coordinator:run plans the articles after each group's last range every
TAIL_EVERY_SECONDS (120): one LIST ACTIVE per name prefix gives every
group's newest article, and new articles become ranges at priority -1,
leased and prefetched before any history. A group gets a tail range once
it has TAIL_MIN_ARTICLES (100k, a chunk) new or its last planning is
TAIL_MAX_DELAY_MINUTES (15) ago, so quiet groups don't make a range every
round. History (planAll, hourly) is now only for groups not planned yet.
ranges:plan --tail plans one round.
Found on the way: with XFEATURE COMPRESS GZIP on, Eweka compresses LIST
ACTIVE's reply too ("215 NewsGroups Follow [COMPRESS=GZIP]"), and
listActive waited for a plain end that never came (groups:add --top too).
It reads the compressed reply now; the fake server in the test does both.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Connections fetch chunks round-robin from several starting points of one
group at once, so each era sees the same server load; per era: seconds
per command (mean, median), articles/s per connection, bytes per article.
On a.b.boneless (2 connections, 10 min): 2008 24.9k, 2011 28.0k, 2019
22.6k, 2026 17.9k articles/s per connection; medians 2.3-3.0 s, means set
by a tail of slow commands (up to 41 s).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
libdeflate inflates a chunk whole, in half the time of zlib-rs (1.09 ->
0.51 s per boneless range); memchr finds line ends and tabs with SIMD.
The per-article timers are gone. Boneless 2026 ~700k headers/s, ath 2013
~920k; all three captures still identical to the TypeScript.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each with its regex/chrono version as the reference in tests; check.ts
still finds all three captures identical. Boneless 2026 ~500k headers/s,
ath 2013 ~650k.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The suffix and remainder regexes were most of a new file's description
(333k on boneless 2026): now plain byte matching, with the regex versions
kept as the reference in a test of tricky names and subjects. mimalloc
for the many small strings. Boneless 2026: 360k -> 495k headers/s;
check.ts still finds the outputs identical on all three captures.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The RangeIndexer path of edge:bench (inflate, scan, owner rule, segments,
file aggregation, cut-off, rules, release identity, chunk encoding v2,
probe candidates) as a Rust binary over a captured range, with JS
semantics where they matter (\s, trim, \b, parseInt, lastIndexOf, UTF-16
lengths). check.ts runs the TypeScript RangeIndexer on the same capture
and compares every output: identical on all three captures (releases and
their decoded payloads, samples, candidates, rule counts, fragments),
after mapping Date headers ending in UTC to GMT (chrono follows RFC
5322; Date.parse takes UTC, a quarter of a.b.boneless). Compressed
payloads differ by <0.1% in size (zstd framing), not in content.
First run on boneless 2026: 360k headers/s and 228 MB peak (TypeScript:
~217k/s, ~0.7-0.8 GB).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
For workers on a shared line (poweredge-kantoor: 160 Mbit/s): the range
stream is read at most at the cap, TCP holds the coordinator back. The
cap is the node's, split over its processes (also EDGE_MAX_MBIT).
REBUILD.md: where workers run (connections set the speed; remote workers
take CPU off homelab).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On homelab (1 Gbit, 20 connections, ~630k articles/s) the coordinator
used 1.3-1.7 cores, mostly GC of the text it inflated only to know a
compressed reply was complete, plus a copy of every reply. A raw reply
is now complete when it ends in ".\r\n" (no copy): 0.4-0.5 cores at
the same rate.
".\r\n" can also end a network read inside the compressed data (~1 in
16M reads, a few dozen times over the whole target). Such a chunk
doesn't inflate on the worker: it reports it (POST ranges/:id/corrupt),
the spooled chunk is removed and the range goes back to the queue; the
prefetcher fetches it again. The rest of that reply breaks the
connection's next command, so the connection is replaced.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The coordinator's event loop was 16.5% busy at ~10 MB/s of compressed
XOVER (socket reads, buffers, GC, spool writes; nothing stands out), so
one process tops out around 60 MB/s: less than 50 connections bring
(~1M articles/s, ~70-86 MB/s). A shard prefetches the ranges with
id % n = i - 1 with its own NNTP_CONNECTIONS; only shard 1 plans, merges
and settles. Two shards of 5 connections fetched what one of 10 did.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It checked only after a data frame, so a worker waiting for its range's
spool (keep-alive frames only) never stopped on SIGTERM.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It waited for every chunk in flight, and a connection the server held
back (after killed coordinators left sessions open) waits up to the
10-minute reply timeout: it never seemed to stop. It now waits at most
15 s, then closes all connections, busy ones too; the greeting and login
time out after 30 s. And run() ends with terminate(), as edge:node does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The coordinator logs a line a minute: chunks fetched, articles/s,
seconds a chunk, how busy its connections were. It showed the
workstation's runs are bound by its 100 Mbit line now (10 connections
99% busy, ~89 Mbit/s: compressed XOVER is ~69 B an article on recent
boneless), not by the code.
The check that a compressed reply is complete (inflating it) ran on the
coordinator's event loop, ~60 ms a chunk; now on a thread-pool thread.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A lease was only granted for a range with spooled chunks, and the
prefetcher kept ranges ahead per active lease: once the spool ran dry a
worker got none and slept 30 s, fewer ranges were leased, fewer
prefetched. Three workers held one lease between them (~50k headers/s).
A lease now takes a spooled range first, else the next planned one
(leased ranges are fetched first; the stream waits).
crypto.hash instead of createHash().update().digest() for file and
release keys: same digests, ~2.5x faster (~4.6% of a worker's time).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
decodeFrames joined each network piece (~64 KB) onto the frame so far:
a chunk of MBs copied over and over, 5.4% of a worker's time in an
end-to-end profile (the bench reads files, so it never showed). Pieces
are now joined once a frame is complete. A batch post's gzip (1.6%) runs
on a thread-pool thread.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The coordinator ran the prefetch tick in the same loop as the merger
(up to 30 s a minute): its connections went idle once their targets
were fetched, and the first leases after a start found nothing spooled.
The prefetcher now ticks in a loop of its own.
The prefetcher took the first planned ranges with nothing spooled, even
ones with nothing left to fetch (next_id past the end): two such ranges
(edge:bench's, set back to planned) kept it from fetching anything. It
and the lease query now skip those, and edge:bench marks its range done.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The merger's neighbour query matched "end_id = ? OR start_id = ?", which
no index serves: it read every range of the group per file, 5.5 s a
round and most of Postgres's time in an end-to-end run. Now the range
holding the article before and after, by (group_id, start_id): 0.08 ms.
The prefetcher's leased and started ranges (every 2 s) get partial
indexes: 15-25 ms -> <0.1 ms with 145k planned ranges.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Waiting files counted as work, so the merge loop re-read them for its
whole 30 s budget every minute: Postgres used ~a core even without
workers. Now ~4% at rest, and storing a boneless range costs Postgres
~0.7–1.9 s of CPU (it was 33 s with the loop running). Chunk payloads
are stored EXTERNAL (zstd already; no pglz attempt).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A slice of a chunk's text keeps the whole text alive (tens of MB) for as
long as the file it belongs to stays open: message-ids, file details and
cache keys are now copies. edge:bench on boneless: peak RSS 1,137 → ~660
MB, external memory 570 → 40–75 MB, the same speed (~190k headers/s);
the same results on 800k real lines. edge:bench prints the breakdown.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Hourly in coordinator:run: a release unchanged for 6 hours is settled,
complete when every file its [i/N] counter announced was found, unknown
without a counter (phase 2 reads par2 sets for those). Partial index on
unsettled releases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
lib/nzb (buildNzb, parseNzb) from the current app; NzbService builds a
release's NZB from its chunks; GET /api/v1/releases?q= (every word of the
name) and GET /api/v1/releases/:id/nzb behind the api guard. A test stores
a batch, finds the release and reads its NZB back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nyuu's 12 capital+small letter pairs in 15 bytes (10 bits a pair), its
timestamp as a delta within the file, the suffix once per file;
ngPost-style 32 hex digits in 16 bytes; other ids as zstd text as before.
Recent boneless: 25.2 → 20.3 B per segment (and 25.0 → 19.5 on the next
range); a 2013 ath range stays at 17.9 (JBinUp ids are text). Every chunk
of three real ranges decodes to itself (edge:bench --verify). Encoding is
no slower than v1 (less text for zstd). Version 1 chunks stay readable.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The neighbour check was three queries per file (~2,000 files/minute); now
one per round of 20k files, with a cursor so waiting files don't take
every round's places, and rounds until nothing is left or 30 s. On the
dev backlog (old-style fragments): ~18k files/minute in the 30 s budget;
at full throughput the new fragment rule leaves ~1,200 files/minute.
A group that fails GROUP (gone from the server) no longer stops planning
the others; edge:bench's groups are inactive.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Files cut by a range's edge are put together once the ranges around them
are done (coordinator:run, every minute): a complete file is judged by the
rules and written as a release chunk with probe candidates; one still
missing parts while both neighbouring ranges are done is dropped and
counted (incomplete/edge). A neighbour not planned yet counts as not done,
except before the group's oldest article. BatchStore.writeReleases is the
shared write path.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured on two neighbouring boneless ranges (merging their fragments):
- 99.8% of fragment files were single articles of a random-named
'file' of >100 parts: a random subject per article, which never
completes from headers. They now count as per-article.
- A file is kept as a fragment when its first part seen is above 1 and
comes within the start window (the idle window: uploads in busy groups
spread over minutes, and a 50k-article edge zone is ~30 s of boneless),
or when it was still open at the range's end with later parts missing.
Anything else missing parts is incomplete and only counted.
Per range: 70,308 → 130 fragment files; merged, 47 of the 48 files met
complete (57,558 segments, ~8% of the range's kept segments).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Eweka takes ~1.5–2 s per XOVER command whatever its size. One connection
on recent a.b.boneless: 9–15k lines/s in chunks of 25k, ~20k in 50k,
~30k in 100k–200k. Four connections at once: 15k each at 50k, 22k each
at 100k (the same as one connection alone at that moment): connections
don't slow each other. CHUNK_ARTICLES defaults to 100k; edge:node posts a
batch every --batch-articles (500k) instead of every N chunks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
XoverClient joined all received pieces into one buffer on every arrival
(quadratic for a 2.6 MB reply), and turned the inflated check into text
the raw path doesn't use. The coordinator under load (3 workers, 4
connections): busy 40% → 13% of a core, Buffer.concat gone from the
profile. (master's relay uses the same client.)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On the 4-core workstation (shared with coordinator, web server and
database): 3 processes indexed 6 boneless ranges (12M headers) in 37 s,
~324k headers/s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
End to end (workstation, recent a.b.boneless ranges from the spool): a 2M
range in 10–12 s, 166k–202k headers/s per worker process (edge:bench on
one core: 215k).
- Prefetcher: spooled_at was set against a total counted from the range's
resume point, which shrinks as the worker advances: the last chunks were
never fetched and the worker waited. Now counted against the whole
range, only new chunks counted, and leased ranges are always checked
against the files. Planned ranges ahead are bounded (it had spooled ever
more); the spool size is measured every 30 s (a stat per file).
- The stream sends an empty frame every 20 s while waiting for a chunk
(Node's fetch gave up on a quiet body after 5 minutes).
- A range taken over midway: the first batch starts at the resume point
(it started at the range's start and was refused).
- edge:node inflates the next chunk on the thread pool while indexing the
current one, and doesn't wait for a post before indexing on (posts stay
in order; a failed post stops the range).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
edge:bench (one core): boneless 193k → 215k headers/s, ath 303k → 343k.
Same results on 800k real lines.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The worker reads the XOVER reply's text directly: fields by position, the
owner from the Xref without arrays (crossposts leave before any subject
parsing), a subject parsed once per file (cached without its segment
counter), a file's details as one shared object, dates as numbers.
Garbage collection had been ~22% of the time.
Same results as parseOverviews + segmentFromHeader on 800k real lines.
edge:bench (one core): boneless 137k → 193k headers/s (CPU 21 → 12.8 s),
ath (2013) 143k → 303k; peak RSS 1.15 / 0.87 GB.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The file key and release identity (two SHA-1 hashes and the release rules)
were ~25% of the edge worker's time. edge:bench: boneless 100k → 137k
headers/s on one core, ath (2013) 70k → 143k.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
End to end on the workstation (coordinator:run, the API, edge:node, real
NNTP): a.b.e-book.french and a.b.cd.image.playstation2.dvdiso ranges.
- Prefetcher: every connection takes the next missing chunk (leased
ranges first, in order); one connection per range (~26k lines/s) kept
the worker waiting
- executable rule: releases of bare executables (malware floods named
after films and software) are dropped; they were 92% of the catalog rows
of the test ranges, 21% of the segments
- fragments per file (details once) and gzipped batch bodies; the JSON
limit (1 MB by default) raised to 256 MB, empty strings kept (a rule
count's key is '')
- release_chunks.batch_from is a bigint (boneless article numbers)
- tests get their own database; edge:bench --store times BatchStore:
1.3 s for a boneless range (1.24 ms per release chunk)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- groups:add / ranges:plan / nodes:create / coordinator:run commands
- the node API: lease, the range's chunks as one framed stream from the
spool, batches, heartbeat, finish, release (node guard, token per node)
- edge:node: needs COORDINATOR_URL and NODE_TOKEN only (placeholders for
the rest), posts a batch every 10 chunks, gives its range back on SIGTERM
- releases are gathered per batch: 1.04 chunk rows per release at 10
chunks per batch (1.35 when built per XOVER chunk)
- partitions are checked per batch instead of remembered
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A batch is stored in one transaction and only at its range's resume
point, so a retried batch can't count twice. Garbage is judged before
per-article: a dropped tool isn't worth reading bodies for either.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Overview chunks in, per batch out: releases as encoded chunks, probe
candidates, counts per rule (crossposts per owner group, obfuscated,
garbage, per-article, incomplete), a sample of what was dropped, and
fragments near the range's edges.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
lib/nntp (XOVER client, Xref, GROUP), lib/headers (subjects, file keys,
segments from overview lines; the yEnc path stays behind for phase 2),
lib/release (identity, random names, now in lib), lib/files (aggregator,
volume names), lib/rules (posting tools, owner by priority, release
verdicts), lib/chunks (release chunk codec). Their unit tests came along;
new ones for codec, rules and volumes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The rebuild branch starts over (REBUILD.md): the API starter kit, with
Postgres instead of SQLite (its own database, DB_SEARCH_PATH for a
shared one in development) and the #lib alias. Phase 1's code is
imported from the current app module by module.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
