[marcin-kordas-hoc]

Context

The PR adds an advisory workflow that compares base and head by the CPU instructions each performance scenario executes, instead of by time.

• Why: on shared runners, wall-clock noise lets the timing comparison (HF-150: alternate base and head runs in the performance workflow #1805) show only regressions of about 5% or more per scenario, and it still reports false regressions. Instruction counts do not depend on the load of the runner.
• What changed:
  • .github/workflows/performance-instructions.yml: installs valgrind, counts the instructions of every scenario for base and head (one after the other, from the same build directory), writes the comparison to the job summary and prints the raw counts into the log. Paths follow the monorepo layout.
  • The job never fails the pull request (continue-on-error) and posts no comment. performance.yml is unchanged.
  • When the paired tests branch has no performance/run-instructions.ts, the job skips the counts and the comparison with a notice instead of failing with module-not-found.
  • hyperformula/dev-docs/PERFORMANCE.md: one sentence about this workflow.
• Result: every PR gets a per-scenario instruction-count comparison that can see changes well under 1%.
• Role: the intended direction for the performance check, done in-house with valgrind on standard runners and without a vendor service. The wall-clock comparison stays as a temporary advisory layer, because instruction counts do not see GC pauses, memory or cache effects.

How did you test your changes?

• Paired tests branch: feat/hf-150-instruction-count in hyperformula-tests.
• This PR changes no engine code, so every scenario should come out at about 0%. CI runs:

Run	Largest change over 13 scenarios	Job time
before the monorepo merge	0.0075%	7m39s
after the monorepo merge (63b4b92)	0.00022%	10m37s
with the runner check and the scenario table (7ded1cc, tests dae23a2)	0.0056%	11m58s

• A throwaway control PR that added a fixed loop to setCellContents moved every scenario that calls it by +0.06% to +3.97%, and every other scenario by at most 0.013%.
• Locally, in the monorepo layout: one scenario counted end to end under valgrind.
• The runner check, run as a script against three tests refs: develop and feat/hf-150-perf-gate (no runner) print the notice and set no output, so the three guarded steps skip; feat/hf-150-instruction-count sets present=true.

Compatibility with Excel and Google Sheets

Not applicable.

Public API changes

None.

Out of scope and follow-ups

• A blocking check.
• Scale scenarios: stacked PR HF-150: run the performance workflows against 100k-row scale scenarios #1838 (100k rows in the wall-clock suite, 30k in the instruction count).

Types of changes

• Breaking change (existing functionality no longer works as before) – agreed with the maintainers
• New feature or improvement (a non-breaking change that adds functionality)
• Bug fix (a non-breaking change that fixes an issue)
• Additional language file, or a change to an existing language file (translations)
• Change to the documentation

CI tooling only, plus one sentence in dev-docs/.

Related issues

1. HF-150 (internal task)

Checklist

• My pull request meets the Definition of Done. External contributors: see also Contributing to HyperFormula.
• I have signed the Contributor License Agreement.
• I read my own diff end to end before requesting a review.
• I added a CHANGELOG.md entry, written by the changelog rules – or none is needed (docs, tests or tooling only).
• I updated the documentation and JSDoc to match the change.
• Breaking change only: I added a section to the migration guide.
• I updated the dev-docs/ pages this change affects – or none needed it.

🤖 Generated with Claude Code

---

Note

Low Risk
CI and dev-docs only; no engine or API changes, and the job cannot fail the pull request.

Overview
Adds a prototype GitHub Actions workflow Performance (instruction count) that runs on pull requests alongside the existing wall-clock performance job.

The new job checks out base and head, syncs the shared hyperformula-tests suite to both trees, and uses valgrind/callgrind to count CPU instructions for each performance scenario on each revision (sequentially, on the same Node 22 build). It writes a base-vs-head comparison to the job summary and logs raw JSON counts. The job is advisory only: continue-on-error: true, no PR comment, and no change to performance.yml.

If the paired tests branch lacks performance/run-instructions.ts, guarded steps are skipped with a notice instead of failing.

hyperformula/dev-docs/PERFORMANCE.md gains one sentence documenting this workflow and its prototype status.

^{Reviewed by Cursor Bugbot for commit 2f13c8b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	hyperformula-docs	d0ac897	Commit Preview URL Branch Preview URL	Nov 04 2026, 10:44 AM

[github-actions]

Performance comparison of head (d0ac897) vs base (b3f2f68) with the tests at 6d7003d

Scenario	base median (ms)	head median (ms)	change	threshold	noise (CV base/head)	verdict
Sheet A	449.1	445.1	-0.88%	±5.0% †	2.3% / 2.1%	ok
Sheet B	197.4	194.8	-1.30%	±7.4% †	3.5% / 4.4%	ok
Sheet T	156.6	159.3	+1.68%	±8.0% †	3.6% / 4.9%	ok
Column ranges	451.0	447.6	-0.76%	±5.3% †	1.3% / 1.3%	ok
Sorted lookup	11316.4	11437.4	+1.07%	±50.0% ‡	1.3% / 1.5%	ok
Sheet A: change value, add/remove row/column	132.0	172.9	+30.96%	±30.3%	22.6% / 38.0%	inconclusive
Sheet B: change value, add/remove row/column	166.4	176.2	+5.90%	±13.3% †	7.4% / 5.7%	ok
Column ranges - add column	147.3	148.1	+0.55%	±6.4% †	4.5% / 3.1%	ok
Column ranges - without batch	389.4	404.5	+3.88%	±6.3% †	5.0% / 1.2%	ok
Column ranges - batch	120.4	121.4	+0.81%	±5.1% †	3.7% / 3.5%	ok
Recalculation fan-out - one edit, 10k dependents	42.7	45.8	+7.25%	±6.4% †	7.1% / 3.5%	regression
Recalculation fan-out - 5 edits without batch	125.9	130.1	+3.32%	±9.0% †	5.5% / 3.6%	ok
Recalculation fan-out - 5 edits in batch	43.9	45.8	+4.37%	±6.8% †	4.1% / 2.9%	ok

Threshold of an unmarked scenario = max(5%, 3 x the standard error of the difference of the two medians). † = the band of changes observed on identical code in earlier runs, instead of that rule. ‡ = a fixed threshold for a scenario that guards against a change of complexity class. This check is advisory: it reports regressions but does not fail the build.

• Scenario 'Sheet A: change value, add/remove row/column': the merged rounds of one side differ by more than 20%, so it is not judged (a slow phase of the runner, not the code, is the likely cause).

[Tobiadefami]

Nothing checks that performance/run-instructions.ts exists in the paired tests branch. A tests branch cut before hyperformula-tests#91, or created from a develop that doesn't have the runner yet, makes this step stop with a ts-node module-not-found error and leaves the summary empty. If this PR merges before #91, every PR gets that failing job.

Could we guard it the way performance.yml guards noise-history.json, and skip with a visible notice? Add this step before the count steps:

      - name: Check for the instruction-count runner
        id: runner
        run: |
          if [ -f head/hyperformula/test/hyperformula-tests/performance/run-instructions.ts ]; then
            echo "present=true" >> "$GITHUB_OUTPUT"
          else
            echo "::notice::The paired hyperformula-tests branch predates the instruction-count runner; skipping."
          fi

Then add if: steps.runner.outputs.present == 'true' to the two count steps and to "Compare the counts".

[marcin-kordas-hoc]

Added the "Check for the instruction-count runner" step; the two count steps and "Compare the counts" now run only when it sets present=true (7ded1cc). #1852 has the same check in the instruction-count job of performance.yml (e36fd13).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (feat/hf-150-perf-gate@b3f2f68). Learn more about missing BASE report.

Additional details and impacted files

@@                   Coverage Diff                    @@
##             feat/hf-150-perf-gate    #1814   +/-   ##
========================================================
  Coverage                         ?   97.36%           
========================================================
  Files                            ?      209           
  Lines                            ?    16524           
  Branches                         ?     3670           
========================================================
  Hits                             ?    16089           
  Misses                           ?      427           
  Partials                         ?        8           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.