Antalya 25.8 Backport of #2020 - Fix crash merging _part_offset projection on row-combining engines

[mkmkme]

25.8: Fix crash merging _part_offset projection on row-combining engines

Same as #2020, but for Antalya 25.8

Changelog category (leave one):

• Critical Bug Fix (crash, data loss, RBAC)

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

Fixes a bug where OPTIMIZE FINAL on row-combining engines could crash the server (ClickHouse#106929)

Documentation entry for user-facing changes

...

CI/CD Options

Exclude tests:

• Fast test
• Integration Tests
• Stateless tests
• Stateful tests
• Performance tests
• Aarch64 tests
• All with ASAN
• All with TSAN
• All with MSAN
• All with UBSAN
• All with Coverage
• All Regression
• Disable CI Cache

Regression jobs to run:

• Fast suites (mostly <1h)
• Aggregate Functions (2h)
• Alter (1.5h)
• Benchmark (30m)
• ClickHouse Keeper (1h)
• Iceberg (2h)
• LDAP (1h)
• OAuth (5m)
• Parquet (1.5h)
• RBAC (1.5h)
• SSL Server (1h)
• S3 (2h)
• S3 Export (2h)
• Swarms (30m)
• Tiered Storage (2h)

[github-actions]

Workflow [PR], commit [a635191]

[mkmkme]

@blau-ai

[blau-ai]

CI triage — 4 failing checks: 0 PR-caused, 3 infra/flaky, 1 aggregate gate

Short version: none of the red checks are caused by this backport. The diff only flips one boolean in MergeTask.cpp (all non-Ordinary merge modes now rebuild _part_offset projections instead of merging them) and adds a stateless test — it cannot affect builds timing out, ParallelReplicas connectivity, or swarm-cluster node-failure tests. Positive signal: the PR's own regression test passed, and the debug build/BuzzHouse are green (see below). Recommendation: re-run the three failing jobs, no code change needed.

1. Build (amd_release) — infra / timeout, not PR-caused

Report: Job got terminated with an error, exit code [-15] after 2h6m28s (job). Exit -15 = SIGTERM, i.e. the runner killed the job (build-time limit / spot reclaim), not a compile error. A one-line logic change can't make a build hang. The amd_debug build in the same run compiled the identical change fine.
→ Re-run.

2. Stateless (amd_binary, ParallelReplicas, s3 storage, parallel) — flaky, non-blocking

The single failure is unrelated to this PR:

03156_nullable_number_tips.sql (query 9, line 11: SELECT DISTINCT dumpColumnStructure(*) FROM test;)
Code: 279. DB::Exception: Can't connect to any replica chosen for query execution:
While executing Remote. (ALL_CONNECTION_TRIES_FAILED)

This is a ParallelReplicas connectivity flake in a dumpColumnStructure test — nothing to do with merges/projections. CI itself treats it as non-blocking (NOTE: Failed 1 tests - do not block pipeline, exit with 0) and the GitHub Actions job concluded success (job); only the S3 commit-status is red.
→ Ignore / re-run.

3. Regression x86 Swarms — flaky, not PR-caused

3 features / 2 scenarios failed (job):

✘ /swarms/feature/node failure/check restart swarm node
✘ /swarms/feature/node failure/swarm out of disk space
✘ /swarms/feature/swarm joins
Code: 394. DB::Exception: ... Query '…' is killed in pending state. (QUERY_WAS_CANCELLED)
  (on a static_swarm_cluster query over a datalake/iceberg table, sleepEachRow(1))

These are object-storage swarm-cluster resilience tests (deliberately killing nodes / filling disk mid-query), reading external tables — no code path overlaps with MergeTree projection merges. QUERY_WAS_CANCELLED during a forced node kill is a known race. Swarms was green on the latest base-branch commit (a6839be → RegressionTestsRelease / Swarms / swarms: success), so this is intermittent, not a regression from the diff.
→ Re-run.

4. PR — aggregate status

Red only because it summarizes the three jobs above. Clears once they pass.

---

Health check / positive evidence

• The PR's own regression test passed: 03786_projection_part_offset_reducing_merge: [ OK ] in Stateless tests (amd_debug, parallel) (job).
• Build (amd_debug) and BuzzHouse (amd_debug) are green.
• The change is a minimal, well-scoped backport of 25.8: Fix crash merging _part_offset projection on row-combining engines #2020 / upstream fix — one condition (mode != Ordinary) plus a targeted test. No divergence concerns.

Bottom line: safe to re-run the three failing jobs; I'd expect green. No fix required from this PR.

_{🤖 Generated by @blau-ai · evidence from CI run 30079153187}