mirror of
https://github.com/block/buzz.git
synced 2026-10-27 04:30:52 +00:00
5b9dbcb4c2
When new messages arrive in a mobile channel, the app updates the list of loaded messages on the UI thread. It currently sorts that list twice. The second sort is needed only when the app has retained messages opened through a link to a specific message. This change uses the first sorted list when there are no retained messages to merge. It reduces work during message bursts and keeps the existing merge and ordering behavior for message links. A controlled debug test with 2,000 loaded messages and 1,000 incoming messages reduced processing time from 446 ms to 266 ms. This measures the message update code, not screen rendering or release-device performance. All 26 focused tests pass, including a regression test that fails if the extra pass returns. Tests also cover message order and retained messages. Repository checks and downstream test/build suites passed, apart from unrelated Rust timing failures in the full local run; each passed an isolated retry. GitHub CI passed after retrying an unrelated desktop attachment test failure. --------- Signed-off-by: Tom Brow <tomb@block.xyz>
31 lines
1.5 KiB
Markdown
31 lines
1.5 KiB
Markdown
# Mobile performance diagnostics
|
|
|
|
With the repository's Hermit environment active, run from `mobile/`:
|
|
|
|
```sh
|
|
flutter test --dart-define=BUZZ_RUN_BENCHMARKS=true test/benchmark/channel_live_burst_test.dart
|
|
```
|
|
|
|
The diagnostic is skipped by default. It does not use a relay, credentials, or
|
|
community content.
|
|
|
|
The buffered-channel diagnostic sends synthetic events through the real
|
|
`RelaySessionNotifier` event buffer and `ChannelMessagesNotifier` live callback.
|
|
Only connection state and history responses are replaced. It measures one
|
|
synchronous flush and a zero-delay timer queued just before that flush. The timer
|
|
shows how long other event-loop work waits behind the consumer callbacks.
|
|
|
|
It varies retained history (50, 500, 2000 events) and live burst size (50, 200,
|
|
1000 events). Each case warms up twice before emitting four JSON measurement
|
|
records in microseconds. Compare medians on the same host under similar load.
|
|
No wall-clock threshold is asserted because debug/JIT compilation and host load
|
|
make it unsuitable for a stable CI performance gate. Final event counts are
|
|
asserted to ensure the measured work actually completes.
|
|
|
|
This is a bounded stress diagnostic, not a representative release benchmark.
|
|
Retained history is modeled as one synthetic page instead of real pagination.
|
|
It excludes widgets, media, network traffic, signature verification, and other
|
|
subscriptions. It can identify synchronous consumer costs, but cannot establish
|
|
that a measured burst occurred in a real community or predict release frame
|
|
latency.
|