Fix distributed global right join by ianton-ru · Pull Request #2315 · Altinity/ClickHouse

ianton-ru · 2026-09-04T13:34:53Z

Changelog category (leave one):

Bug Fix (user-visible misbehavior in an official stable release)

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Fix distributed global right join

Documentation entry for user-facing changes

Swap left and right tables for distributed right join.
Fix for object_storage_cluster_join_mode='global' - it was broken after frontports and executed as local.

CI/CD Options

Exclude tests:

Regression jobs to run:

The expected result matches initiator-side join semantics; the test currently fails because shards emit extra unmatched right-side rows. Related: #1664 Co-authored-by: Cursor <cursoragent@cursor.com>

Broadcast the original left table and keep the preserved right side sharded, so unmatched right rows are not duplicated on every shard. Related: #1664 Co-authored-by: Cursor <cursoragent@cursor.com>

After ClickHouse#94748 the planner always wrapped the left cluster source, so object_storage_cluster_join_mode='global' never saw the JOIN and ran on the initiator. Keep wrapping for allow/local and for RIGHT/FULL to avoid duplicated unmatched rows. Co-authored-by: Cursor <cursoragent@cursor.com>

github-actions · 2026-09-04T13:36:11Z

Workflow [PR], commit [449a143]

ianton-ru · 2026-09-07T08:00:24Z

@blau-ai

blau-ai · 2026-09-07T08:04:52Z

CI triage for #2315 (@ianton-ru)

Verdict: 0 of the failures are caused by this PR. All red checks are pre-existing flakiness / infra / runner-environment issues in subsystems your change never touches. This PR only edits src/Planner/PlannerJoinTree.cpp plus two tests, and its own new tests pass (05058_distributed_global_right_join is green in every stateless run that executed it; nothing from test_s3_cluster appears in the failure set). Re-running the failed jobs should clear them.

Head SHA: 5ca17e5.

Per-failure breakdown

1. Source upload — infra (not PR)
Checkout submodules step failed: fatal: could not read Username for 'https://github.com' while cloning contrib/* (llvm-project, aws, …), retried 3× and aborted (exit code 123). Pure network/credential failure on the runner, unrelated to source. → Re-run.

2. Integration tests (amd_asan_ubsan, db disk, old analyzer, 4/8) — pre-existing/infra (not PR)
All 12 failures are test_storage_iceberg_with_spark/test_schema_inference.py::test_schema_inference[...] (every s3/azure/local parametrization). Iceberg + Spark schema inference has nothing to do with analyzer JOIN planning; these are heavy external-dependency (Spark/MinIO/Azure) tests that fail as a block when the Spark side is unhealthy. → Re-run.

3. Integration tests (amd_asan_ubsan, db disk, old analyzer, 7/8) — pre-existing/infra (not PR)
test_database_iceberg_lakekeeper_catalog/test.py::test_auth_token_profile_events — Iceberg/Lakekeeper REST catalog auth, again unrelated to the planner. → Re-run.

4. Stateless tests (amd_binary, cas s3 storage, parallel) — flaky storage layer (not PR)
02265_column_ttl: Code: 210 ... did not prove it still holds the manifest it offered for part ... by relink ... (NETWORK_ERROR) from DataPartsExchange::Fetcher::relinkPartToDisk. This is the CAS/shared-storage part-exchange path (Antalya feature), not the query planner. The harness self-diagnosis even labels it flaky. → Re-run.

5. Stateless tests (amd_asan_ubsan, cas s3 storage, parallel, 2/2) — environment / flaky (not PR)

04105_limit_by_into_sort_explain: result diff is only stream counts (MergingSortedTransform 12 → 1 vs 6 → 1, 24→12, etc.). That is the runner's CPU/max_threads count, not query semantics — an environment-sensitive EXPLAIN reference, independent of the JOIN change.
04033_tpc_ds_q60: Code: 241 ... memory limit exceeded (14.55 GiB > 14.40 GiB). Passed on rerun (0/3 reruns failed) — memory pressure. Flaky.
→ Re-run.

6. Stateless tests (amd_asan_ubsan, cas s3 storage, parallel, 1/2) — server death, collateral (not PR)
Server died; the 6 UNKNOWN tests (00051_min_max_array, 02122_parallel_formatting_JSONStrings, …) are just the ones that never ran after the server went down. Same CAS-s3-storage config as #4 above; a source-only planner change cannot make the server die at this scale, and the affected tests are unrelated. → Re-run.

7. Stress test (amd_asan_ubsan) — infra (not PR)
Cannot start clickhouse-server → Check failed. The server never started, so no stress workload ran. Builds all passed and Stress test (amd_debug) is green, so this is a startup/environment flake, not the diff. → Re-run.

Bottom line

Nothing here needs a code change. The PR-relevant signal is all green: builds pass on every arch, Fast test / Stateless tests (amd_debug, parallel) / distributed-plan variants pass with 0 failures, and the regression suites (swarms, s3_export, parquet, oauth) are green. The failures are the usual Iceberg/Spark, CAS-s3-storage, thread-count-EXPLAIN, memory-limit, and submodule-checkout flakes. I'd just re-run the seven failed jobs.

(No fix pushed — there's nothing PR-caused to fix. Happy to re-check after a re-run if any of these persist.)

— @blau-ai

Port of the review fixes made for the `master` pull request ClickHouse#118611 `tryRewriteGlobalRightJoinAsLeftJoin` only recognised a narrow shape of the query it was meant to fix, so a `GLOBAL RIGHT JOIN` whose left table is a sharded `Distributed` table still emitted the rows preserved from the broadcast side once per shard in these cases: * `USING` and `NATURAL` joins were skipped, because only `ON` was accepted. A `USING` key stores its sides positionally, so the key pairs are now swapped together with the table expressions, and a key that does not hold a plain column per side leaves the query alone. * The rewrite required the right side to be a sharded `Distributed` table too. Only the left table fans the query out across shards, so the gate is now on the left table alone, and a local table, a one shard `Distributed` table, a subquery or a table function on the right is handled as well. `Any`, `Semi` and `Anti` join over to the swapped sides the same way `All` does and are accepted; `Asof` and `RightAny` do not mirror and stay out. * Only the topmost join was examined. Join trees are left deep, so the walk now descends to the join that reads the leftmost table. A `RIGHT` or `FULL` join that survives the rewrite, such as the one in `(A INNER C) RIGHT B`, cannot be swapped without moving a join into the right table expression. Those now read the leftmost table through a subquery, which keeps the join on the initiator, the only remaining way to emit the preserved rows once. `getDistributedStorageFromTableExpression` also missed `remote` and `clusterAllReplicas` on the left, because it only looked at `TableNode`. Co-authored-by: Cursor <cursoragent@cursor.com>

…istributed tables. An `Alias`, `Buffer`, `MaterializedView` or `StorageProxy` in front of a sharded `Distributed` table still fans the query out across shards, so unmatched right rows would otherwise be duplicated once per shard. Co-authored-by: Cursor <cursoragent@cursor.com>

…al-right-join-test

ianton-ru and others added 3 commits September 4, 2026 13:30

Add a regression test for global RIGHT JOIN on Distributed tables

cba62d6

The expected result matches initiator-side join semantics; the test currently fails because shards emit extra unmatched right-side rows. Related: #1664 Co-authored-by: Cursor <cursoragent@cursor.com>

Rewrite global RIGHT JOIN of Distributed tables as LEFT JOIN

3ec16e4

Broadcast the original left table and keep the preserved right side sharded, so unmatched right rows are not duplicated on every shard. Related: #1664 Co-authored-by: Cursor <cursoragent@cursor.com>

ianton-ru changed the title ~~Bugfix/antalya 26.6/distributed global right join test~~ Fix distributed global right join Sep 7, 2026

ianton-ru added antalya antalya-26.6 labels Sep 8, 2026

ianton-ru and others added 2 commits September 9, 2026 13:09

Merge branch 'antalya-26.6' into bugfix/antalya-26.6/distributed-glob…

449a143

…al-right-join-test

ianton-ru marked this pull request as ready for review September 10, 2026 08:02

Navigation Menu

Uh oh!

Fix distributed global right join - #2315

Fix distributed global right join#2315
ianton-ru wants to merge 6 commits into
antalya-26.6from
bugfix/antalya-26.6/distributed-global-right-join-test

ianton-ru commented Sep 4, 2026

Uh oh!

github-actions Bot commented Sep 4, 2026 •

edited

Loading

Uh oh!

ianton-ru commented Sep 7, 2026

Uh oh!

blau-ai commented Sep 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

ianton-ru commented Sep 4, 2026

Changelog category (leave one):

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Documentation entry for user-facing changes

CI/CD Options

Exclude tests:

Regression jobs to run:

Uh oh!

github-actions Bot commented Sep 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ianton-ru commented Sep 7, 2026

Uh oh!

blau-ai commented Sep 7, 2026

CI triage for #2315 (@ianton-ru)

Per-failure breakdown

Bottom line

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

github-actions Bot commented Sep 4, 2026 •

edited

Loading