vitest-driven benches at `packages/bridge/bench/recv.bench.ts` and
`packages/onebot/bench/send.bench.ts`. Picks up proton's vite plugin
so call sites get the inlined codecs (running via tsx + proton
runtime-fallback would throw without a pre-built runtime map).
Run via `pnpm -F @snowluma/bridge bench` and
`pnpm -F @snowluma/onebot bench`.
What's measured
---------------
Receive chain:
- parseMsgPush: pure decode + dispatch (no SQLite, no event emit)
* group_message text ~ 0.29 μs/op (3.4M ops/sec)
* group_message 5 mixed ~ 0.50 μs/op (2.0M ops/sec)
* group_member_join ~ 0.25 μs/op (4.1M ops/sec)
- IncomingPacketPipeline.process: full pipeline (with SQLite via
rememberGroupMemberIdentity, BridgeEventBus emit)
* group_message text ~ 0.42 μs/op (2.4M ops/sec)
* group_message 5 mixed ~ 0.67 μs/op (1.5M ops/sec)
* group_member_join ~17.5 μs/op ( 57K ops/sec) ⚠
Send chain:
- parseMessage (OneBot input → MessageElement[])
* CQ string text ~ 0.31 μs/op
* CQ string mixed ~ 1.22 μs/op
* segment array text ~ 0.32 μs/op
* segment array 5 ~ 1.07 μs/op
- messageStore SQLite writes:
* storeMeta ~17.0 μs/op ( 59K ops/sec) ⚠
* storeEvent (~250B JSON) ~23.2 μs/op ( 43K ops/sec) ⚠
- messageStore reads:
* findMeta hit ~ 5.4 μs/op (186K ops/sec)
* findMeta miss ~ 3.7 μs/op (272K ops/sec)
* findEvent hit ~ 3.8 μs/op (262K ops/sec)
Headline finding
----------------
Pure decode/encode is already cheap (sub-μs). The real cost per
sent message is the SQLite write pair (storeMeta + storeEvent =
~40 μs). On receive, group_message itself does NOT trigger
identity writes; only join/leave/admin events do — but those go
through 4-6 db.prepare(sql).run/get(...) calls each, all re-parsing
SQL on every invocation (the canonical SQLite anti-pattern).
The follow-up commits in this series target:
1. Replace `node:sqlite` (experimental, API may shift) with
`better-sqlite3` (stable, sync, ~30-50% faster wire) + cache
prepared statements as readonly fields. Single commit, both
wins together — better-sqlite3's API encourages stmt caching
by design. Expected: storeMeta / storeEvent drop to ~3-5 μs,
group_member_join drops to ~3-5 μs.
2. Skip rememberGroupMemberIdentity DB write when the in-memory
uidByUin / uinByUid cache already has the mapping unchanged.
3. Drop the `Buffer.from(pkt.body)` allocation in msg-push
context (pkt.body is already Uint8Array; proton accepts it).
4. Make `elementsToJson` sync-fast-path when no async resolvers
are wired (saves per-element microtask overhead).
Each commit re-runs these benches and the message should quote the
before/after numbers so the deltas are auditable.
Note: tsx is NOT involved — vitest's bench mode owns the run, and
the proton vite plugin is configured in each `vitest.config.ts`.
Absolute numbers reflect the production code path.
DrSmoothl
f3f5d27c65