[kurok]

Closes #242

The bug

LatencyBreakdown.available was not serialized, so the master rebuilt every breakdown with the default full phase set. aggregate_breakdowns intersects available to omit phases a backend cannot measure — so a --http2 --latency-breakdown distributed run reported:

dns:     {avg: 0.0, min: 0.0, max: 0.0, p50: 0.0, p95: 0.0, count: 5}
connect: {avg: 0.0, ...}
tls:     {avg: 0.0, ...}

Five samples of a measured-zero connection phase that never happened. That reads as an impossibly fast connection rather than an unknown one — which is exactly the failure mode the available mechanism was built to prevent.

Single-node --http2 got this right. Only the cluster path lied.

The fix

available travels with each sample and is restored on the master.

A worker from an older release does not send the key. The full phase set is what it would have meant, so a mixed-version cluster still merges rather than silently dropping every phase. A malformed available is rejected by name, consistent with the rest of _deserialize_stats after #237.

Tests

tests/test_distributed.py::TestBreakdownAvailableRoundTrip, all four verified against the unpatched source:

• test_breakdown_available_round_trips — the issue's named case. Pre-fix: Tuples differ: ('dns', 'connect', 'tls', 'ttfb', 'transfer', 'total') != ('ttfb', 'transfer', 'total').
• test_an_httpx_worker_does_not_contribute_phantom_connection_phases — runs the restored samples through aggregate_breakdowns and asserts dns/connect/tls are absent. This is the one that tests the behaviour rather than the field: pre-fix it prints the phantom {'avg': 0.0, 'count': 5} blocks in the failure message.
• test_a_worker_that_does_not_send_available_means_the_full_set — the mixed-version path.
• test_a_malformed_available_is_named.

Checks

ruff format/check clean, mypy clean, 1928 passed.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!