25.8 stable backport of #83096: A bunch of micro-optimizations to speed up small queries

[ilejn]

Changelog category (leave one):

• Performance Improvement

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

A bunch of micro-optimizations to speed up small queries

References ClickHouse#81043

(ClickHouse#83096 by @Algunenano)

CI/CD Options

Exclude tests:

• Fast test
• Integration Tests
• Stateless tests
• Stateful tests
• Unit 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)
• CAS (content-addressed storage; Antalya only)
• 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 [1d8f092]

[mkmkme]

@blau-ai

[blau-ai]

CI triage for #2535 (backport/stable-25.8/83096 → stable-25.8)

Verdict: 0 PR-caused, 5 not-PR-caused (1 flaky, 2 infra, 2 pre-existing image CVEs). Nothing in the red checks points at this backport. This PR is a pure perf micro-optimization (swapping std::stack/std::deque → boost::container::devector in the pipeline executor, plus an IMergeTreeReader column-handling refactor); none of the failures touch that code. My recommendation is to re-run the failed jobs; no code change is warranted.

---

1. Stateless tests (amd_ubsan, parallel) — FLAKY, not PR-caused
Single failure out of 7967: 03221_merge_profile_events.

 Vertical 1 20000 1 2 480000 1 1 1 1 1 1 1 1   (reference)
 Vertical 1 20000 1 2 480000 1 1 1 1 1 1 0 1   (got)

The only differing column is ProfileEvents['UserTimeMicroseconds'] > 0, which returned 0 for the small (20 000-row) Vertical merge. UserTimeMicroseconds is OS thread-CPU accounting (getrusage/taskstats); for a tiny, fast merge it can round to 0, especially on a virtualized UBSan runner. All the merge-logic columns (algorithm, rows, columns, bytes, stage timings) match exactly — only the CPU-time counter flipped. The PR does not touch profile-event or merge CPU accounting, so this is a timing flake.
→ Safe to re-run.

2. Stress test (amd_debug) & (amd_ubsan) — INFRA, not PR-caused
Both report check_status.tsv doesn't exists; the real cause is in test_results.tsv:

Cannot start clickhouse-server  FAIL
<Error> Application: Caught exception while loading metadata:
  Azure::Storage::StorageException ... Code 111, Connection refused

The server couldn't start because the Azurite (Azure blob) emulator was unreachable (connection refused during metadata load), so the stress harness never produced a status file. This is an environment/sidecar-container issue, unrelated to the diff.
→ Safe to re-run.

3. Grype Scan (clickhouse-keeper) / GrypeScanKeeper — PRE-EXISTING, not PR-caused
"20 high/critical vulnerabilities" in the clickhouse-keeper container image. These are OS/base-image package CVEs surfaced by the image scanner; this PR changes only C++ source and cannot affect them. This gate is failing image-wide on the 25.8 line, independent of the backport.
→ Needs an image/package bump (separate from this PR), not a change here.

4. "PR" check — aggregate gate. Red only because it rolls up the checks above; no independent failure.

---

Head SHA analyzed: 1d8f092 · report: result_pr.json on the praktika S3 bucket. If a re-run of the stateless + stress jobs comes back green, this backport is clear to merge.