Files
amethyst/relayBench/README.md
T
Claude 8b29c06c13 feat(relayBench): deep resumable --download for million-event corpora
The corpus downloader previously sampled ~100k events max (40-page cap
per kind bucket) and held everything in memory with no failure recovery.
Rework it for full-depth timeline pulls:

- page the latest events newest-first with an inclusive until cursor
  (id-dedup absorbs the same-second overlap) instead of kind buckets
- no page cap: keep paging until the --limit target is met
- stream every unique event to an on-disk NDJSON spill instead of RAM
- checkpoint the pagination cursor per relay; interrupted downloads
  resume where they left off
- reconnect with exponential backoff on socket drops/timeouts
- filter deterministically droppable events (kind-5 deletions are ~40%
  of a live firehose, ephemerals, oversize) at page time so they never
  count toward the download goal

Verified with a 1M-event pull from relay.damus.io (~2.1 GB raw,
4100 pages, ~22 min through a proxy) feeding a full geode vs strfry
run; corpus prepared to exactly 1,000,000 events, fingerprint
141e746599d901f5.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012EZeWww5TJnzBZKPoc6mvU
2026-07-04 03:52:54 +00:00

133 lines
7.1 KiB
Markdown

# relayBench
Head-to-head benchmark for Nostr relay implementations. Boots each relay as a
real external process on loopback under an equivalent setup — persistent
storage, signature verification on, no auth, stock limits — replays the same
event corpus into each, and renders a side-by-side report.
```bash
./relayBench/run.sh # geode vs strfry, 10k-event synthetic corpus
./relayBench/run.sh --quick # 2k-event smoke run
./relayBench/run.sh --real # replay the checked-in real-event dump (2024, ~30k events)
```
`run.sh` builds `:geode:installDist` and the harness, and resolves strfry from
`$STRFRY_BIN`, the `PATH`, or by building it from source into
`relayBench/.cache/` (first run only). `SKIP_STRFRY=1` / `SKIP_GEODE=1` skip a
side.
## What is measured
**Ingest — receipt ➜ queryable.** One connection publishes an event; from the
same instant a second connection hammer-polls `REQ {"ids":[id]}` until the
event comes back. This measures exactly "how long after the relay receives an
event can a REQ return it", which is not the same thing as the OK ack — the
report also shows OK latency and what fraction of events were already
queryable when their OK arrived.
**Ingest — throughput.** The corpus is replayed over N connections (default 4)
with a bounded number of unacked EVENTs in flight, wall-clocked from first
send to last OK. Accepted/rejected counts come from the OKs.
**Queries.** Filters modeled on what real clients send, derived from the
corpus itself so they hit meaningful data: global feed, profile hydration,
home feed (150 follows), hottest thread, notifications for the most-mentioned
pubkey, hashtag feed, 100-id batch fetch, recent time window. Each runs
warmup + measured rounds on one connection (time-to-first-event / time-to-EOSE
percentiles) and then from 8 connections at once (aggregate events/second).
All filters stay inside strfry's default limits, and the number of events each
relay returns is cross-checked — a ⚠ in the report means the relays disagree
about the result set, which invalidates the speed comparison for that row.
**NIP-77 negentropy sync — every pair of relays.** Both sides get an 80% slice
of the corpus (60% overlap); the harness plays the `strfry sync` role with one
side's dataset as its local set and measures: initial reconciliation against
each relay as server (time, NEG-MSG rounds, wire bytes), the delta transfer to
convergence, and the steady-state reconcile of identical sets. Convergence is
verified, so this doubles as an interop test.
**Storage.** On-disk footprint after full ingest (LMDB vs SQLite vs whatever).
## Corpora
The corpus is the controlled variable: every relay sees the same events in
the same order, and reports carry a `fingerprint` (sha256 over event ids) so
two runs are comparable only when fingerprints match.
| source | flag | notes |
|---|---|---|
| **synthetic** (default) | `--events N --seed S` | Deterministic to the byte: seeded keys, fixed timestamps, seed-derived BIP-340 nonces. Same spec ⇒ identical NDJSON on any machine. Zipf-popularity authors, threads, reactions, reposts, zap pairs, hashtags. |
| **real dump** | `--real` | The `quartz` test fixture `nostr_vitor_startup_data.json.gz` — ~31k unique real events from 2024 with a rich kind mix (notes, chats, DMs, zaps, reports, communities). |
| **contact lists** | `--corpus contact-lists.gz --limit 100000 --max-event-bytes 1048576 --max-tags 20000` | 2.1M real kind-3 contact lists (heavy events, ~1.3 kB avg, up to 100+ kB). Grab it with `pip install gdown && gdown 1yyC93xY9sDsEsa351ZAMhtAXwBUh3LYT`. Raising the size/tag caps reconfigures strfry to match, so both relays still accept the full stream. |
| **any dump** | `--corpus FILE` | NDJSON or a single JSON array, gzipped or plain (sniffed by magic bytes). |
| **fresh download** | `--download [urls]` | Pages the latest events out of public relays newest-first (damus/nos.lol/primal by default), `--limit N` deep — built for million-event pulls: streams to an on-disk spill, checkpoints the pagination cursor, resumes interrupted downloads, reconnects on drops. E.g. `--download wss://relay.damus.io --limit 1000000`. |
Every source goes through the same preparation: dedup by id, drop unsigned
events (NIP-17 rumors), kind-5 deletions and ephemerals (order-dependent or
unqueryable — they would make relays disagree for reasons unrelated to
performance), drop events over the size/tag caps, verify every Schnorr
signature in parallel, sort chronologically. The prepared corpus is cached in
`relayBench/.corpus-cache/` as NDJSON next to a `manifest.json` with the
fingerprint and kind histogram — that file pair is a shareable, citable
benchmark artifact.
> **Why this corpus matters:** there is no de-facto community benchmark corpus
> today. Existing relay benchmarks (rnostr's and privkeyio's nostr-bench,
> mattn's scripts) each synthesize their own events with unspecified
> distributions, so published numbers aren't reproducible corpus-controlled;
> the only shared dataset (Wellorder's early-1m) is frozen in January 2023.
> relayBench's synthetic spec ("seed 1, n=10000, v1" ⇒ byte-identical corpus)
> and manifest/fingerprint convention are designed so other relay authors can
> run the exact same workload and publish comparable numbers.
## Feature parity: NIP-50 search
geode maintains a NIP-50 full-text index by default; strfry has no search
at all. That skews the *ingest* comparison — geode tokenizes every
searchable event into the FTS index (~a quarter of its write cost) for a
feature strfry isn't providing. `--geode-no-search` runs geode with
`--no-search` (no FTS index, NIP-11 stops advertising 50) for the
apples-to-apples write path:
```bash
./relayBench/run.sh --geode-no-search # geode(no NIP-50) vs strfry
```
To see the price of search inline, run both geode flavors side by side:
```bash
GEODE=geode/build/install/geode/bin/geode
./relayBench/run.sh --geode-no-search \
--relay "geode-fts=$GEODE --host 127.0.0.1 --port {port} --db {dir}/geode.sqlite"
```
## Adding another relay
No code needed if the relay can be launched from a command line:
```bash
./relayBench/run.sh --relay 'nostr-rs-relay=/usr/bin/nostr-rs-relay --db {dir} --port {port}'
```
`{port}` and `{dir}` are substituted at launch; the process must listen on
`127.0.0.1:{port}` with persistent storage under `{dir}`, verification on and
no auth. If the relay needs a config file, point the template at a small
wrapper script that writes one (see `StrfryRelay` in
`relays/RelayUnderTest.kt` for the pattern — adding a first-class subclass is
~20 lines).
## Output
The terminal report shows each metric as name / bar / value rows with the
winner starred, followed by a head-to-head summary. Every run also writes:
- `relayBench/results/<timestamp>/report.md` — shareable Markdown
- `relayBench/results/<timestamp>/results.json` — raw numbers for tooling
## Direct harness invocation
`run.sh` is a thin wrapper; the harness itself is
`relayBench/build/install/relaybench/bin/relaybench` — see `--help` for all
options (`--samples`, `--publishers`, `--window`, `--query-rounds`,
`--query-conns`, `--no-sync`, `--out`, `--keep-data`, …).