mirror of
https://github.com/vitorpamplona/amethyst.git
synced 2026-10-30 15:48:36 +00:00
Claude
fd6662ca30
Found while attributing the small-REQ wire floor (backlog item 6, latency half): geode's new WireReqFloorBenchmark measured a flat 43.7 ms per REQ round trip that survived every server-side change — store configs, dispatchers, the pump — and then vanished when the round's preceding CLOSE was dropped. Root cause is client-side: OkHttp does not set TCP_NODELAY, relays never answer a CLOSE (NIP-01), so its bytes sit unACKed for the peer's ~40 ms delayed-ACK window and Nagle holds the next REQ behind them. CLOSE-then-REQ is a Nostr client's hottest pattern — every feed/filter switch. relayBench's harness client already shipped a no-delay socket factory (which is why benchmark numbers never showed the stall) but the production clients did not. New TcpNoDelaySocketFactory (quartz jvmAndroid, next to BasicOkHttpWebSocket) is now used by the Android relay pool factory, the Desktop relay client, amy's relay connections, and geode's mirror worker. Direct connections only — SOCKS/Tor paths are untouched. With the factory, the benchmark puts geode's ~21-row REQ at ~1.25 ms on the wire (matching relayBench): ~0.6 ms Ktor CIO+OkHttp loopback floor, ~0.5 ms per-REQ server work (already investigated). Per-frame burst cost measured negligible and the pump adds ~nothing, so the send-path latency angle of backlog item 6 is closed as not-a-problem; its ingest-CPU share remains a separate throughput question. Findings recorded in quartz/plans/2026-07-04-small-req-floor.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TtDNpayEYvJH7QuPswND3A
78 lines
3.6 KiB
Markdown
78 lines
3.6 KiB
Markdown
# Small-REQ dispatch floor — investigated, inline fast path reverted
|
||
|
||
**Status: closed (negative result recorded).** Backlog item 2 of the
|
||
relay performance campaign.
|
||
|
||
## The gap
|
||
|
||
relayBench at 50k events: geode WINS most 500-event query scenarios
|
||
(hashtag 5.4 vs 8.4 ms, recent-window 4.1 vs 8.5) but loses ~2.5× on
|
||
small results — author-archive (19 events) 1.2–1.7 ms vs strfry's
|
||
~0.5–0.6, thread (23) likewise — and the @8conn throughput inverts
|
||
(strfry 2–3× geode). With ~20-row responses, throughput ≈ 1/latency:
|
||
there is a fixed per-REQ floor.
|
||
|
||
## Decomposition (SmallReqFloorBenchmark, kept in jvmTest)
|
||
|
||
In-process at 50k events, ~21 rows/REQ, medians of 400:
|
||
|
||
| stage | ms |
|
||
|---|---:|
|
||
| A raw store query (SQL + row decode) | 0.18 |
|
||
| B + live machinery (FilterIndex reg/unreg, dedupe set) | 0.36 |
|
||
| C + session dispatch (parse, launch, frames) | 0.60 |
|
||
|
||
Note: in-memory DBs have no reader pool (`useReader` falls back to the
|
||
writer mutex), so absolute numbers are conservative vs the file-DB
|
||
bench setup.
|
||
|
||
## What was tried and why it was reverted
|
||
|
||
An inline fast path (`SessionBackend.queryRawInline`): REQs with
|
||
provably bounded replays (limit or ids-count summing ≤ 512) ran their
|
||
stored replay on the receive coroutine and kept only a live-tail
|
||
handle — no per-REQ `launch`, no Job, no dispatcher handoffs. It cut
|
||
in-process time-to-EOSE ~17% (0.60 → 0.50 ms) with full wire-behavior
|
||
parity (stored→EOSE order, live tail, CLOSE, same-subId replacement).
|
||
|
||
Three relayBench runs (baseline, cap-256 [path not engaged — bench
|
||
filters carry limit=500 or no limit], cap-512 [engaged for
|
||
author-archive/by-ids/500-limit feeds]) showed **no movement outside
|
||
the container drift band** — strfry's own numbers drifted ±30% run to
|
||
run, and inline-eligible scenarios moved the same as ineligible ones.
|
||
Reverted per the keep-only-winners rule.
|
||
|
||
## Where the floor actually is (corrected after WireReqFloorBenchmark)
|
||
|
||
The follow-up wire benchmark (geode's `WireReqFloorBenchmark`, Ktor CIO
|
||
+ OkHttp on loopback) attributed the full path:
|
||
|
||
| leg | ms |
|
||
|---|---:|
|
||
| bare Ktor CIO echo round trip (1 or 22 frames — same) | ~0.9–1.0 |
|
||
| geode NOTICE (inline, full pump + Ktor send) | ~0.5 |
|
||
| geode empty REQ (launch + SQL, 0 rows) | ~0.6–0.8 |
|
||
| geode ~21-row REQ, wire | ~1.25 (= relayBench's number) |
|
||
|
||
**geode's websocket send path has no latency problem** — per-frame burst
|
||
cost is negligible (echo-22 ≈ echo-1), the pump adds ~nothing (NOTICE ≈
|
||
0.5 ms), and the residual vs strfry (~0.5 ms/REQ) is the per-REQ server
|
||
work already investigated above. Frame batching / permessage-deflate
|
||
would not move these numbers. Backlog item 6's remaining open angle is
|
||
the INGEST-side CPU share (13–25% in the JFR profile) — a throughput
|
||
question, not this latency one.
|
||
|
||
**The real find was client-side.** The first wire measurements showed a
|
||
flat 43.7 ms per REQ — which turned out to be the benchmark's own OkHttp
|
||
client: OkHttp does not set TCP_NODELAY, and the CLOSE-then-REQ pattern
|
||
(every feed/filter switch!) nagles the REQ behind the unACKed CLOSE
|
||
(relays never answer CLOSE) for the ~40 ms delayed-ACK window.
|
||
relayBench's harness client already carried a no-delay socket factory —
|
||
which is why bench numbers never showed it — but the production clients
|
||
(Android relay pool, Desktop, amy, geode's mirror) did not. Fixed by
|
||
`TcpNoDelaySocketFactory` (quartz jvmAndroid), now used by all of them.
|
||
|
||
**Do not retry** relay-side latency work for the small-REQ gap; the
|
||
addressable remainder is the ~0.4 ms of per-REQ dispatch machinery this
|
||
doc's revert already covers, and it does not show on the wire.
|