Support read in reverse order with virtual row

[vdimir]

Changelog category (leave one):

• Performance Improvement

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

• Support the read_in_order_use_virtual_row optimization for reverse-order reads.

---

Note

Medium Risk
Touches MergeTree read-in-order execution paths and primary-key index access, which could affect query correctness/performance for reverse-order reads if edge cases around marks/index sizes are missed.

Overview
Enables the read_in_order_use_virtual_row optimization when reading MergeTree parts in reverse order (ReadType::InReverseOrder) by selecting the appropriate mark position from the part’s primary-key index and guarding against missing PK values (e.g., final mark / index bounds).

Updates the 03711_read_in_order_through_join stateless test to run the reverse-order join case with the same log_comment for both LEFT and INNER joins and adjusts expected OK rows accordingly.

^{Written by Cursor Bugbot for commit 7ecea2c. This will update automatically on new commits. Configure here.}

[clickhouse-gh]

Workflow [PR], commit [7ecea2c]

Summary: ❌

job_name	test_name	status	info	comment
Stress test (amd_msan)		failure
	Logical error: There is a race condition between creation and removal of replicated table. It's a bug (STID: 2508-1e6e)	FAIL	cidb, issue	ISSUE CREATED
Upgrade check (amd_release)		failure
	Error message in clickhouse-server.log (see upgrade_error_messages.txt)	FAIL	cidb	IGNORED

---

AI Review

1) Summary

This PR extends read_in_order_use_virtual_row to reverse-order reads by taking the virtual-row key from the range end mark when final marks are available, and updates 03711_read_in_order_through_join expectations so reverse-order INNER join also keeps read-in-order optimization. High-level verdict: the change is coherent and I did not find high-confidence correctness/safety issues in the final diff.

2) Missing context (if any)

• No Praktika CI report links or failed-job logs were provided in the task context.
• No benchmark results were provided for reverse-order read-in-order behavior.

3) Findings (by severity)

• Blockers: None.
• Majors: None.

4) Tests & Evidence

• Coverage assessment: the updated stateless test now validates both LEFT and INNER join cases for reverse-order read-in-order path and checks query-log behavior under a fixed log_comment, which is directly relevant to this change.
• Negative/error-path tests: not added in this PR (e.g., mixed parts with and without final marks).
• Additional tests to consider (follow-up, non-blocking): add a targeted case with parts lacking final marks to confirm safe fallback (no virtual row for those parts) and stable read-in-order correctness with joins.

5) ClickHouse Compliance Checklist (Yes/No + short note)

• Data deletions logged? Yes — no deletion-path changes in this PR.
• Serialization formats versioned? Yes — no serialization/protocol format changes.
• Experimental setting gate present? Yes — no new feature flag required; this extends existing guarded virtual-row behavior.
• Settings exposed for constants/thresholds? Yes — no new hardcoded tunables introduced.
• Backward compatibility preserved? Yes — behavior remains conditional on existing settings and part metadata.
• SettingsHistory.cpp updated for new/changed settings? Yes — no settings added/changed in final diff.
• Existing tests untouched (only additions)? No — one existing test was adjusted (expected for behavior extension).
• Docs/user-facing notes updated? N/A — internal optimizer behavior/test update only.
• Core-area change got extra scrutiny? Yes — reviewed ReadFromMergeTree read-in-order virtual-row logic specifically.

6) Performance & Safety Notes

• Hot-path impact appears low and localized: one additional branch for reverse-order virtual-row eligibility and mark-position selection.
• Memory/concurrency risk in the final diff is minimal (no new shared state, allocations, or threading primitives introduced).
• Benchmark data is missing; minimal reproducible benchmark: compare ORDER BY ... DESC LIMIT N with INNER JOIN on multi-part MergeTree datasets (with/without final marks), capturing read_rows and elapsed time.

7) User-Lens Review

• User-visible behavior becomes more intuitive: reverse-order INNER join can now keep the same read-in-order optimization path as LEFT in covered scenarios.
• Error/log surface is unchanged; no regressions in operator observability were found in the diff.

8) Final Verdict

• Status: Approve

[clickhouse-gh]

LLVM Coverage Report

Metric	Baseline	Current	Δ
Lines	83.20%	83.80%	+0.60%
Functions	23.60%	23.80%	+0.20%
Branches	75.30%	76.30%	+1.00%

PR changed lines: PR changed-lines coverage: 100.00% (21/21)
Diff coverage report
Uncovered code

[Fgrtue]

The change looks good in my opinion. The only thing that concerns me is the Stress test failure. @vdimir do you think it could be related?

[vdimir]

The change looks good in my opinion. The only thing that concerns me is the Stress test failure. @vdimir do you think it could be related?

I doubt this change can cause it, I created new issue but probably it's duplicate of #85771, since same error happens in CI with different stack trace