Range 161820 failed with "doesn't inflate, also fetched plain" though
the chunk was likely the compressed copy again: in flight, or spooled
before the prefetcher's targets were refreshed, when the plain flag was
set. The range now fails only on the third report.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ranges with a chunk or two left take fewer than perRange connections:
with loops/perRange + 2 ranges, 3.2 of 40 loops were idle on average
over 15 minutes (up to 9.5).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Writers lock their group only, so two groups' batches can number the
same identity key at once (a spam post crossposted to 67 groups:
duplicate key ...:65, HTTP 500, the batch retried). Releases are now
inserted with ON CONFLICT (release_key) DO NOTHING, and those skipped
take the next number (up to 20 tries).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
a.b.0day.stuffz sends steadily but slowly (~400 kB a minute, 5-7
minutes per 100k chunk): the fixed deadline (90 s, doubling up to 360 s)
gave its chunks up again and again, and two of its ranges' worker
streams timed out. A stall is a request that got less than
NNTP_STALL_KB_PER_MINUTE (100) in the last minute (the stuck replies
got ~25 kB a minute); then its connection is closed and the chunk
fetched again. Replaces NNTP_CHUNK_DEADLINE_SECONDS.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At 90 s a chunk that always takes longer would be given up forever
(range 155437's worker gave up waiting on its stream). Each deadline a
chunk hits doubles its next one (90, 180, 360 ... up to 16x), reset
once it comes in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With 2 connections per range the prefetcher needs 20+ ranges with
chunks left; fully spooled ranges counted as ready (SPOOL_AHEAD_PER_LEASE)
but take no connection, so on small ranges 30 of 40 connections were
busy (13-15 ranges). Fresh ranges now also fill up to loops/perRange + 2
ranges being fetched.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A worker waited on one chunk of a leased range for an hour (range
146477: five 15-minute stream timeouts, then the range took 63 s; 88
"stream: Peer disconnected" overnight). The silence timeout doesn't
catch a reply that trickles: a chunk still coming in after
NNTP_CHUNK_DEADLINE_SECONDS (300) is given up, its connection closed,
and fetched again. The report names the oldest request in flight (range,
chunk, age, bytes in) and counts deadlines.
Batches failed with "Maximum call stack size exceeded" (8 overnight):
Math.min(...segments) on a fragment with more segments than the stack
takes arguments.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At most PREFETCH_PER_RANGE (2) connections on a range and
PREFETCH_PER_GROUP (4) on a group; SPOOL_AHEAD_PER_LEASE 2 -> 4 so there
are targets enough. The report says how many ranges and groups are in
flight.
Homelab, 40 connections, alternated 12 min each: unlimited (~10 ranges
in flight) 281k and 167k articles/s; spread (~22 ranges) 319k and 296k,
chunks 11.5 s against 13-24 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Once a second: pool connections with a command in flight, loops
fetching/writing/idle; per request GROUP time, XOVER time to the status
line and the transfer, transfers with a silence over 5 s. The old
"connections busy" (fetch time of the chunks finished in the minute)
said 59-75% while all 40 connections were in flight all the time:
dropped.
Homelab: 39-40 of 40 in flight, no idle loops; the status line comes at
once, the whole 8-35 s is the transfer, with data flowing (few gaps
over 5 s): ~3 MB per chunk at 0.1-0.4 MB/s per connection.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A chunk not in after NNTP_HEDGE_AFTER_SECONDS (10) is asked again on one
of NNTP_HEDGE_CONNECTIONS spare connections; the first answer wins and
the other connection is closed (lib/nntp/hedge.ts, tested). Connections
can be retired after NNTP_CONNECTION_MAX_MINUTES (±25%). The prefetch
report counts hedges and slow first requests by connection age.
Homelab, 15 min each: hedging on (40 + 6) 280k articles/s, off 273k.
The second request won about half of its races, but ~25-35% of first
requests were slow (>10 s) at every connection age, so neither helps:
Eweka is slow under our load, not on particular connections. Both
off by default; the age statistics stay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BatchStore.store held the range's row (SELECT … FOR UPDATE) for the
whole batch transaction (15-20 s on the HDD), so the prefetcher's
spooled_chunks updates and the heartbeats of that range waited on it
(0.5 s on average): the prefetcher's NNTP connections sat idle. Now an
advisory lock per range serializes its batches, and the resume point is
checked again in the final, conditional update.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TABLESAMPLE SYSTEM (0.5) read 30 s of random pages on the HDD at 7M
releases, growing with the catalog; a refresh that never finished kept
the header's totals 3.5 hours old (it blocked every later refresh).
Now TABLESAMPLE SYSTEM_ROWS (5000) (tsm_system_rows), every 5 minutes,
and a refresh is given up after 2 minutes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A picture group's 500k articles made batches over the server's 256 MB
body limit: HTTP 413, the range given back and leased again in a loop
(148597: 39 times in 30 minutes). The worker now also posts once
100,000 completed or edge-cut files wait (RangeIndexer.pending).
A stats refresh that failed kept the previous totals silently; it
logs the error now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A picture group's batch has 30-55k release keys; an advisory lock each
ran Postgres out of shared memory (max_locks_per_transaction), and those
batches failed with HTTP 500 (the ranges went back, nothing stored).
A chunk only matches releases of its range's group, so one lock per
group serializes exactly the writers that could both start an upload.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Eweka sends some ranges' compressed XOVER truncated, the same bytes
every time, while the plain reply is whole (a.b.models 3317215-3367214:
1,675,600 bytes that don't inflate, twice; 50,000 lines plain;
tools/false_end.ts). Not a false end of reply. The worker's corrupt
report removed the chunk and the range was fetched again, with the same
bytes: one range was reported 540 times. (Before this morning's change
the inflate check waited for more data until the 10-minute timeout,
like voyager 1's 'chunk never answered' ranges.)
A reported chunk's start goes into ranges.plain_chunks; the prefetcher
fetches those over one connection without XFEATURE COMPRESS GZIP
(NntpPool.usePlain, opened on demand) and deflates them here. A plain
chunk reported again sets the range aside (status failed, error). The
capture (edge:bench --capture) fetches such a chunk plain too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A poster re-uploading under the same name made one release of all
uploads: in a sample, 24% of releases with 2+ files had a file name more
than once (4.3% of files extra copies), 70% of those copies posted a day
or more apart; their NZBs listed each copy.
- File assembly keeps a re-upload's copy apart: a part already in every
open copy, posted more than 3 h from them, starts another copy; a
missing part fills the nearest copy (a re-send of what was missing),
a re-sent part near a copy that has it is ignored as before.
- A key's files in a batch split into uploads where posting pauses for
more than 3 h, counted from the latest article posted so far (a file
that takes hours to post doesn't split it); each upload is judged and
chunked on its own (ReleaseChunk.upload, .postedLast; candidates carry
their upload).
- The catalog matches a chunk to the release of its key whose articles
the range's group owns and whose posting window is within 3 h, else
starts one (key, key:2, …): releases get identity_key, group_id,
last_posted_at. Keys are locked in order (advisory locks), so
concurrent batches neither deadlock nor both start one; the merger
writes in key order too. The same upload multi-posted to other groups
is a release per group.
- The gap: 91% of uploads post within 3 h; 73% of re-uploaded copies
come later (releases:upload-gaps). releases:duplicates measures the
duplicate names.
Both implementations; check.ts compares the windows and uploads too
(identical on all captures; Rust tests mirror the TypeScript cases).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- releases.complete is computed with every batch (files >= the
counter's total) instead of only when a release settles after 6 quiet
hours: the 'complete' filter matched nothing in a fresh catalog.
- A release's file total comes from any counter form (counterTotal:
'(01/10)', 'File 27 of 132', …), not only [i/M]; the post rule's key
keeps the [i/M] total, so keys are unchanged. Both implementations.
- groups.tail: a group's live tail can be paused (its history and owner
priority stay), e.g. to let other groups' history go first.
- Stats: history by articles, and the current pass (history ranges
before those deferred 2000 priorities behind) as its own figure; the
header showed 0.1% of ranges counting 145k deferred ones.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- /api (Newznab: caps, search, tvsearch, movie, music, book, get) for
Prowlarr, Sonarr and Radarr: voyager 1's controller and XML on the new
catalog; readable releases not known to miss files; NEWZNAB_API_KEY.
- /api/v1/releases (q, cat, readable, complete, source, poster, size,
sort, paging), /api/v1/releases/:id with its files (decoded from the
chunks; NzbService.files, shared with the NZB), /api/v1/stats
(estimates, plus history and tail progress), /health. Open, as voyager
1's catalog; the web interface uses it.
- release_category(name): voyager 1's categorizer without the archive
contents, a generated column with a (category, posted_at) index.
- @adonisjs/static for the web interface (next).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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>
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 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>
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>
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>
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>
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>
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>
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>
Every probe rewrote its file's row on the HDD (content_type, archive):
a new row version on a random page and an entry in each of the five
indexes. They now live in file_details on the NVMe; files rows are
inserted once. files:move-details copies the existing values after
the deploy; a later migration drops the old columns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
passwordFromSpot finds 30 more passwords in the stored descriptions (480
→ 510): words between keyword and separator, trailing "!!", the value on
the next line, "wachtwoord is x", "… wachtwoord. Deze is: x", "WW x" as a
whole line.
Listed spots were read from their overview only, so the NZB's own
password and the description were unknown for 2.27M of them.
SpotService.readListing reads the spot's article and NZB without
importing files: details, the NZB password (over a spot-text one), the
spot-text password, placeholder NZBs. worker:spots does a batch when
idle; spot:postprocess works through the history (nzb_read_at).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two probes of one release can be applied at once; when one deletes the
release as garbage, the other's inserts (e.g. a par2 listing) violated
the foreign key and the node got a 500.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Random-named uploads that never get a real name or password were ~90% of
what seeding stores. Two filters:
- before storing (seeding, ingest): random-named uploads of a dropped
signature (posting tool from the Message-ID, first group) or of a
release dropped after its probe aren't stored; 2% of a dropped
signature's releases are still kept and probed. Stored files of dropped
signatures aren't probed.
- after a probe: a random-named release that is still unnamed, without
password, names from files or spot, with all its probes done, and
encrypted, a container or missing its first article, is deleted and
tombstoned in dropped_releases.
Signatures are learned by gc:scheduler from a sample of probed releases
(signature_probes); garbage:learn --seed fills it from existing files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
5.4M waiting inspection jobs took 3.3 GB of Redis and most of its single
thread. Producers now add file ids to inspection_queue; index-scheduler
feeds up to INSPECTION_FEED_MAX (50k) into BullMQ. inspection:from-redis
moves the existing backlog over once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Creating a partition deadlocked with file inserts 11 times in an hour, and
each failure ended the GC sweep before its deletes. The parent is now
locked first (lock_timeout 10 s), four partitions are kept ahead, and an
error is logged while the sweep goes on.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Eweka answers 'no such article' after ~450 ms; a quarter of inspection
requests were for missing articles and each failed job was retried up
to three times. Files get content_type 'missing' instead; homelab also
records it from failure reports of nodes not rebuilt yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SeedService looked up every file of a batch across all files partitions
after inserting it (15% of Postgres' active time); the merge job read
every release on each GC sweep.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Par2 naming needs the fingerprint inspection stores on a binary, and the
GC deleted seeded binaries (6 h unfingerprinted, 48 h stale) long before
the inspection queue reached them. binaries.awaiting_inspection protects
them for up to 7 days; the GC deletes use partial indexes so they don't
read past waiting rows. Seeding no longer creates binaries for files that
aren't inspected (~85% of a.b.ath), which only churned the table.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Half the live leases go to higher-priority ranges, the rest to the others,
so a slow recent gap runs beside older history. With no article
connections the article relay answers 503 instead of waiting forever.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured with 20 relay connections, plain and gzip both gave ~1.5 chunks/s
for 2026 a.b.ath articles; plain used many times the CPU and traffic.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
