Two benchmark-integrity bugs that could invalidate sync numbers while a run
reports success:
- CorpusServerMain keyed its serve-existing decision only on port + row
count, so re-running a port with a different corpus/maxCount, or after a
load crashed mid-way, silently served a stale/partial DB. Gate reuse on a
completion sentinel keyed on corpus identity (path + byte length) and
maxCount, written only after a full load; on any mismatch the prior DB is
dropped and reloaded.
- CorpusDownloader treated CLOSED identically to EOSE, so a relay ending a
sub early (rate-limit/policy) after sending a partial page advanced the
cursor past the unsent tail — silent corpus loss. Treat CLOSED as a soft
failure (null → reconnect and retry the same cursor; the id dedup set
absorbs the re-fetch), distinct from EOSE which means the page is
complete.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012EZeWww5TJnzBZKPoc6mvU
Report, per sync pair: fetch coverage (did the peer REQ return every needed id,
without duplicates?), publish acks (accepted/rejected/unacked per target), and —
for any event a relay ends up missing — whether it was in the delta batch
delivered to that relay. In-batch-but-absent isolates ingest loss
(ack-without-persist) from a fetch/read gap, turning a vague "missing N" into a
pinned layer.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012EZeWww5TJnzBZKPoc6mvU
When the identical-set reconcile shows a side missing events the reference set
has, log those events grouped by kind (plus a few samples with their tag keys)
instead of leaving a bare "did not converge". The harness already holds every
event by id, so it can resolve exactly what a relay dropped — turning a vague
verdict into an actionable ingest-semantics diagnosis (NIP-09 deletion / NIP-40
expiration / validation).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012EZeWww5TJnzBZKPoc6mvU
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
A NIP-77 server session rebuilt its reconciliation structure from
scratch on every NEG-OPEN: full id+created_at scan, per-entry hex
decode into a fresh StorageVector, O(n log n) seal. That cost grows
with the corpus and is paid even when nothing changed — the exact
shape of a periodic mirror's heartbeat, where N peers reconcile the
same broad filter over and over. relayBench measured 342 ms per
identical-set reconcile at 50k events vs strfry's 26 ms off its
always-current LMDB tree.
Reconciliation only *reads* the sealed storage, so one instance can
back any number of concurrent sessions:
- NegentropyServerSession now accepts a pre-sealed IStorage (the
List<IdAndTime> constructor remains and delegates).
- SessionBackend.sealedNegentropyStorage() builds + seals (null when
the set exceeds maxSyncEvents); LiveEventStore overrides it with a
single-slot cache keyed by (filter set, write generation) plus a 30s
TTL. The generation bumps on every accepted ingest; the TTL bounds
staleness from delete paths the counter can't see (expiration
sweeps, admin purges) — negentropy snapshots are point-in-time sets,
so seconds of staleness only means a peer briefly re-offers ids.
- NegSessionRegistry.open consumes the shared sealed storage;
over-cap NEG-ERR behavior unchanged (strfry parity).
relayBench gains a 'heartbeat' measurement — the identical-set
reconcile repeated immediately with no writes in between. At 50k
events: geode 342 ms -> 27.8 ms vs strfry 21.1 ms (near parity; was
13x). Cold reconciles (first open after a write) are unchanged.
Also fixes the GeodeVsStrfryNegentropySyncTest fixture to write
'nofiles = 0' so the opt-in interop test can boot strfry inside
containers with a low RLIMIT_NOFILE hard cap; the interop suite passes
against strfry v1-b80cda3 with the cache in place.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NeoCvXnTxsKzqurkmjdC46
Reviewed strfry's LMDB indices (golpe.yaml) against geode's SQLite set:
the two are nearly isomorphic — time, id, kind+time, author+kind+time,
tag+time, plus conditional deletion/expiration/replaceable entries
(geode's are partial indexes, so ordinary events don't pay for them).
Nothing to drop. One real hole: strfry maintains a plain
pubkey(+created_at) index and geode had none, so an authors-only filter
(no kinds) — archive pulls, account-migration tools, 'everything by
these pubkeys' — degraded to a full walk of the time index. EXPLAIN
confirmed: SCAN query_by_created_at_id.
- quartz: IndexingStrategy.indexEventsByPubkeyAlone (default false —
clients query their supported kinds and can skip it) gates a new
query_by_pubkey_created index; DATABASE_VERSION 2→3 with an
idempotent migration that backfills it for opted-in strategies.
- geode: relayIndexingStrategy turns it on.
- relayBench: new 'author-archive' scenario — every kind by 3 *quiet*
pubkeys. Quiet is the point: prolific authors are dense in the time
index and a scan finds them quickly, which is why the suite never
caught this; sparse authors force the full walk.
Measured (50k corpus): author-archive EOSE p50 42.8 ms -> 3.6 ms (12x,
and the old path grows linearly with table size); ingest 5,337 -> 5,156
events/s (~3%, the one extra B-tree per event). strfry reference on the
same scenario: 0.52 ms.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NeoCvXnTxsKzqurkmjdC46
strfry implements no NIP-50 at all, so geode's default setup pays FTS
tokenization on every searchable event for a feature the other side of a
benchmark isn't providing. The new switch removes it cleanly:
- geode: --no-search CLI flag and [options].full_text_search TOML key
(default true). When off, the store skips the FTS index entirely,
NIP-11 stops advertising 50 (an explicit [info] nips list stays
operator-authoritative), and search filters match nothing via the
existing QueryBuilder guard. RelayIndexingStrategy becomes
relayIndexingStrategy(fullTextSearch) with the stock val kept.
- relayBench: --geode-no-search runs the geode entry with the flag (relay
named geode-nosearch in reports); README documents the apples-to-apples
rationale and a recipe for benchmarking both geode flavors side by side.
Alternating A/B on the 10k corpus: 4,767/4,792 ev/s without search vs
4,111/3,977 with — NIP-50 costs ~18% of ingest throughput at current
write-path speed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NeoCvXnTxsKzqurkmjdC46
Main's giant-REQ fix (18bd3600) replaced query()'s copy-on-add immutable
dedupe set with a spin-locked HashSet, but the rebase left queryRaw —
now the default REQ path — on the old pattern, which would have
reintroduced the O(n²) crawl for large replays. Both paths now share
the mutable-set-under-spinlock shape.
relayBench's sync driver moves to NegentropySession.fromEvents(),
following the session's new List<IdAndTime> constructor.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NeoCvXnTxsKzqurkmjdC46
New :relayBench module that boots relay implementations as real external
processes under equivalent setups (persistent storage, sig verification on,
no auth) and compares them on the same corpus:
- Ingest: receipt->queryable-by-REQ latency (publish on one connection,
hammer-poll REQ{ids} on another), OK-ack latency, and pipelined corpus
replay throughput over N connections.
- Queries: client-realistic filters derived from the corpus (home feed,
thread, notifications, profiles, hashtag, by-ids, ...), time-to-EOSE
percentiles plus aggregate events/sec under concurrency; result-set
counts are cross-checked between relays and mismatches flagged.
- NIP-77 negentropy sync between every relay pair: 80%/80% slices with 60%
overlap, reconcile timing/rounds/wire-bytes per server side, delta
transfer, steady-state identical-set reconcile, convergence verified.
- Storage footprint after full ingest.
Corpora (all cached as NDJSON + manifest with a sha256 id fingerprint so
results are comparable across runs and machines):
- synthetic (default): deterministic to the byte — seeded keys, fixed
timestamps, seed-derived BIP-340 aux nonces — with a realistic social
shape (zipf authors, threads, reactions, reposts, zap request/receipt
pairs, hashtags);
- the checked-in real dump (quartz test fixture, ~31k unique 2024 events);
- external dumps (NDJSON or JSON array, gzip sniffed by magic bytes),
e.g. the 2.1M contact-list archive, with --max-event-bytes/--max-tags
raising both the corpus filter and the strfry config together;
- live download from public relays.
Every source runs through the same preparation: dedup, drop unsigned/
ephemeral/kind-5, enforce relay ingest caps, parallel Schnorr verify,
chronological sort.
relayBench/run.sh is the one-command entry point: builds geode + harness,
resolves strfry (STRFRY_BIN, PATH, or source build into .cache), runs the
suite and renders an ANSI report with per-metric bars and winners, plus
report.md and results.json under relayBench/results/<timestamp>/.
The harness client disables Nagle (TCP_NODELAY): with the JDK default, a
REQ following the previous round's CLOSE stalls a full delayed-ACK
interval and every latency floors at ~44 ms against both geode and strfry
(verified: ~0.3 ms with it off).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NeoCvXnTxsKzqurkmjdC46
Claude