External/buzz

Fork 0

mirror of https://github.com/block/buzz.git synced 2026-08-18 06:50:31 +02:00

Files

T

History

01f76ec971 fix(desktop): cut steady-state relay traffic from polls and read-state echo (#5879 )

## Problem

Desktop webview CPU stayed high after the presence-scope fix (#5830) and
the shared useNow ticker (#5861). A per-kind byte tap hot-patched into
`relayClientSession.ts` on a live desktop (~500 channels, large agent
fleet; 850 s capture correlated with CPU sampling) showed the remaining
steady-state relay traffic is mostly self-inflicted:

| kind | what | share of inbound bytes | shape |
|------|------|-----------------------|-------|
| 30078 | read-state | **34%** | our own ~44 KB nip44 blob echoed back
every ~10-30 s while reading |
| 30030 | emoji union | **33%** | 2-min poll refetching every member's
full set (~300 KB burst) |
| 30175 | persona catalog | **13%** | same 2-min backstop pattern, ~150
KB per walk |

CPU tracked the bursts directly: 3-5% in quiet 10 s buckets vs 44-54% in
buckets containing a poll burst or read-state echo. (The kind-24200
observer-frame theory was tested and disproven by the same tap: 9.7% of
bytes, steady trickle.)

## Outcome

- **Read-state echo drop.** `ReadStateManager` remembers the ids of
events it just published (FIFO set capped at 64) and drops their relay
echoes before the nip44-decrypt + `JSON.parse` step. Ids are recorded
*before* publishing so relay fan-out can't race the OK. The drop
consumes the id, so a reconnect replay of the same event still parses
normally. Events from other clients of the same pubkey are untouched.
- **Poll backstops stretched 2 min → 20 min** for the emoji union and
persona catalog queries. The live subscriptions (invalidate on any new
30030/30175) and the reconnect invalidations remain the freshness paths;
the poll only exists to cover a silently dropped live event. Behavior on
publish, focus, and reconnect is unchanged.
- Mechanical: localStorage identity helpers moved to
`readStateIdentity.ts` (no behavior change) to keep
`readStateManager.ts` under the file-size ratchet.

Expected effect on the measured profile: the poll stretch cuts the
30030/30175 bursts (46% of inbound bytes) by 10x; the echo drop removes
the recurring ~44 KB nip44-decrypt + parse per publish cycle (the echo
still arrives on the wire — nostr filters cannot exclude own-author
events — so this is a CPU/IPC saving, not a bandwidth one).

## Acceptance

- New tests: echo dropped **before** decrypt (mutation-checked:
disabling the drop fails the test), replayed duplicate of the same id
still parses, foreign-client events always parse, published-id set stays
capped when publishes fail (never-echoed ids).
- Full desktop suite **4794/4794**, `tsc --noEmit` clean, `pnpm check`
(biome + ratchets) clean at head.

## Not addressed (follow-ups)

- The 44 KB blob itself (one read-state event carries all ~500 channels;
a delta or per-channel-shard format is a protocol change).
- Duplicate delivery of the same events on concurrent `history-`
subscriptions (relay/client dedupe).
- Webview RSS of 12.5 GB observed on the same machine — retention hunt
is separate work; shrinking the heap multiplies the value of this PR
since the GC floor scales with live-heap size.

Signed-off-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Co-authored-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>

2026-08-14 14:38:33 -04:00