test.py issue_8243 --repeat 8 --use-whole-schedule-line, 70/70 passed. This does not attribute or resolve the sporadic reference-DDL failure; no serialization or assertion suppression has been applied. The prior PG15 query-generator timeout did not recur in ce607 CI, but that is not timeout attribution or clearance of known deterministic wrong results.

New exact-head runs: full including advisories, PR, packaging. q13/q70 remain on HOLD with separate root-cause ownership. Dependencies/images/versions unchanged; no merge or stable promotion.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@ihalatci

Copy link
Copy Markdown
Contributor

Published 8b35613da1eb885d3c57605dbb288a6ba2e5c2c8: the existing EXPLAIN filter now recognizes optional indented Planning: labels, matching the independently confirmed release14 fix. Only three text-mode labels and their row counts changed; plan nodes, memory values, planning/execution times and structured-format output are preserved. The actual a020 PR artifact28 showed precisely this extra label (14→15 rows).

Controlled helper input retained Seq Scan, Memory, Planning Time and Execution Time while excluding both label indentations: 7/7 setup/control tests passed on PG15/16/17. Full run_test.py pg17 --repeat 8 --use-whole-schedule-line passed 24/24 on each major, including unchanged pg17_json cases.

Both a020 required CI gates passed. PR finished 180 success/9 failure/2 skipped (31/32 advisory passes; eight PG17 N-1 failures). Full finished 178 success/10 failure/2 skipped/one cancelled (29/32 advisory passes; two actual failures: unsupported-view dependency warning choice and issue_8243 reference-DDL error). The cancelled advisory22 ran pytest's 16 cluster-node-health items for six hours before cancellation; its log emitted 14 progress dots, no individual remaining test identities, and the uploader found no matching files. Its raw job log is retained; no artifact22 exists. This was not cancelled by a feature push, and no generic retry or timeout change was made. All 31 existing full advisory artifacts are retained.

The verified correction was pushed only after both prior runs completed. New exact-head full, PR, packaging. Remaining diagnostic/reference-DDL/hang causes are not declared resolved or masked. q13/q70 HOLD remains; no dependency/image changes, merge or stable promotion.

@ihalatci

Copy link
Copy Markdown
Contributor

Final current-head (8b35613da) outcomes: PR required CI passed, but full required CI failed. PR: 179 success/10 failure/two skipped. Full: 177 success/12 failure/two skipped. Packaging: 23/23 passed. Both advisory matrices are 30/32 passed, all 64 artifact payloads retained. No nested-EXPLAIN failure or six-hour pytest cancellation occurred this round; that does not establish clearance of prior hangs.

The full required failure is PG15 background_rebalance_parallel_reference_tables: representative/debug ordering and per-task shard labels changed. Actual expected/result comparison confirms 44 joined graph rows on each side, identical numeric task-dependency edges and identical colocation-slot/reference-dependency multisets; 12 shard-label rows differ. The earlier controlled metadata-heap-order experiment retained the same colocation/shard/placement state but switched the representative table. Existing #8789 already proposes sorting the representative candidate list by OID. It remains open and was inspected read-only; no runtime-policy backport or expectation normalization was applied without the separate scope decision.

The remaining advisory failures are actual reference-DDL errors (full issue_5248: relation OID17420 does not exist; PR issue_8243: table1 is not distributed) and the unsupported-view dependency warning naming either local_table_5 or local_table_6. These remain unresolved/unmasked. Eight PG17 N-1 failures remain in each run. q13/q70 HOLD remains, independently of CI results. No merge, stable promotion or dependency/image change.

@ihalatci

Copy link
Copy Markdown
Contributor

Final scope handoff: the user selected keeping #8789 separate from this PostgreSQL update. Its runtime representative-order policy will not be ported into #8977; no expectation masking or q13/q70 fix is included here. Existing validated fixture corrections remain at unchanged head 8b35613da1eb885d3c57605dbb288a6ba2e5c2c8.

Live state verified: PR required CI green; full-dispatch required CI red on the independently disclosed rebalance representative nondeterminism. Exact graph evidence remains preserved: 44 joined dependency rows on both sides, identical numeric dependency edges and identical colocation-slot/reference-dependency multisets, with 12 differing shard-label rows. Both advisory matrices remain 30/32 passed, packaging 23/23 passed. Residual reference-DDL failures and unsupported-view warning selection are unclassified and unmasked. q13/q70 root investigation remains separately owned and on HOLD. No merge or stable promotion.

Pair with the-process merge 7148bd3cb9f8b40b168b2021a5cb4c1644b05b61, published by workflow 45002632140. Leave the separately gated query-local planner correction unpublished.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants