The routing table records a row each time a prefix and origin AS gets its first peer (announced) or loses its last (withdrawn), with the time and how many peers carry the prefix just after. The prefix handler writes the rows to a new origin_changes table, apart from routes. Every hour the database maintainer keeps each prefix's rows from the last 72 hours or its newest 500, whichever keeps more. The prefix page lists them newest first, at most 5,000.
At most 100,000 rows wait in memory to be written; past that the oldest are dropped and logged. A stop loses the rows past its last write.
Model: deepseek-v4-pro (implementation); opus-5-5 (commit message)
The table's PeerDown visits every prefix under the write lock, which
takes over a second on a full table. The prefix handler called it
inline, so its queue filled and the streamer dropped updates.
PeerDown now holds the lock for 10,000 prefixes at a time, then lets
the lock go for as long again. The handler runs PeerDown on a separate
goroutine, one peer at a time. The handler holds back the down peer's
later messages until that PeerDown returns. Then the handler applies
them in order, one at a time, and takes new messages in between. It
holds back at most 20,000 messages. Stop waits for all of this before
its last flush.
Model: opus-5-5
HandleMessage added one entry to the batch for every AS of every path,
and waited for the write in progress once the batch held 300,000. In a
burst of the feed its queue then filled, and the streamer dropped
messages.
The batch is now a map from AS number to the earliest and the latest
time of its updates. It holds one entry per AS, so it no longer grows
with the feed's rate. HandleMessage never waits for a write.
asnBatchSize now counts ASes. Stop logs the entries of the batch being
written and of the next batch under "batch_entries". An AS in both
counts twice.
Model: opus-5-5
Five prefix handler tests gave HandleMessage two seconds to apply a full
batch of 25,000 prefixes, which under the race detector on a loaded host
takes longer. A test now waits up to stepTimeout, 15 seconds, for each step
it expects.
The slowest tests filled production limits, such as the peering handler's
500,000 paths. The peering handler now takes its path limit, and the prefix
handler its limit on one write of the table's changes, as they take their
timer interval. Those tests give small limits, and the tests of the maximum
batch size lower it to one batch. This also fixes
#215.
The test helpers bound every wait, Stop included, so a test that hangs in
one fails with its own message.
Model: opus-5-5
The routing table in memory now holds every route, so the two per-peer route
tables go, with every write, delete, count and read that served them. The
prefix handler applies each update to the table, writes the table's changes
every second, and stores each new prefix with its first-seen time. The peer
delete, its retry on the next flush and the updates held for it go too. The
types and errors of the server's answers move to internal/server.
The image's memory settings fit a full table of about 1 GB: GOMEMLIMIT of
2.5 GiB, a 512 MiB page cache for the writing connection, and SQLite heap
limits of 768 MiB and 1 GiB. The README describes the new table, its
database copy and the memory budget.
Model: opus-5-5
The prefix, AS and peer handlers skipped a timed write unless one
interval had passed since the last write ended. The next tick came a
little less than one interval after that, so every other tick was
skipped. Each tick now writes. A full batch's write restarts the timer,
so the next timed write comes one interval after that write began.
Comparing clock readings would not do: two ticks one interval apart can
read a hair under it. flushBatch no longer reports whether it wrote a
batch, since nothing reads that. A new test runs the loop on a fake
clock and fails on the old code.
Model: opus-5-5
A write can run past the stop deadline while a full batch's wake-up or a
tick waits beside Stop's context. The select picked one at random, so about
half the time the AS, peer or peering handler started another write after
the deadline. Each loop now takes a waiting stop first, as the prefix
handler's does. The prefix, AS and peer handlers' flush loops were
near-copies, so they are now one function, flushLoop. A new test for each
handler fails about half its runs on the old code.
The AS and peer handler tests give the flush timer an interval longer than
the test. A timer flush during the fill can no longer hang
TestPeerHandlerWaitsAtTheMaximumBatchSize.
Model: opus-5-5
When a write ran past the stop deadline, a full batch's wake-up or a tick
could wait beside Stop's context. The flush goroutine's select picked one of
them at random, so about half the time it wrote that batch after the
deadline. That was the intermittent failure of
TestPrefixHandlerSkipsItsLastWriteAfterTheStopDeadline. The flush goroutine
now takes a waiting stop first. A new test hands over a full batch while a
write is held; it fails about half its runs on the old code.
TestPrefixHandlerSkipsItsLastWriteAfterTheStopDeadline and
TestPrefixHandlerStopLogsTheTableChangesLeftUnwritten also failed when the
flush timer fired between two of their steps. Both now give the handler a
timer interval longer than the test.
Model: opus-5-5
The prefix handler applies each update to the routing table in memory as
it handles it, sorted into announcements and withdrawals as for the live
route tables. Each flush first writes at most 20,000 of the table's
changed pairs, with every new peer; the rest follow in the next flushes.
A failed write puts its changes and new peers back in the table. Stop
makes one last flush, and logs the changed pairs left.
Startup reads the table before the streamer connects, and logs its pairs,
peers and time. The per-peer counts are a slice, so the read takes about
a third less time. The health check's start period is 60 seconds, so a
full table's read fits before upaas looks.
Model: opus-5-5
Each handler took the latest time per key in a batch and inserted it as
a new row's first_seen, so a key updated twice in one batch, seconds
apart, was stored first seen at the later time, and the IP lookup could
answer with a prefix first_seen after the route's last_seen.
The prefix handler now keeps the earliest time of all of the batch's
updates for each prefix. The AS, peer and peering handlers keep the
earliest and the latest time per key in database.SeenTimes, and
GetOrCreateASNBatch, UpdatePeerBatch and RecordPeering store a new row
first seen at the earliest and last seen at the latest. A stored row
still has only its last_seen set, to the latest.
Model: opus-5-5
Shutdown passes its context to every stage: the streamer, the four
handlers, the WHOIS fetcher, the database maintainer and the HTTP server
each stop waiting when it ends. A handler then skips its last write and
logs how many updates it did not write, and the HTTP server closes the
connections still open. The database is checkpointed and closed last,
and left open once the deadline has passed, since a write may still
hold it. The stop hook gives Shutdown 8 seconds and fx 9, inside the 10
seconds upaas and docker stop allow, so a stop signal exits 0.
Model: opus-5-5
A path ending in an AS set, such as [64496, 64497, [64498, 64499]], was
flattened into the sequence, so the route was stored as originated by the
last set member and the peering handler recorded peerings the path does not
show. ASPath now keeps the sequence and a trailing set of several members
apart; Origin() gives the last AS of the sequence, or 0 (no known origin)
for such a set. The handlers use it; peerings come from the sequence only.
The pages show origin 0 as no known origin, and the IP lookup asks no WHOIS
lookup for it. Empty paths are unchanged.
Model: opus-5-5
Takes .golangci.yml, .dockerignore, .gitignore, .editorconfig, the
workflow and REPO_POLICIES.md from sneak/prompts 0653579, with this
repository's own entries after the shared content. Lint (golangci-lint
v2.14.0) and test are Dockerfile phases that script/lint and script/test
build alone with a new CACHEBUST value, keeping Go's and golangci-lint's
caches; the image build depends on both. make fmt and make fmt-check also
format the markdown with prettier, pinned in yarn.lock. The agent
guidance moves to AGENTS.md. The Go code is fixed for what the new
configuration finds, and the AS and prefix detail reads now return the
rows' error after their loops.
Model: opus-5-5
clawbot