Performance

Generated page

Written by tools/perf_report.py render from the nightly benchmark run (.github/workflows/performance-nightly.yml) and redeployed to the dev docs version. Editing it by hand is pointless — the next nightly overwrites it.

This is Phase 0 of #131: the measurement the later phases are supposed to be argued from. It is a dashboard, not a gate — nothing in CI fails because a number here moved, because hosted runners are shared VMs and a nightly that cried wolf would be muted inside a week. Read it for trends, not for single runs.

Latest run

• When: 2026-08-20T18:00:23.371329+00:00
• Commit: 188767e1fb98 (refs/heads/master)
• Runner: seed (local, macOS arm64), Python 3.12.11
• Recorded runs in history: 1

This is the seed run committed with #132 — the trend charts appear once the nightly has recorded a second one.

Percentages compare against the baseline recorded on seed (local, macOS arm64) (2026-08-20, 188767e1fb98).

EventBus.publish

The floor under everything else. publish snapshots both subscriber lists defensively, which is the allocation #131 Phase 1 proposes to replace with copy-on-write — the 0-subscriber row is what that would be measured against.

Benchmark	Mean	vs baseline
test_bench_bus_publish[0]	160 ns	+0.0%
test_bench_bus_publish[10]	499 ns	+0.0%
test_bench_bus_publish[1]	184 ns	+0.0%
test_bench_bus_publish[50]	1.84 µs	+0.0%
test_bench_bus_publish_to_firehose	182 ns	+0.0%

LogPipeline.process (full backend)

One replay of 60 in-game seconds of traffic through the parser chain and every handler subscribed to it. Divide by the line count in the table for per-line cost; the driver thread has 100 ms per poll to absorb whatever arrived in it.

Benchmark	Mean	vs baseline
test_bench_pipeline_cold_backend	115.29 ms	+0.0%
test_bench_pipeline_corpus	1.56 ms	+0.0%
test_bench_pipeline_profile[group]	9.97 ms	+0.0%
test_bench_pipeline_profile[raid]	27.03 ms	+0.0%
test_bench_pipeline_profile[solo]	3.81 ms	+0.0%

Per line: test_bench_pipeline_cold_backend: 633 lines, 182.14 µs/line; test_bench_pipeline_corpus: 93 lines, 16.77 µs/line; test_bench_pipeline_profile[group]: 440 lines, 22.67 µs/line; test_bench_pipeline_profile[raid]: 1264 lines, 21.38 µs/line; test_bench_pipeline_profile[solo]: 167 lines, 22.84 µs/line.

Plugin subscriber dispatch

The same publish, but through HostPluginContext — so the per-plugin try/except guard and the telemetry gate are both inside the number. off and on are collection disabled and enabled: the gap is the entire cost of #132's measurement on the dispatch path.

Benchmark	Mean	vs baseline
test_bench_plugin_dispatch[1-off]	203 ns	+0.0%
test_bench_plugin_dispatch[1-on]	484 ns	+0.0%
test_bench_plugin_dispatch[10-off]	745 ns	+0.0%
test_bench_plugin_dispatch[10-on]	3.32 µs	+0.0%
test_bench_plugin_dispatch[50-off]	3.06 µs	+0.0%
test_bench_plugin_dispatch[50-on]	16.64 µs	+0.0%

Plugin parser in the chain

A plugin parser appended after the built-in chain, over a raid burst. It never consumes a line, which is the worst case: every line reaches it.

Benchmark	Mean	vs baseline
test_bench_plugin_parser[off]	18.94 ms	+0.0%
test_bench_plugin_parser[on]	14.36 ms	+0.0%

Per line: test_bench_plugin_parser[off]: 633 lines, 29.93 µs/line; test_bench_plugin_parser[on]: 633 lines, 22.69 µs/line.

Qt bridge (buffer + coalesced flush)

1000 events published from a worker thread and delivered in one coalesced GUI-thread flush. burst has a per-event slot connected, batch_only just the bulk signal.

Benchmark	Mean	vs baseline
test_bench_qt_bridge_batch_only	1.05 ms	+0.0%
test_bench_qt_bridge_burst	1.19 ms	+0.0%

Map canvas render

The heaviest widget in the app, coalescing 120 location fixes into one render.

Benchmark	Mean	vs baseline
test_bench_map_location_burst	348.55 µs	+0.0%

End-to-end latency (log append -> UI slot)

A real LogDriver tailing a real file, so this includes the 100 ms poll interval it is dominated by. That is the honest end-to-end figure — it is what the user waits.

Benchmark	Mean	vs baseline
test_bench_latency_append_to_ui	102.05 ms	+0.0%
test_bench_latency_parse_to_ui	72.24 µs	+0.0%

Running it yourself

QT_QPA_PLATFORM=offscreen uv run pytest -m benchmark --benchmark-only

Benchmarks are excluded from the default uv run pytest (see the addopts in pyproject.toml) because timing assertions in CI are noise. The fixtures they replay are in tests/perf/profiles.py: solo, group and raid traffic composed from the EQtoolsTests line corpus.