Update two-stacks thresholds of promql function deriv()

[vitlibar]

Changelog category (leave one):

• Performance Improvement

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

Update two-stacks thresholds of promql function deriv()

---

Workflow [PR]
Sync PR [sync-upstream/pr/118987]

Version info

• Merged into: 26.9.1.1073 (included in 26.9 and later)

[clickhouse-gh]

Workflow [PR], commit [46d27f5]

Summary: ✅

• Performance Comparison: Performance dashboard

---

AI Review

Summary

This PR lowers the two-stack selection thresholds used by the shared linear-regression time-series aggregators and updates the stateless/performance tests so they still exercise the intended recompute, average-density, and hard-cap paths. I did not find a code-level correctness or coverage problem in the implementation, but the PR template metadata does not describe the actual change accurately.

PR Metadata

• Changelog category is not correct for this change. This is a performance-tuning change for existing experimental functions, not a new experimental feature. Use Performance Improvement.
• A Changelog entry is required for Performance Improvement, and the current entry is too vague and only names deriv() even though the shared linear-regression trait also changes timeSeriesPredictLinearToGrid.
• Suggested replacement:

### Changelog category (leave one):
- Performance Improvement

### Changelog entry (a [user-readable short description](https://github.com/ClickHouse/ClickHouse/blob/master/docs/changelog_entry_guidelines.md) of the changes that goes into CHANGELOG.md):
Lowered the threshold that switches experimental `timeSeriesDerivToGrid` and `timeSeriesPredictLinearToGrid` from window recomputation to the two-stack sliding aggregator, improving finalize performance for windows with 4 to 11 populated buckets.

Final Verdict

✅ No technical blockers in the code or tests. Fix the PR metadata before merge.

LLVM Coverage Report

⚠️ No coverage measurement for commit 46d27f5: incomplete coverage measurement: 1 of 21 shard profiles are missing: LLVM_COVERAGE_FILE_it_7.profdata.

[vitlibar]

Recent measurements after our optimizations showed different values than what were measured originally:

ClickHouse/src/AggregateFunctions/examples/timeseries_to_grid_two_stack_vs_recompute.out

Lines 23 to 24 in a05aaa1

	AVG_POPULATED_BPW_TO_ENABLE_TWO_STACKS (two-stacks first wins) = 4
	BPW_TO_FORCE_TWO_STACKS (two-stacks > 2x faster) = 12

So these thresholds in the code should be updated

[vitlibar]

Also the performance and stateless tests were adjusted to follow their claimed scenarios.

[clickhouse-gh]

Build profile diff (arm_release)

Comparing 46d27f54c with master a05aaa178 (stripped binary size, per-symbol sizes and ThinLTO time; compile times per translation unit against the most recent warmup build that recompiled it).

✅ No significant changes.

Binary sizes

programs/clickhouse-stripped: smaller than the master baseline by the known offset between the two builds, so the difference is not shown. A delta that differs from the offset by more than 50% of it is shown, in either direction.

The official master build is compiled with -g and a pull request build is not, and XRay counts debug instructions towards its instrumentation threshold, so master instruments thousands of functions more and its binary is ~0.4% larger no matter what the pull request does.

Compile time of recompiled translation units

8 translation units recompiled, 45 s compile time in total, 8 of them have a recent master baseline.

Job report