Pause a registry's WHOIS server after it refuses a query (closes #218 )

The fetcher kept one interval for every registry, so a registry that refused a query was asked again within seconds. RIPE blocked the test host's address that way. The WHOIS client now pauses an RIR's or NIC.br's server for an hour after the server refuses a query. A refused or reset connection is a refusal, and so is an %ERROR:201: or %ERROR:202: line. Lookups at the other servers go on. A lookup whose RIR is paused leaves the AS stale, and counts as neither a success nor a failure. IANA is never paused, because a pause of IANA would stop every lookup. NIC.br's hour-long skip after its over-limit answer is now the same pause. Model: opus-5-5
2026-11-02 01:28:50 +00:00
22 changed files with 983 additions and 2208 deletions
@@ -10,231 +10,156 @@
 # Status
-RouteWatch is pre-1.0, with no git tags. `main` last took `next` on 2026-10-27,
+pre-1.0. No git tags. `main` last took `next` on 2026-10-06, through
-through https://git.eeqj.de/sneak/routewatch/pulls/166.
+https://git.eeqj.de/sneak/routewatch/pulls/6. The linter and the tests are
-https://git.eeqj.de/sneak/routewatch/pulls/202 carries the commits since then
+phases of the Docker build, and the Gitea workflow runs `script/cibuild` (the
-into `main`. It is labelled `merge-ready` and waits for sneak.
+format check of the Go code and the markdown, both phases, then the image) on
-
+every push. The image sets memory ceilings for a 5 GiB container (README
-The linter and the tests are phases of the Docker build. The Gitea workflow runs
+"Memory") and the README says how to run it under upaas (README "Running under
-`script/cibuild` on every push. That script runs the two phases and checks the
+upaas"). A 25-hour run of `a9b6860` on the live feed peaked at about 0.6 GiB
-formatting of the Go code and the markdown. Then it builds the image.
+(https://git.eeqj.de/sneak/routewatch/issues/3). A run of `825afa2` on the live
-
+feed found the prefix handler dropping about a quarter of its updates because
-The image sets memory ceilings for a 5 GiB container (README "Memory"). README
+its route writes were slower than the feed
-"Running under upaas" says how to run it.
+(https://git.eeqj.de/sneak/routewatch/issues/96); the cheaper writes since then
-
+are measured by a benchmark, not yet on the live feed.
 A five-day run of `1c65656` on the live feed, from 2026-10-27 to 2026-11-01,
 found three problems:
 - The prefix handler dropped updates while a peer's session went down
  (https://git.eeqj.de/sneak/routewatch/issues/216).
 - The AS handler dropped updates when the feed burst
  (https://git.eeqj.de/sneak/routewatch/issues/217).
 - The WHOIS fetcher kept asking a registry that had refused it
  (https://git.eeqj.de/sneak/routewatch/issues/218).
 All three are fixed on `next`. So is
 https://git.eeqj.de/sneak/routewatch/issues/222, which the third fix brought up.
 A run of `26d939a` on the live feed showed the prefix handler still dropping
 updates while the home page's cards were worked out.
 https://git.eeqj.de/sneak/routewatch/issues/226 fixed that on `next`.
 https://git.eeqj.de/sneak/routewatch/issues/225, also on `next`, takes the peers
 whose sessions go down together out of the table in one walk. A 24-hour run of
 `9bd0e91` on the live feed is checking whether any handler still drops updates.
 Two questions wait for sneak: which history the prefix page keeps
 (https://git.eeqj.de/sneak/routewatch/issues/161), and which fields the IP
 lookup answer gains (https://git.eeqj.de/sneak/routewatch/issues/203).
 The order of the project's remaining work is kept on
 https://git.eeqj.de/sneak/routewatch/issues/86.
 # Next Step
-Post the outcome of the 24-hour run of `9bd0e91` on
+Run `next` on the live feed, as README "Running under upaas" describes, until
-https://git.eeqj.de/sneak/routewatch/issues/226: whether any handler still drops
+the routing table stops growing. Post its size and memory figures on
-updates, and the process's peak memory against 5 GiB. File an issue for each
+https://git.eeqj.de/sneak/routewatch/issues/122. The same run makes the final
-problem the run finds.
+check https://git.eeqj.de/sneak/routewatch/issues/96 names: whether the prefix
 handler still drops updates after the first hour. If it does, a new issue
 follows with the numbers. Then an audit of the code at the commit that run
 tested.
 # Completed Steps
 - 2026-11-03: the streamer drops an update for a handler only when that
  handler's queue is full. It used to drop a random share of the updates once
  the queue was half full. The prefix handler's queue now holds 100,000
  messages, about 100 MiB. It used to hold 20,000. While the routing table's
  `Overview` walks every prefix, the prefix handler applies updates more slowly.
  In a run on the live feed, the feed paused during such a walk and then sent
  its backlog. The streamer then dropped about 10,000 of the prefix handler's
  updates (closes #229)
 - 2026-11-02: `README.md` is rewritten from scratch, sentence by sentence,
  against the rules of https://git.eeqj.de/sneak/routewatch/issues/157. Every
  statement was checked against the code on `next` (closes #157)
 - 2026-11-02: `TODO.md` is rewritten from scratch, sentence by sentence, against
  the rules of https://git.eeqj.de/sneak/routewatch/issues/157. Every Completed
  Steps entry keeps its date, its facts and its issue. Status and Next Step
  describe the repo as the tracker shows it (part of #157)
 - 2026-11-02: when the sessions of several peers go down, the prefix handler
  takes the peers out of the routing table together. The table's `PeersDown`
  takes several peers out of every prefix in one walk. One walk runs at a time.
  The next one takes every peer whose session went down since the previous one
  began. Each peer used to have a walk of its own. A peer's messages stayed held
  back while the walks queued ahead of its own ran. The handler holds back at
  most 150,000 messages, about 150 MiB. It used to hold back at most 20,000. The
  status page shows how many messages the handler holds back (closes #225)
 - 2026-11-02: working out the home page's cards no longer stops the prefix
  handler. The walk over the routing table holds the table's lock for 10,000
  prefixes at a time, then lets the lock go for as long again, as `PeerDown`
  does. It used to take the lock back at once. The prefix handler then applied
  about one update per 10,000 prefixes walked. During a walk, its queue could
  fill, and the streamer then dropped updates for it (closes #226)
 - 2026-11-02: A skipped WHOIS lookup now makes the fetcher wait longer before
  the next one, as a failed lookup does. A lookup is skipped when its RIR is
  paused. An RIR is one of the five regional Internet registries. The longest
  wait is one minute. The status page does not count a skip as a failure. A skip
  used to leave the wait as it was. Each lookup asks IANA which RIR holds the AS
  before it finds that RIR paused. So while most stale ASes belonged to a paused
  RIR, the fetcher could ask IANA every second (closes #222)
 - 2026-11-02: A peer's session going down no longer stops the prefix handler.
  The handler runs the routing table's `PeerDown` on a goroutine of its own, one
  peer at a time. Meanwhile it goes on taking messages. `PeerDown` holds the
  table's lock for 10,000 prefixes at a time, then lets the lock go for as long
  again. The handler holds back the down peer's later messages until that peer's
  `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. At that
  number, `HandleMessage` waits until one of them is applied. When `Stop` runs
  past its deadline, it logs how many held-back messages were not applied
  (closes #216)
 - 2026-11-02: RouteWatch pauses an RIR's or NIC.br's WHOIS server for an hour
-  after the server refuses a query. NIC.br is Brazil's registry. A refused or
+  after the server refuses a query. A refused or reset connection is a refusal.
-  reset connection is a refusal. So is an answer with an `%ERROR:201:` or
+  So is an answer with an `%ERROR:201:` or `%ERROR:202:` line. Lookups at the
-  `%ERROR:202:` line. Lookups at the other servers go on. An AS whose RIR is
+  other servers go on. An AS whose RIR is paused is looked up after the pause.
-  paused is looked up after the pause. While NIC.br is paused, a Brazilian AS
+  IANA is never paused, because a pause of IANA would stop every lookup. A
-  gets the details from LACNIC's answer. IANA is never paused, because a pause
+  refusal by IANA is an ordinary failure. NIC.br is also paused for an hour
-  of IANA would stop every lookup. A refusal by IANA is an ordinary failure.
+  after it says RouteWatch has passed NIC.br's query limit. Before, the fetcher
-  NIC.br is also paused for an hour after it says RouteWatch has passed its
+  asked a server again within seconds of a refusal (closes #218)
-  query limit. Before, the fetcher asked a server again within seconds of a
+- 2026-11-02: the AS handler never waits for a database write. Its batch holds
  refusal (closes #218)
 - 2026-11-02: The AS handler never waits for a database write. Its batch holds
  one entry per AS, with the earliest and the latest time of its updates. So the
  batch no longer grows with the feed's rate. It used to hold one entry for each
-  AS of each path. At 300,000 entries it waited for the write in progress. In a
+  AS of each path and to wait once it held 300,000. In a burst of the feed its
-  burst of the feed its queue then filled, and the streamer dropped messages for
+  queue then filled, and the streamer dropped messages for it (closes #217)
-  it (closes #217)
+- 2026-10-27: the tests of `internal/routewatch` no longer fail on a loaded
 - 2026-10-27: The tests of `internal/routewatch` no longer fail on a loaded
  host. Five of them gave `HandleMessage` two seconds to apply a full prefix
-  batch. Under the race detector that can take longer. A test now waits up to 15
+  batch, which under the race detector can take longer. A test now waits up to
-  seconds for each step it expects. The tests of the handlers' limits now give
+  15 seconds for each step it expects. The tests of the handlers' limits give
-  the handlers small limits. So the package no longer runs past its 90-second
+  the handlers small ones, so the package no longer runs past its 90-second
  limit (closes #213, #215)
- 2026-10-27: The server works out the home page's cards every minute until it
+- 2026-10-27: the server works out the home page's cards every minute until it
-  stops. It used to work them out only once, at startup. The end of the context
+  stops. It used to work them out only once, at startup: the end of the context
-  given to `Start` stopped the refreshes (closes #211)
+  that `Start` is given stopped the refreshes (closes #211)
- 2026-10-27: README "Rationale" says why RouteWatch exists. RouteWatch answers
+- 2026-10-27: README "Rationale" says why RouteWatch exists: to answer our other
-  our other services' questions about IP addresses, from its own routing table
+  services' questions about IP addresses, from its own routing table and the
-  and the WHOIS details it keeps. The section still says what RIS Live does not
+  WHOIS details it keeps. It still says what RIS Live does not give (closes
-  give (closes #204)
+  #204)
- 2026-10-27: Every JSON answer carries `@meta`. A panic on a JSON route answers
+- 2026-10-27: every JSON answer carries `@meta`. A panic on a JSON route answers
  500 with the JSON error body. `/status.json` and the health check answer 405
  with the JSON error body to any method they do not take. The 500 and the 405
  used to have an empty body. The home page's IP lookup shows the whole JSON
  answer, `@meta` included. It used to show only the answer's `data`, or only
  the message of an error answer. A test checks `@meta` on every JSON route, in
  its successes and its errors (closes #201)
- 2026-10-27: Every value with a unit in a JSON answer is a number. The unit is
+- 2026-10-27: every value in a JSON answer that has a unit is a number, with the
-  at the end of the field name. `@meta.execution_time` is now
+  unit at the end of its field name. `@meta.execution_time` is now
-  `execution_time_ms`. In the stats, `uptime` is now `uptime_seconds`,
+  `execution_time_ms`. The stats' `uptime`, `connection_duration` and
-  `connection_duration` is now `connection_duration_seconds`, and `memory_usage`
+  `memory_usage` are now `uptime_seconds`, `connection_duration_seconds` and
-  is now `memory_usage_bytes`. `whois_stats.current_interval` is now
+  `memory_usage_bytes`. `whois_stats.current_interval` is now
-  `whois_stats.current_interval_seconds`. They used to be strings such as "12
+  `current_interval_seconds`. They used to be strings such as "12 ms",
-  ms", "1h18m34s" or "1.1 GB". `connection_duration_seconds` is null while the
+  "1h18m34s" or "1.1 GB". `connection_duration_seconds` is null while the feed
-  feed is not connected. `API_POLICIES.md` is vendored from `sneak/prompts`, and
+  is not connected. `API_POLICIES.md` is vendored from `sneak/prompts`, and
-  `REPO_POLICIES.md` is re-vendored (closes #199)
+  `REPO_POLICIES.md` re-vendored (closes #199)
- 2026-10-27: Every result of the IP lookup has the same fields. A field without
+- 2026-10-27: every result of the IP lookup has the same fields. A field without
-  a value is null. It used to be left out. An address with no route in the
+  a value is null. It used to be left out. An address with no live route has
-  routing table has every route and AS field set to null. `ptr` is an empty list
+  every route and AS field set to null, and `ptr` is an empty list when there
-  when there are no PTR names. The new field `num_origins` gives the number of
+  are no PTR names. New fields give the number of origin ASes of the netblock,
-  origin ASes of the netblock, and `origin_asns` gives which they are (closes
+  in `num_origins`, and which they are, in `origin_asns` (closes #200)
-  #200)
+- 2026-10-27: with `TRUSTED_PROXIES` unset, the server trusts reverse proxies at
 - 2026-10-27: With `TRUSTED_PROXIES` unset, the server trusts reverse proxies at
  the private addresses of RFC 1918. `/ip/`, the home page's `curl` card and the
  request log then give the client's address from `X-Forwarded-For`. They used
-  to give the proxy's address. A list in `TRUSTED_PROXIES` replaces the default.
+  to give the proxy's address. A list in `TRUSTED_PROXIES` replaces the default,
-  `TRUSTED_PROXIES` set but empty trusts no proxy (closes #198)
+  and `TRUSTED_PROXIES` set but empty trusts no proxy (closes #198)
- 2026-10-27: The home page has cards with figures of the routing table. One
+- 2026-10-27: the home page has cards with figures of the routing table: the
-  card shows the newest and the oldest kept route. One shows the newest and the
+  newest and oldest kept route, the newest and oldest updated AS, the AS with
-  oldest updated AS. One shows the AS with the most prefixes. One shows the ASes
+  the most prefixes, the ASes with a single prefix, and fifteen well-known
-  with a single prefix. One shows fifteen well-known networks. Two cards count
+  networks. Two cards count the bogon announcements and the prefixes with more
-  the bogon announcements and the prefixes with more than one origin AS. They
+  than one origin AS, and link to the new pages `/bogons` and
-  link to the new pages `/bogons` and `/multiple-origins`, which list them. A
+  `/multiple-origins` that list them. A new package, `internal/bogon`, tells
-  new package, `internal/bogon`, tells which prefixes and AS numbers are
+  which prefixes and AS numbers are reserved. The server works the figures out
-  reserved. The server works the figures out once a minute (closes #154)
+  once a minute (closes #154)
- 2026-10-27: The per-peer route tables are gone, with the code that wrote,
+- 2026-10-27: the per-peer route tables are gone, with the code that wrote,
  counted and read them. The prefix handler applies each update to the routing
-  table in memory. It writes the table's changes every second. It also stores
+  table in memory and writes the table's changes every second. It also stores
-  each new prefix with its first-seen time. The image's memory limits fit a full
+  each new prefix with its first-seen time. The image's memory limits are set
-  routing table of about 1 GB. `GOMEMLIMIT` is 2.5 GiB. The writing connection's
+  for a full routing table of about 1 GB: `GOMEMLIMIT` is 2.5 GiB, the writing
-  page cache is 512 MiB. SQLite's soft and hard heap limits are 768 MiB and 1
+  connection's page cache 512 MiB, and SQLite's heap limits 768 MiB and 1 GiB
-  GiB (closes #122)
+  (closes #122)
- 2026-10-27: The pages, the API and the stats read every route from the routing
+- 2026-10-27: the pages, the API and the stats read every route from the routing
  table in memory. For a prefix with several origin ASes, the IP lookup gives
  the one most peers carry. Its `last_seen` is the newest kept route's time. The
-  prefix page and API list every peer that carries the prefix. Only the kept
+  prefix page and API list every peer that carries the prefix, with an AS path,
-  routes show an AS path, a next hop and a time. The stats count the origins
+  next hop and time only for the kept routes. The stats count the origins whose
-  whose kept routes were all withdrawn, in `origins_without_kept_route` (part of
+  kept routes were all withdrawn, in `origins_without_kept_route` (part of #122)
 - 2026-10-27: the prefix, AS and peer handlers make a timed write on every tick
  of their timers, every 1, 2 and 2 seconds. They used to skip every other tick.
  A full batch's write restarts the timer, so the next timed write comes one
  interval after that write began (closes #192)
 - 2026-10-27: once the stop deadline has passed, the AS, peer and peering
  handlers start no write either. A full batch or a tick of the timer that
  waited while a write ran past the deadline could still start one. The flush
  goroutines of the prefix, AS and peer handlers now run the same code (closes
  #189)
 - 2026-10-27: once the stop deadline has passed, the prefix handler starts no
  flush. A batch that filled while a write ran past the deadline could still be
  written after it (closes #188)
 - 2026-10-27: the prefix handler applies each update to the routing table in
  memory, and writes the table's changes to the database every second. A write
  takes at most 20,000 changed pairs, and the rest follow in the next writes. A
  failed write is written again later. Startup reads the table before the
  streamer connects. The pages still read the per-peer route tables (part of
  #122)
- 2026-10-27: The prefix, AS and peer handlers make a timed write on every tick
+- 2026-10-27: the database keeps a copy of the whole routing table in a new
  of their timers. The prefix handler's timer ticks every second. The other two
  tick every two seconds. They used to skip every other tick. A full batch's
  write restarts the timer. So the next timed write comes one interval after
  that write began (closes #192)
 - 2026-10-27: Once the stop deadline has passed, the AS, peer and peering
  handlers start no write either. Before, a full batch or a timer tick could
  wait while a write ran past the deadline. Either could then start one more
  write. The flush goroutines of the prefix, AS and peer handlers now run the
  same code (closes #189)
 - 2026-10-27: Once the stop deadline has passed, the prefix handler starts no
  flush. Before, a batch that filled during a late write could still be written
  after the deadline (closes #188)
 - 2026-10-27: The prefix handler applies each update to the routing table in
  memory. It writes the table's changes to the database every second. A write
  takes at most 20,000 changed pairs of prefix and origin AS. The rest follow in
  the next writes. A failed write is written again later. Startup reads the
  table before the streamer connects. The pages still read the per-peer route
  tables (part of #122)
 - 2026-10-27: The database keeps a copy of the whole routing table in a new
  `routes` table, one row per prefix and origin AS. The peers' numbers are in a
-  new `route_peers` table. One function writes the pairs that changed in memory.
+  new `route_peers` table. One function writes the pairs that changed in memory,
-  Another reads the whole table back into memory at startup. Nothing uses them
+  and another reads the whole table back into memory at startup. Nothing uses
-  yet (part of #122)
+  them yet (part of #122)
- 2026-10-27: A new package, `internal/routetable`, holds the whole routing
+- 2026-10-27: a new package, `internal/routetable`, holds the whole routing
  table in memory. For each prefix and origin AS it keeps the peers that carry
-  it, one bit per peer. It also keeps the two newest of their routes. Nothing
+  it, one bit per peer, and the two newest of their routes. Nothing uses it yet
-  uses it yet (part of #122)
+  (part of #122)
- 2026-10-27: Each result of the IP lookup lists the public blocklists the
+- 2026-10-27: each result of the IP lookup lists the public blocklists the
  address is on, in `blocklists`. The lists are DShield, Feodo Tracker, two
-  FireHOL lists, two Spamhaus DROP lists and the Tor exit list. RouteWatch
+  FireHOL lists, Spamhaus DROP and the Tor exit list. RouteWatch downloads each
-  downloads each list on its own schedule and answers from copies in memory. The
+  list on its own schedule and answers from copies in memory. The copies are
-  copies are also kept in the state directory's `blocklists` directory. IP
+  also kept in the state directory's `blocklists` directory. IP lookup answers
-  lookup answers carry `Cache-Control`. An address with no route now has a
+  carry `Cache-Control`. An address with no live route now has a result too,
-  result too, without route fields. So the IP lookup no longer answers 404
+  without route fields, so the IP lookup no longer answers 404 (closes #158)
-  (closes #158)
+- 2026-10-27: the home page has an IP information API card. It shows the `curl`
- 2026-10-27: The home page has an IP information API card. It shows the `curl`
+  command that looks up the visitor's own address at `/ip/{addr}`, and that
-  command that looks up the visitor's own address at `/ip/{addr}`. It also shows
+  command's live answer (closes #156)
-  that command's live answer (closes #156)
+- 2026-10-27: on the AS page, a click on the heading of the IPv4 or the IPv6
 - 2026-10-27: On the AS page, a click on the heading of the IPv4 or the IPv6
  prefix list folds or unfolds the list. Both lists start open (closes #155)
- 2026-10-27: The AS page lists its IPv4 and its IPv6 prefixes by mask length,
+- 2026-10-27: the AS page lists its IPv4 and its IPv6 prefixes by mask length,
-  shortest first. Prefixes of one mask length are listed by address. They used
+  shortest first, and prefixes of one mask length by address. They used to be
-  to be sorted by address first, then by mask length (closes #153)
+  listed by address (closes #153)
- 2026-10-27: The connection status on `/` and `/status` gets a spinner. It
+- 2026-10-27: on `/` and `/status`, a spinner beside the status turns one frame
-  turns one frame each time the stats show more messages than the time before.
+  each time the stats show more messages than the time before. A page that has
-  After three refreshes with no new messages, the page shows "stale" in red
+  not seen more messages for three refreshes shows "stale" and the time since
-  instead. Next to it is the time since the last fresh answer (closes #152)
+  then instead, in red (closes #152)
- 2026-10-27: The WHOIS details of a Brazilian AS keep their accented letters.
+- 2026-10-27: the WHOIS details of a Brazilian AS keep their accented letters.
-  LACNIC's answer for such an AS is NIC.br's answer with each accented letter
+  LACNIC's answer for such an AS comes from NIC.br, Brazil's registry, with each
-  replaced. NIC.br is Brazil's registry. The details now come from NIC.br's own
+  accented letter replaced. The details now come from NIC.br's own server,
-  server, `whois.registro.br`. When that query fails, the details from LACNIC's
+  `whois.registro.br`. When that query fails, the details from LACNIC's answer
-  answer are kept and a warning is logged. RouteWatch stops asking NIC.br for an
+  are kept and a warning is logged. After NIC.br answers that RouteWatch is over
-  hour after NIC.br says RouteWatch is over its query limit (closes #174)
+  its query limit, RouteWatch stops asking NIC.br for an hour (closes #174)
 - 2026-10-27: `make asupdate` stops with an error when the download of `as.csv`
  breaks off. The generator used to retry the failed read forever. A record that
  does not parse is still logged and skipped (closes #177)
@@ -243,192 +168,182 @@ problem the run finds.
  one AS. A failed download used to replace the list with an empty file. The
  generator reads `as.csv` from `ipverse/as-metadata`, the source's new name
  (closes #173)
- 2026-10-27: The WHOIS details of an AS at RIPE, APNIC or AFRINIC come only
+- 2026-10-27: the WHOIS details of an AS at RIPE, APNIC or AFRINIC come only
  from its `aut-num` object and the organisation object it names. RIPE and APNIC
  start their answer with the `as-block` object of the AS's block. Its dates
  were stored as the AS's (closes #168)
 - 2026-10-27: AS names keep their accented letters. The AS list built into
  RouteWatch is rebuilt from today's `as.csv`, which added a country code
  column. That file no longer lists AS 0 or the documentation AS numbers. A
-  WHOIS answer that is not valid UTF-8 is decoded as Latin-1 before it is parsed
+  WHOIS answer that is not valid UTF-8 is decoded as Latin-1, which RIPE and
-  and stored. RIPE and APNIC send Latin-1 (closes #165)
+  APNIC send, before it is parsed and stored (closes #165)
- 2026-10-27: The Age column on the AS and prefix pages reads in words, such as
+- 2026-10-27: the Age column on the AS and prefix pages reads in words, such as
-  "5 minutes ago" or "2 weeks ago". It used to print Go durations such as
+  "5 minutes ago" or "2 weeks ago", instead of Go durations such as "360h0m0s
-  "360h0m0s ago". An age under a minute reads "now" or in seconds. An age over a
+  ago". An age under a minute reads in seconds, and one over a month in months
-  month reads in months or years instead of a date (closes #163)
+  or years instead of a date (closes #163)
- 2026-10-26: A write to a stored AS, peering or peer keeps the earlier of the
+- 2026-10-26: a write to a stored AS, peering or peer keeps the earlier of the
-  two first-seen times. It keeps the later of the two last-seen times. An old
+  two first-seen times and the later of the two last-seen times. An old update
-  update from a collector catching up on a backlog no longer moves last seen
+  from a collector catching up on a backlog no longer moves last seen back
-  back (closes #162)
+  (closes #162)
- 2026-10-26: The status page shows "Route Churn/min". That is the number of
+- 2026-10-26: the status page shows "Route Churn/min", the number of prefix
-  prefix announcements and withdrawals from all peers in the last full clock
+  announcements and withdrawals, from all peers, in the last full clock minute.
-  minute. `/api/v1/stats` reports it as `route_churn_per_min` in `stream`. It
+  `/api/v1/stats` reports it as `route_churn_per_min` in `stream`, which
  replaces the moving average `route_churn_per_sec`. The messages and updates
  per second on `/` and `/status` are whole numbers (closes #160)
- 2026-10-26: The WHOIS fetcher backs off more slowly. A failure makes its wait
+- 2026-10-26: the WHOIS fetcher backs off more slowly. A failure makes its wait
  1.5 times as long, up to one minute instead of five. Its queries to RIPE,
-  APNIC and AFRINIC leave out person and role objects. RIPE counts those objects
+  APNIC and AFRINIC leave out person and role objects, which RIPE counts against
-  against a daily limit per address. A refused query is logged with the RIR's
+  a daily limit per address. A refused query is logged with the RIR's error line
-  error line (closes #159)
+  (closes #159)
- 2026-10-26: The home page's AS name search works. It lists the ASes whose
+- 2026-10-26: the home page's AS name search lists the ASes whose handle or
-  handle or description holds the text, as links to their pages. The new
+  description holds the text, as links to their pages, from the new
-  `/api/v1/search/as` answers it. The search ignores case. When the database has
+  `/api/v1/search/as`; the search ignores case. When the database has no record
-  no record of an AS, its page takes the handle and description from the AS list
+  of an AS, its page takes the handle and description from the AS list built
-  built into RouteWatch. It still lists the AS's prefixes (closes #151)
+  into RouteWatch and still lists the AS's prefixes (closes #151)
- 2026-10-19: README "HTTP Endpoints" lists every route with what it answers.
+- 2026-10-19: README "HTTP Endpoints" lists every route with what it answers,
-  README "Database Schema" matches `schema.sql`. The rest of the README matches
+  "Database Schema" matches `schema.sql`, and the rest of the README matches the
-  the code. This file is brought up to date (closes #72)
+  code; this file brought up to date (closes #72)
- 2026-10-19: Route writes are cheaper. The live route and prefix tables lose
+- 2026-10-19: route writes are cheaper: the live route and prefix tables lose
-  their random UUID keys. The prefix tables lose their duplicate index. A route
+  their random UUID keys and the prefix tables their duplicate index, a route
-  update sets only the columns that change. Every write goes through one
+  update sets only the columns that change, and every write goes through one
  connection with a page cache of 1 GiB (README "Memory"). A benchmark writes
  the prefix handler's largest batch faster than the feed sends it (closes #96)
- 2026-10-19: The checkpoint of the WAL (SQLite's write-ahead log), the
+- 2026-10-19: the WAL checkpoint, the incremental vacuum and ANALYZE take the
-  incremental vacuum and `ANALYZE` take the database lock that every batch write
+  database lock every batch write holds, so a vacuum no longer fails with
-  holds. So a vacuum no longer fails with "database is locked". An `ANALYZE` no
+  "database is locked" and an ANALYZE no longer makes a batch write lose its
-  longer makes a batch write lose its updates. `ANALYZE` runs as
+  updates; ANALYZE runs as `PRAGMA optimize` with an analysis limit (closes #94)
-  `PRAGMA optimize` with an analysis limit (closes #94)
+- 2026-10-19: shutdown stops the streamer first, and each handler writes
- 2026-10-19: Shutdown stops the streamer first. Each handler then writes
+  everything left on its queue before it stops, so the updates waiting there are
  everything left on its queue before it stops. So the updates waiting there are
  no longer lost (closes #69)
- 2026-10-19: A path that ends in an AS set of several members has origin AS 0,
+- 2026-10-19: a path that ends in an AS set of several members has origin AS 0,
-  which means no known origin. It used to have the set's last member as its
+  no known origin, instead of the set's last member, and the set adds no
-  origin. The set adds no peerings. The pages show such a route as having no
+  peerings; the pages show such a route as having no known origin (closes #75)
  known origin (closes #75)
 - 2026-10-19: `bgp_peer_count` drops to 0 when the last peer goes down (closes
  #92)
- 2026-10-19: `bgp_peer_count` counts every peer the feed has heard from. It
+- 2026-10-19: `bgp_peer_count` counts every peer the feed has heard from, not
-  used to count only the peers whose session opened since startup. It drops a
+  only those whose session opened since startup, and drops a peer whose session
-  peer whose session goes down (closes #74)
+  goes down (closes #74)
- 2026-10-19: The shared files are re-vendored from `sneak/prompts` at
+- 2026-10-19: the shared files are re-vendored from `sneak/prompts` at
-  `0653579`. The linter is `golangci-lint` v2.14.0 with its configuration. Lint
+  `0653579`: golangci-lint v2.14.0 and its configuration, lint and test as
-  and test are `Dockerfile` phases that keep Go's and `golangci-lint`'s caches.
+  `Dockerfile` phases that keep Go's and golangci-lint's caches, markdown
-  `make fmt` also formats the markdown with `prettier`. The agent guidance moves
+  formatted with prettier, agent guidance in `AGENTS.md`; the code is fixed for
-  to `AGENTS.md`. The code is fixed for what the new linter configuration finds
+  what the new linter configuration finds (closes #52)
-  (closes #52)
+- 2026-10-19: the streamer reads feed lines up to 1 MiB, so a line over 64 KiB
- 2026-10-19: The streamer reads feed lines up to 1 MiB. A line over 64 KiB used
+  no longer ends the connection (closes #76)
-  to end the connection (closes #76)
+- 2026-10-19: a peer's live routes are removed when its BGP session goes down;
- 2026-10-19: A peer's live routes are removed when its BGP session goes down.
+  `STATE` messages are never dropped for backpressure, and one that finds its
  `STATE` messages are never dropped when a queue is full. One that finds its
  queue full waits for room ahead of the handler's other messages (closes #56)
- 2026-10-19: The prefix-length pages pick their sample with a bounded number of
+- 2026-10-19: the prefix-length pages pick their sample with a bounded number of
-  index lookups. They used to read every route of that length. They list each
+  index lookups instead of reading every route of that length, and list each
  prefix once (closes #66)
- 2026-10-19: A request past its time limit answers 504 instead of 408.
+- 2026-10-19: a request past its time limit answers 504 instead of 408, and
  `/api/v1/stats` answers 504 instead of 500 when its read runs out of time
  (closes #70)
- 2026-10-19: The AS, peer and prefix handlers write each batch outside the lock
+- 2026-10-19: the AS, peer and prefix handlers write each batch outside the lock
-  their message handling needs. So messages keep coming off their queues during
+  their message handling needs, so messages keep coming off their queues during
-  a write. Meanwhile the next batch grows to at most ten batch sizes (closes
+  a write; meanwhile the next batch grows to at most ten batch sizes (closes
  #65)
- 2026-10-19: A live route is identified by prefix and peer. So an announcement
+- 2026-10-19: a live route is identified by prefix and peer, so an announcement
-  with a new origin AS replaces the peer's route. A withdrawal removes it
+  with a new origin AS replaces the peer's route and a withdrawal removes it
  whatever its origin AS (closes #55)
- 2026-10-14: The streamer reconnects when RIS Live ends the stream or sends
+- 2026-10-14: the streamer reconnects when RIS Live ends the stream or sends
-  nothing for 60 seconds. The health check reports the feed disconnected
+  nothing for 60 seconds, and the health check reports the feed disconnected
  meanwhile (closes #53)
- 2026-10-14: The prefix page takes the names of the origin ASes from the prefix
+- 2026-10-14: the prefix page takes the names of the origin ASes from the prefix
-  query. It used to read every route of each origin AS (closes #67)
+  query instead of reading every route of each origin AS (closes #67)
- 2026-10-14: The daemon exits with status 1 when its web server cannot listen
+- 2026-10-14: the daemon exits with status 1 when its web server cannot listen
-  on its port. It used to run on without a web server (closes #73)
+  on its port, instead of running on without it (closes #73)
- 2026-10-14: The prefix pages parse the prefix before looking it up. One that
+- 2026-10-14: the prefix pages parse the prefix before looking it up: one that
-  does not parse answers 400. Any valid way of writing a prefix finds its routes
+  does not parse answers 400, and any valid way of writing a prefix finds its
-  (closes #71)
+  routes (closes #71)
- 2026-10-14: The IP lookup ignores routes shorter than /8 for IPv4 and /16 for
+- 2026-10-14: the IP lookup ignores routes shorter than /8 for IPv4 and /16 for
-  IPv6. So a default route is no longer named as an address's owner (closes #57)
+  IPv6, so a default route is no longer named as an address's owner (closes #57)
- 2026-10-14: A prefix batch keeps each peer's last update for a prefix. It used
+- 2026-10-14: a prefix batch keeps each peer's last update for a prefix, instead
-  to keep one update per prefix across all peers (closes #54)
+  of one update per prefix across all peers (closes #54)
 - 2026-10-14: `reconnect_count` in `/api/v1/stats` counts every reconnect
  (closes #68)
- 2026-10-03: `/api/v1/stats` serves the prefix distribution from memory. The
+- 2026-10-03: `/api/v1/stats` serves the prefix distribution from memory, seeded
-  distribution is seeded at startup and adjusted on every live-route write. So a
+  at startup and adjusted on every live-route write, so a request no longer
-  request no longer reads every live route (closes #30)
+  reads every live route (closes #30)
- 2026-10-03: A new database is created with `auto_vacuum` set to incremental.
+- 2026-10-03: a new database is created with `auto_vacuum` set to incremental,
-  The setting goes in the connection string, so it is set before the file is
+  through the connection string so it is set before the file is first written,
-  first written. Each periodic incremental vacuum now returns up to 1000 free
+  and each periodic incremental vacuum now returns up to 1000 free pages; the
-  pages. The late `PRAGMA auto_vacuum` in `Initialize`, which SQLite ignored, is
+  late `PRAGMA auto_vacuum` in `Initialize`, which SQLite ignored, is gone
-  gone (closes #43)
+  (closes #43)
- 2026-10-03: Looking up an IP address no longer reads every IPv6 route. Both
+- 2026-10-03: looking up an IP address no longer reads every IPv6 route: both
  families find the most specific live route with at most 33 or 129 lookups on
-  the prefix index. The IPv4 range columns are gone. Prefixes from the feed are
+  the prefix index, and the IPv4 range columns are gone. Prefixes from the feed
-  stored in one text form. So an IPv6 withdrawal, which the feed sends
+  are stored in one text form, so an IPv6 withdrawal, which the feed sends
  uncompressed, now removes its route (closes #48)
- 2026-10-02: A plain `docker build .` stamps the commit's tag or short commit
+- 2026-10-02: a plain `docker build .` stamps the commit's tag or short commit
-  into the page footer. It used to stamp `unknown`. The value comes from
+  (`git describe --tags --always`) into the page footer instead of `unknown`:
-  `git describe --tags --always`. `.dockerignore` sends `.git` without
+  `.dockerignore` sends `.git` without `.git/config`, a `VERSION` build argument
-  `.git/config`. A `VERSION` build argument takes precedence. `make build`
+  takes precedence, and `make build` stamps the same value (closes #46)
-  stamps the same value (closes #46)
+- 2026-09-29: the entrypoint creates the data directory if it is missing and
- 2026-09-29: The entrypoint creates the data directory if it is missing. It
+  stops the start if a step fails; README "Running under upaas" no longer asks
  stops the start if a step fails. README "Running under upaas" no longer asks
  for the host directory to be created first (closes #42)
 - 2026-09-29: `.dockerignore` keeps `.git`, local build output, local databases
-  and `.env` out of the Docker build context. So they stay out of the source
+  and `.env` out of the Docker build context, and so out of the source archive
-  archive in the image (closes #39)
+  in the image (closes #39)
- 2026-09-29: The README's first line names the MIT licence and the author. The
+- 2026-09-29: README first line names the MIT license and the author; the
-  License section says MIT and links `LICENSE`. An Author section is added. This
+  License section now says MIT and links `LICENSE`, and an Author section was
-  file is brought up to date (closes #38)
+  added; this file brought up to date (closes #38)
- 2026-09-29: `LICENSE` is added, with the MIT licence (closes #1)
+- 2026-09-29: MIT `LICENSE` (closes #1)
- 2026-09-28: Stopping the daemon while the feed is flowing no longer panics
+- 2026-09-28: stopping the daemon while the feed is flowing no longer panics
-  with "send on closed channel". The read loop checks for a stop just before
+  with "send on closed channel": the read loop checks for a stop just before
-  handing a message to the handler queues. A second `Stop` no longer closes the
+  handing a message to the handler queues, and a second `Stop` no longer closes
-  queues again (closes #34)
+  the queues again (closes #34)
 - 2026-09-28: `docker stop` no longer kills the daemon 2 seconds after the stop
-  signal. The entrypoint switches to the `routewatch` user with `setpriv`
+  signal: the entrypoint switches to the `routewatch` user with `setpriv`
-  instead of `runuser`. So the daemon receives the signal itself. It gets the
+  instead of `runuser`, so the daemon receives the signal itself and gets the
  whole wait `docker stop` allows, up to its own 60-second limit (closes #33)
- 2026-09-28: RouteWatch is ready to run under upaas. A set but invalid `PORT`,
+- 2026-09-28: ready to run under upaas: a set but invalid `PORT`,
-  `XDG_DATA_HOME` or `MALLOC_ARENA_MAX` stops the start. The health check
+  `XDG_DATA_HOME` or `MALLOC_ARENA_MAX` stops the start, the health check
-  follows `PORT`. README gains a "Running under upaas" section (closes #31)
+  follows `PORT`, README "Running under upaas" section (closes #31)
- 2026-09-22: The database statistics are kept in memory. The counts are seeded
+- 2026-09-22: realtime in-memory database statistics: counts seeded at startup
-  at startup and adjusted on every write. The oldest and newest route times come
+  and adjusted on every write, oldest/newest route timestamps via index-end
-  from a lookup at each end of an index. `/api/v1/stats` no longer scans the
+  lookups; `/api/v1/stats` no longer scans the tables (closes #27)
-  tables (closes #27)
+- 2026-09-21: batch writes take the write lock when their transaction begins
- 2026-09-21: Batch writes take the write lock when their transaction begins
+  (`_txlock=immediate`), so they wait out a WAL checkpoint instead of failing
  (`_txlock=immediate`). So they wait out a WAL checkpoint instead of failing
  with "database is locked" (closes #25)
- 2026-09-21: `MALLOC_ARENA_MAX=2` in the image caps the number of glibc malloc
+- 2026-09-21: `MALLOC_ARENA_MAX=2` in the image caps glibc malloc arenas, so
-  arenas. So memory outside the Go runtime no longer grows with the core count
+  memory outside the Go runtime no longer grows with the core count (closes #23)
-  (closes #23)
+- 2026-09-21: `GOMEMLIMIT=1536MiB` in the image; README Memory section with the
 - 2026-09-21: The image sets `GOMEMLIMIT=1536MiB`. README "Memory" gives the
  memory budget and the 5 GiB container limit (closes #13)
- 2026-09-21: The four handler queues hold at most 20,000 messages each, down
+- 2026-09-21: the four handler queues hold at most 20,000 messages each, down
  from 100,000 (closes #11)
- 2026-09-21: Two goroutine leaks are fixed. The stats handlers leaked after a
+- 2026-09-21: two goroutine leaks fixed: the stats handlers after a timeout and
-  timeout. The streamer's tickers leaked on every reconnect (closes #12)
+  the streamer's tickers on every reconnect (closes #12)
- 2026-09-21: Parsed RIS messages no longer keep the unused `Community` and
+- 2026-09-21: parsed RIS messages no longer keep the unused `Community` and
  `Raw` fields (closes #9)
- 2026-09-21: `.editorconfig` is added. A Gitea workflow runs `script/cibuild`
+- 2026-09-21: `.editorconfig`, and a Gitea workflow that runs `script/cibuild`
  on every push (closes #14)
- 2026-09-21: The peering handler's AS-path map holds at most 500,000 paths.
+- 2026-09-21: the peering handler's AS-path map holds at most 500,000 paths and
-  Every 30 seconds it is swapped for an empty one instead of copied (closes #10)
+  is swapped for an empty one every 30 seconds instead of copied (closes #10)
- 2026-09-21: SQLite's memory is bounded across the whole connection pool. Each
+- 2026-09-21: SQLite memory bounded across the whole connection pool: a 64 MiB
-  connection has a 64 MiB page cache. The heap limits are 1 GiB soft and 1.5 GiB
+  page cache on each connection, 1 GiB soft and 1.5 GiB hard heap limits (closes
-  hard (closes #8)
+  #8)
- 2026-09-21: The Docker build runs the format check, the linter and the tests.
+- 2026-09-21: the Docker build runs the format check, the linter and the tests,
-  Linting runs in a separate stage on a `golangci-lint` image pinned by digest
+  linting in a separate stage on a golangci-lint image pinned by digest (closes
-  (closes #5)
+  #5)
- 2026-09-21: `make test` skips the live feed test, with `-short`. So
+- 2026-09-21: `make test` skips the live-network feed test (`-short`), so
  `make check` no longer depends on the network (closes #2)
- 2026-07-07: The scripts-to-rule-them-all layout is adopted. `script/` holds
+- 2026-07-07 Adopted scripts-to-rule-them-all: `script/` entrypoints, Makefile
-  the entrypoints. The `Makefile` targets `check`, `test`, `lint`, `fmt` and
+  shims, README Entrypoints section
-  `fmt-check` call them. So do `bootstrap`, `setup`, `docker` and `hooks`.
+- 2026-02-22: repo policy compliance: required policy files, .gitignore update,
-  README gains an "Entrypoints" section
+  Makefile fmt-check/check/docker/hooks targets, gofmt pass
- 2026-02-22: The repo is brought in line with the repo policies. The required
+  (repo-policies-compliance, unmerged)
-  policy files are added. `.gitignore` is updated. The `Makefile` gains
+- 2026-01-01: WAL checkpointing: periodic (5s), on startup with logging,
-  `fmt-check`, `check`, `docker` and `hooks` targets. The Go code is formatted
+  TRUNCATE mode; fixed slow queries
-  with `gofmt`. The work was on the branch `repo-policies-compliance`, unmerged
+- 2025-12-31: status page: navbar, home page with search, oldest/newest route
-  when this entry was written
+  timestamps, NULL handling in WHOIS stats; container runs as routewatch user
- 2026-01-01: The WAL is checkpointed every 5 seconds, and at startup with a log
+  with proper state dir
-  line. Checkpoints run in `TRUNCATE` mode. This fixed slow queries
+- 2025-12-30: structured HTTP request logging, increased timeouts, CIDR prefix
- 2025-12-31: The pages get a navigation bar and a home page with a search. The
+  URL routing fixes, status page metrics and footer
-  status page shows the oldest and newest route times. The WHOIS stats query
+- 2025-12-29: Dockerfile multi-stage build with source archive; SQLite
-  handles `NULL`. The container runs as the `routewatch` user with the right
+  incremental vacuum for non-blocking space reclamation
-  state directory
+- 2025-07-27: initial RouteWatch BGP stream monitor
 - 2025-12-30: HTTP requests are logged as structured log lines. The HTTP
  timeouts are longer. Prefix URLs with a slash in them route correctly. The
  status page gets more metrics and a footer
 - 2025-12-29: The `Dockerfile` builds in stages and puts a source archive in the
  image. SQLite reclaims space with incremental vacuum, which does not block
 - 2025-07-27: The first commit: RouteWatch, a BGP stream monitor
 # Future Steps
 None.
@@ -273,7 +273,7 @@ func TestVacuumReturnsFreePages(t *testing.T) {
 	writeTableChanges(t, db, table)
 	// The peer going down removes every prefix, and so every row.
-	table.PeersDown([]string{peerA})
+	table.PeerDown(peerA)
 	writeTableChanges(t, db, table)
 	var before int
@@ -107,7 +107,7 @@ func TestReadRouteTableGivesBackWhatWasWritten(t *testing.T) {
 	// A peer moves to another origin, one goes down and a new one comes.
 	announceTo(source, tablePrefixV4, asB, peerA, 60)
-	source.PeersDown([]string{peerC})
+	source.PeerDown(peerC)
 	announceTo(source, tablePrefixV6, asB, peerE, 70)
 	writeTableChanges(t, db, source)
@@ -133,7 +133,7 @@ func TestReadRouteTableGivesATableThatTakesLaterUpdates(t *testing.T) {
 	// another one goes down, and a new peer comes.
 	for _, table := range []*routetable.Table{source, read} {
 		table.Withdraw(netip.MustParsePrefix(tablePrefixV4), peerA)
-		table.PeersDown([]string{peerB})
+		table.PeerDown(peerB)
 		announceTo(table, tablePrefixShortV4, asB, peerD, 50)
 	}
@@ -10,6 +10,10 @@ import (
 	"git.eeqj.de/sneak/routewatch/internal/ristypes"
 )
 // walkBatch is how many prefixes the walk over the whole table reads in one
 // hold of the read lock. The updates waiting for the lock get it in between.
 const walkBatch = 10000
 // Overview holds figures about the whole table. Working them out takes a walk
 // over every prefix, so a caller asks for them now and then and keeps them.
 type Overview struct {
@@ -79,9 +83,8 @@ type OriginPeers struct {
 // Overview works out the table's overview. It lists at most maxListed bogon
 // announcements and at most maxListed prefixes with more than one origin AS:
-// the first in prefix order. It visits every prefix in the table, in batches
+// the first in prefix order. It walks over every prefix, but holds the read
-// of walkBatch. After each batch, it lets the table's lock go for as long as
+// lock for walkBatch prefixes at a time.
 // it held the lock for the batch.
 func (t *Table) Overview(maxListed int) Overview {
 	var overview Overview
@@ -111,18 +114,15 @@ func (t *Table) Overview(maxListed int) Overview {
 }
 // walkPrefixes calls visit with every prefix of the table and its origins. It
-// holds the read lock for walkBatch prefixes at a time. After each batch, it
+// holds the read lock, but lets it go after every walkBatch prefixes, so the
-// lets the table's lock go for as long as it held the lock for the batch.
+// updates waiting for it get in. A prefix that is added or removed in such a
-// During a walk over many batches, updates can then take the lock about half
+// gap may or may not be visited, as Go allows for a map that changes during a
-// of the time. The walk may or may not visit a prefix that is added or removed
+// range over it. visit must not keep origins once it returns.
 // while the lock is let go. Go allows this in a range over a map that changes.
 // visit must not keep origins once it returns.
 func (t *Table) walkPrefixes(visit func(prefix netip.Prefix, origins []*origin)) {
 	t.mu.RLock()
 	defer t.mu.RUnlock()
 	visited := 0
 	batchStart := time.Now()
 	for _, f := range []*family{&t.ipv4, &t.ipv6} {
 		for _, prefixes := range f.byLength {
@@ -131,13 +131,8 @@ func (t *Table) walkPrefixes(visit func(prefix netip.Prefix, origins []*origin))
 				visited++
 				if visited%walkBatch == 0 {
 					held := time.Since(batchStart)
 					t.mu.RUnlock()
 					time.Sleep(held)
 					t.mu.RLock()
 					batchStart = time.Now()
 				}
 			}
 		}
@@ -131,61 +131,3 @@ func TestWalkPrefixesLetsUpdatesIn(t *testing.T) {
 			"before its second batch")
 	}
 }
 // TestOverviewLetsUpdatesIn applies another peer's 1,000 announcements while
 // Overview walks a table with more prefixes than five of the walk's batches.
 // After each batch, the walk lets the lock go for as long as it held the lock.
 // So all 1,000 announcements must be applied before Overview returns. A walk
 // that takes the lock back at once lets in about one update per batch.
 func TestOverviewLetsUpdatesIn(t *testing.T) {
 	t.Parallel()
 	const (
 		asn       = 1000
 		peer      = "192.0.2.1"
 		otherPeer = "192.0.2.2"
 		// prefixCount is more prefixes than five of the walk's batches.
 		prefixCount   = 5*walkBatch + 1
 		announcements = 1000
 	)
 	routeTime := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
 	table := New(Params{})
 	prefixes := make([]netip.Prefix, 0, prefixCount)
 	addr := netip.MustParseAddr("1.0.0.0")
 	for range prefixCount {
 		prefix := netip.PrefixFrom(addr, addr.BitLen())
 		table.Announce(prefix, asn, Route{Peer: peer, Time: routeTime})
 		prefixes = append(prefixes, prefix)
 		addr = addr.Next()
 	}
 	overviewDone := make(chan struct{})
 	go func() {
 		defer close(overviewDone)
 		table.Overview(0)
 	}()
 	// Wait until Overview holds the read lock. TryLock fails while a reader
 	// holds it.
 	for table.mu.TryLock() {
 		table.mu.Unlock()
 		runtime.Gosched()
 	}
 	for _, prefix := range prefixes[:announcements] {
 		table.Announce(prefix, asn, Route{Peer: otherPeer, Time: routeTime})
 	}
 	select {
 	case <-overviewDone:
 		t.Fatal("Overview returned before another peer's 1,000 announcements " +
 			"were applied")
 	default:
 	}
 	<-overviewDone
 }
@@ -227,7 +227,7 @@ func TestCounts(t *testing.T) {
 	})
 	// peerC was the last peer of the IPv6 prefix. It keeps its number.
-	table.PeersDown([]string{peerC})
+	table.PeerDown(peerC)
 	checkCounts(t, table, routetable.Counts{
 		Pairs:                2,
 		Peers:                3,
@@ -27,11 +27,6 @@ const (
 	// kept routes in order: each of its nodes holds up to 63 times.
 	keptTimesDegree = 32
 	// walkBatch is how many prefixes a walk over the whole table visits in
 	// one hold of the lock. Overview and PeersDown each walk the whole table.
 	// Callers waiting for the lock get it between batches.
 	walkBatch = 10000
 	ipv4Bits = 32
 	ipv6Bits = 128
@@ -205,60 +200,24 @@ func (t *Table) Withdraw(prefix netip.Prefix, peer string) {
 	}
 }
-// PeersDown applies the sessions of peers going down: each of the peers leaves
+// PeerDown applies a peer's session going down: the peer leaves every peer
-// every peer set, and every route of each is dropped. When any of them carries
+// set, and every route of it is dropped. When the peer carries any prefix, it
-// a prefix, PeersDown visits every prefix in the table once, in batches of
+// visits every prefix in the table.
-// walkBatch. After each batch, it lets the table's lock go for as long as it
+func (t *Table) PeerDown(peer string) {
 // held the lock for the batch. During a walk over many batches, reads and
 // updates can then take the lock about half of the time. An announcement by
 // one of the peers applied during the walk may be dropped too. So the caller
 // applies those peers' later updates only once PeersDown has returned.
 func (t *Table) PeersDown(peers []string) {
 	t.mu.Lock()
 	defer t.mu.Unlock()
-	// numbers holds the number of each of the peers that carries a prefix.
+	number, known := t.peerNumbers[peer]
-	numbers := make([]uint, 0, len(peers))
+	if !known || t.peerPrefixes[number] == 0 {
 	for _, peer := range peers {
 		number, known := t.peerNumbers[peer]
 		if known && t.peerPrefixes[number] > 0 {
 			numbers = append(numbers, number)
 		}
 	}
 	if len(numbers) == 0 {
 		return
 	}
 	visited := 0
 	batchStart := time.Now()
 	for _, f := range []*family{&t.ipv4, &t.ipv6} {
 		for _, prefixes := range f.byLength {
 			for prefix, origins := range prefixes {
-				for _, number := range numbers {
+				o := originWithPeer(origins, number)
-					o := originWithPeer(origins, number)
+				if o != nil {
-					if o != nil {
+					t.leave(prefix, o, number)
 						t.leave(prefix, o, number)
 						// leave may have removed an origin of the
 						// prefix, or the prefix itself.
 						origins = prefixes[prefix]
 					}
 				}
 				// Go allows the map to change before the next step of
 				// the range over it.
 				visited++
 				if visited%walkBatch == 0 {
 					held := time.Since(batchStart)
 					t.mu.Unlock()
 					time.Sleep(held)
 					t.mu.Lock()
 					batchStart = time.Now()
 				}
 			}
 		}
@@ -1,69 +0,0 @@
 package routetable
 import (
 	"net/netip"
 	"runtime"
 	"testing"
 	"time"
 )
 // TestPeersDownLetsReadsIn reads the table's counts again and again while a
 // peer that carries more than two batches of prefixes goes down. PeersDown lets
 // the lock go after every walkBatch prefixes, so a read must get in while the
 // peer still carries some of its prefixes but no longer all.
 func TestPeersDownLetsReadsIn(t *testing.T) {
 	t.Parallel()
 	const (
 		asn  = 1000
 		peer = "192.0.2.1"
 		// prefixCount is more prefixes than two of walkBatch.
 		prefixCount = 2*walkBatch + 1
 	)
 	routeTime := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
 	table := New(Params{})
 	addr := netip.MustParseAddr("1.0.0.0")
 	for range prefixCount {
 		prefix := netip.PrefixFrom(addr, addr.BitLen())
 		table.Announce(prefix, asn, Route{Peer: peer, Time: routeTime})
 		addr = addr.Next()
 	}
 	// Hold a read lock until PeersDown waits for the lock, so that the first
 	// read below starts once PeersDown holds it. TryRLock fails once a writer
 	// waits.
 	table.mu.RLock()
 	peerDownDone := make(chan struct{})
 	go func() {
 		defer close(peerDownDone)
 		table.PeersDown([]string{peer})
 	}()
 	for table.mu.TryRLock() {
 		table.mu.RUnlock()
 		runtime.Gosched()
 	}
 	table.mu.RUnlock()
 	for {
 		routes := table.Counts().IPv4Routes
 		if routes > 0 && routes < prefixCount {
 			break
 		}
 		select {
 		case <-peerDownDone:
 			t.Fatal("no read got in while PeersDown ran")
 		default:
 		}
 	}
 	<-peerDownDone
 }
@@ -69,10 +69,10 @@ func withdraw(peer string) step {
 	}
 }
-// peersDown is the step of the sessions of peers going down.
+// peerDown is the step of peer's session going down.
-func peersDown(peers ...string) step {
+func peerDown(peer string) step {
 	return func(table *routetable.Table, _ netip.Prefix) {
-		table.PeersDown(peers)
+		table.PeerDown(peer)
 	}
 }
@@ -272,28 +272,21 @@ func TestWithdrawalsAndPeerDowns(t *testing.T) {
 		{
 			name: "a peer going down leaves the prefix",
 			steps: []step{announce(peerA, asA, 10), announce(peerB, asA, 20),
-				peersDown(peerB)},
+				peerDown(peerB)},
 			want: []routetable.Origin{{ASN: asA, Peers: []string{peerA},
 				Routes: []routetable.Route{routeOf(peerA, asA, 10)}}},
 		},
 		{
 			name: "a peer going down after it withdrew changes nothing",
 			steps: []step{announce(peerA, asA, 10), announce(peerB, asA, 20),
-				withdraw(peerB), peersDown(peerB)},
+				withdraw(peerB), peerDown(peerB)},
 			want: []routetable.Origin{{ASN: asA, Peers: []string{peerA},
 				Routes: []routetable.Route{routeOf(peerA, asA, 10)}}},
 		},
 		{
 			name: "peers going down together each leave the prefix",
 			steps: []step{announce(peerA, asA, 10), announce(peerB, asB, 20),
 				announce(peerC, asB, 30), peersDown(peerA, peerB)},
 			want: []routetable.Origin{{ASN: asB, Peers: []string{peerC},
 				Routes: []routetable.Route{routeOf(peerC, asB, 30)}}},
 		},
 	})
 }
-func TestPeersDownLeavesEveryPrefix(t *testing.T) {
+func TestPeerDownLeavesEveryPrefix(t *testing.T) {
 	t.Parallel()
 	table := routetable.New(routetable.Params{})
@@ -310,7 +303,7 @@ func TestPeersDownLeavesEveryPrefix(t *testing.T) {
 	}
 	table.Announce(prefixes[0], asB, routeOf(peerB, asB, 20))
-	table.PeersDown([]string{peerA})
+	table.PeerDown(peerA)
 	want := []routetable.Origin{{ASN: asB, Peers: []string{peerB},
 		Routes: []routetable.Route{routeOf(peerB, asB, 20)}}}
@@ -340,32 +333,6 @@ func TestPeersDownLeavesEveryPrefix(t *testing.T) {
 	}
 }
 // TestPeersDownTakesOutPeersOfDifferentPrefixes takes peerA and peerB out in
 // one PeersDown. peerA carries only prefixV4, and peerB only prefixV6. So
 // peerA's leaving removes prefixV4 before PeersDown looks for peerB there. Both
 // prefixes must then be gone.
 func TestPeersDownTakesOutPeersOfDifferentPrefixes(t *testing.T) {
 	t.Parallel()
 	table := routetable.New(routetable.Params{})
 	prefixes := []netip.Prefix{
 		netip.MustParsePrefix(prefixV4),
 		netip.MustParsePrefix(prefixV6),
 	}
 	table.Announce(prefixes[0], asA, routeOf(peerA, asA, 10))
 	table.Announce(prefixes[1], asB, routeOf(peerB, asB, 20))
 	table.PeersDown([]string{peerA, peerB})
 	for _, prefix := range prefixes {
 		got := table.Origins(prefix)
 		if got != nil {
 			t.Errorf("Origins(%s) = %+v after its only peer went down, want nil",
 				prefix, got)
 		}
 	}
 }
 func TestCountsOriginsWithoutKeptRoute(t *testing.T) {
 	t.Parallel()
@@ -391,7 +358,7 @@ func TestCountsOriginsWithoutKeptRoute(t *testing.T) {
 		{announce(peerC, asA, 30), 0},
 		{withdraw(peerB), 0},
 		{withdraw(peerC), 1},
-		{peersDown(peerA), 0},
+		{peerDown(peerA), 0},
 	}
 	for i, s := range steps {
@@ -497,7 +464,7 @@ func TestTakeChangesHoldsOnlyPairsThatChanged(t *testing.T) {
 	checkChanges(t, table, []change{}, map[string]uint{})
 	// peerA going down changes v4, the one prefix it carries.
-	table.PeersDown([]string{peerA})
+	table.PeerDown(peerA)
 	checkChanges(t, table, []change{
 		{Prefix: prefixV4, OriginASN: asA, Peers: []uint{1, 2},
 			Routes: []routetable.Route{
@@ -506,7 +473,7 @@ func TestTakeChangesHoldsOnlyPairsThatChanged(t *testing.T) {
 	}, map[string]uint{})
 	// peerA, which carries nothing now, goes down again.
-	table.PeersDown([]string{peerA})
+	table.PeerDown(peerA)
 	checkChanges(t, table, []change{}, map[string]uint{})
 }
@@ -650,7 +617,7 @@ func TestConcurrentUse(t *testing.T) {
 		func(round int) { table.Announce(prefix, asA, routeOf(peerA, asA, round)) },
 		func(round int) { table.Announce(prefix, asB, routeOf(peerB, asB, round)) },
 		func(int) { table.Withdraw(prefix, peerA) },
-		func(int) { table.PeersDown([]string{peerB}) },
+		func(int) { table.PeerDown(peerB) },
 		func(int) { table.PutBackChanges(table.TakeChanges(math.MaxInt)) },
 		func(int) { table.ChangeCount() },
 		func(int) { table.Lookup(prefix.Addr()) },
@@ -187,33 +187,25 @@ func (f *ASNFetcher) recordSuccess() {
 	f.statsMu.Unlock()
 }
-// lengthenInterval makes the interval 1.5 times as long, up to maxInterval.
+// recordFailure increases the interval on failed fetch using exponential backoff.
-// A failed lookup and a skipped one both lengthen it.
+func (f *ASNFetcher) recordFailure() {
 func (f *ASNFetcher) lengthenInterval() {
 	f.intervalMu.Lock()
-	defer f.intervalMu.Unlock()
+	f.consecutiveFails++
 	// Exponential backoff: multiply by 1.5, capped at max
 	newInterval := min(
 		f.currentInterval*backoffNumerator/backoffDenominator, maxInterval)
 	if newInterval != f.currentInterval {
-		f.logger.Debug("Increased fetch interval",
+		f.logger.Debug("Increased fetch interval due to failure",
 			"old_interval", f.currentInterval,
 			"new_interval", newInterval,
 			"consecutive_failures", f.consecutiveFails,
 		)
 		f.currentInterval = newInterval
 	}
 }
 // recordFailure counts a failed fetch and lengthens the interval.
 func (f *ASNFetcher) recordFailure() {
 	f.intervalMu.Lock()
 	f.consecutiveFails++
 	f.intervalMu.Unlock()
 	f.lengthenInterval()
 	// Record error time for stats, dropping those older than statsWindow
 	f.statsMu.Lock()
 	now := f.now()
@@ -299,12 +291,9 @@ func (f *ASNFetcher) fetchNextStale(ctx context.Context) {
 // The WHOIS client pauses an RIR's or NIC.br's server for whois.PauseTime
 // after the server refuses a query. Lookups at the other servers go on. IANA
 // is never paused, so a refusal by IANA is an ordinary failure. A lookup
-// whose RIR is paused fails with whois.ErrPaused and is skipped. A skipped
+// whose RIR is paused fails with whois.ErrPaused. Such a lookup leaves the
-// lookup leaves the AS stale and is not counted as a failure. Like a
+// AS stale, and counts as neither a success nor a failure. The lookups that
-// failure, a skipped lookup lengthens the interval. Each lookup asks IANA
+// QueueImmediate asks for also run through fetchAndUpdate.
 // before it finds its RIR paused, so with a one-second interval the fetcher
 // would ask IANA every second. The lookups that QueueImmediate asks for also
 // run through fetchAndUpdate.
 func (f *ASNFetcher) fetchAndUpdate(ctx context.Context, asn int) bool {
 	ctx, cancel := context.WithTimeout(ctx, time.Minute)
 	defer cancel()
@@ -328,7 +317,6 @@ func (f *ASNFetcher) fetchAndUpdate(ctx context.Context, asn int) bool {
 	if errors.Is(err, whois.ErrPaused) {
 		f.logger.Info("Skipped WHOIS lookup, the RIR's server is paused",
 			"asn", asn, "error", err)
 		f.lengthenInterval()
 		return false
 	}
@@ -196,48 +196,6 @@ func TestFetchAndUpdateWaitsOutAPause(t *testing.T) {
 	}
 }
 // TestFetchAndUpdateBacksOffDuringAPause looks up an AS at RIPE while RIPE
 // refuses the query. The refusal pauses RIPE for an hour, so every later
 // lookup in the test is skipped. The test sets the WHOIS client's clock and
 // looks the AS up again each time the interval has passed. Like a failure,
 // each skip must make the interval 1.5 times as long, up to one minute. The
 // skips must not count as failures.
 func TestFetchAndUpdateBacksOffDuringAPause(t *testing.T) {
 	t.Parallel()
 	f, _ := newFetcherWithWHOISAnswers(t, map[string]string{
 		ianaWHOISServer: ianaRefersToRIPE,
 		ripeWHOISServer: ripeAccessDenied,
 	})
 	now := time.Date(2026, 11, 2, 0, 0, 0, 0, time.UTC)
 	f.whoisClient.Now = func() time.Time { return now }
 	f.fetchAndUpdate(t.Context(), testASNA)
 	for i, want := range []time.Duration{
 		33750 * time.Millisecond,
 		50625 * time.Millisecond,
 		time.Minute,
 		time.Minute,
 	} {
 		now = now.Add(f.getInterval())
 		f.fetchAndUpdate(t.Context(), testASNA)
 		if got := f.getInterval(); got != want {
 			t.Errorf("interval after %d skipped lookups = %v, want %v",
 				i+1, got, want)
 		}
 	}
 	got := f.GetStats()
 	if got.ConsecutiveFails != 1 || got.ErrorsLastHour != 1 {
 		t.Errorf("%d failures in a row and %d errors, want 1 and 1",
 			got.ConsecutiveFails, got.ErrorsLastHour)
 	}
 }
 // TestFetchAndUpdateKeepsAskingIANA looks up an AS while IANA's server
 // refuses the connection. The test sets the WHOIS client's clock. The next
 // lookup comes a second before whois.PauseTime has passed. IANA is never
@@ -5,7 +5,6 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
 	"io"
 	"log/slog"
 	"slices"
 	"sync"
@@ -178,14 +177,13 @@ type stopLog struct {
 	Msg          string `json:"msg"`
 	Updates      int    `json:"updates"`
 	TableChanges int    `json:"table_changes"`
 	HeldMessages int    `json:"held_messages"`
 	Paths        int    `json:"paths"`
 	BatchEntries int    `json:"batch_entries"`
 }
 // newErrorLogger returns a logger that writes only errors, as JSON lines, to
 // logs.
-func newErrorLogger(logs io.Writer) *logger.Logger {
+func newErrorLogger(logs *bytes.Buffer) *logger.Logger {
 	return &logger.Logger{Logger: slog.New(slog.NewJSONHandler(logs,
 		&slog.HandlerOptions{Level: slog.LevelError}))}
 }
@@ -15,15 +15,12 @@ import (
 )
 const (
-	// prefixHandlerQueueSize is the most messages the handler's queue holds.
+	// prefixHandlerQueueSize is the queue capacity for prefix tracking
-	// While the routing table's Overview or PeersDown walks every prefix, it
+	// operations, about 4 seconds of feed at peak. The streamer drops updates
-	// holds the table's lock about half of the time. The handler then applies
+	// rather than blocks when a queue is full (a STATE message is held until
-	// updates more slowly. When the feed pauses for some seconds and then
+	// there is room), so this bounds memory. Batches still flush on a timer
-	// catches up, its backlog can arrive faster than that, and the queue
+	// (prefixBatchTimeout).
-	// holds the backlog. The streamer drops an update that finds the queue
+	prefixHandlerQueueSize = 20000
 	// full. A message takes about 1 KiB, so a full queue takes about 100 MiB
 	// (README "Memory").
 	prefixHandlerQueueSize = 100000
 	// prefixBatchSize is the number of prefix updates to batch together
 	prefixBatchSize = 25000
@@ -47,21 +44,6 @@ const (
 	// most one pair. So the million pairs that a full-table peer's session
 	// going down changes are written in about a minute.
 	maxTableChangesPerWrite = 20000
 	// maxHeldBackMessages is the most messages the handler holds back from
 	// peers whose session went down. A peer's messages are held back from
 	// the moment its session goes down until peerDown has applied them. If
 	// peerDown is running at that moment, the peer waits for the rest of
 	// peerDown's current PeersDown call. The peer also waits while peerDown
 	// applies the messages held back from that call's peers, those that
 	// arrive meanwhile included. The next PeersDown call takes the peer out.
 	// Once that call returns, peerDown applies the messages held back from
 	// that call's peers, one peer after another, in the order the call lists
 	// them. A held-back message that says the peer's session went down again
 	// makes the peer's later messages wait for the next call. A message takes
 	// about 1 KiB, so 150,000 take about 150 MiB (README "Memory"). At this
 	// number HandleMessage waits until one of them is applied.
 	maxHeldBackMessages = 150000
 )
 // PrefixHandler keeps the routing table in memory current from the feed. It
@@ -70,17 +52,6 @@ const (
 // does an announcement whose AS path is empty or holds AS 0 anywhere. A peer
 // whose BGP session goes down leaves every prefix.
 //
 // The table's PeersDown applies sessions going down. It visits every prefix,
 // which takes seconds on a full table. So the handler runs PeersDown on a
 // separate goroutine and goes on taking messages meanwhile. One PeersDown runs
 // at a time. The next one takes every peer whose session went down since the
 // previous one began. PeersDown lets the table's lock go about half of the
 // time, so the handler applies other peers' messages while PeersDown runs. The
 // handler holds back a down peer's later messages until peerDown applies them,
 // as the doc of maxHeldBackMessages describes. It applies them in order, one at
 // a time, and takes new messages in between. It holds back at most
 // maxHeldBackMessages messages.
 //
 // Each flush writes the table's changes to the database. It also stores each
 // prefix the updates name with the time it was first seen.
 type PrefixHandler struct {
@@ -96,7 +67,7 @@ type PrefixHandler struct {
 	// give a smaller one.
 	tableChangesPerWrite int
-	// Batching. mu guards batch, writing and writingTableChanges: the
+	// Batching. mu guards batch, writing and writingTableChanges only: the
 	// flush goroutine takes them under mu and writes them after releasing
 	// mu, so HandleMessage waits for a database write only while batch is at
 	// batchMaxSize.
@@ -116,26 +87,6 @@ type PrefixHandler struct {
 	// Stop does not wait for a write in progress to hand it over.
 	stopCh chan context.Context
 	done   chan struct{} // closed when the flush goroutine returns
 	// Peers whose session went down. heldBack, under mu, holds each such
 	// peer's later messages, in order, until the peer has left every prefix
 	// and they are applied. A peer with no such messages yet has an empty
 	// entry. heldCount, under mu, is how many messages heldBack holds.
 	// heldTaken is signalled each time a held-back message is taken out to
 	// be applied; its L is &mu.
 	heldBack  map[string][]*ristypes.RISMessage
 	heldCount int
 	heldTaken sync.Cond
 	// downPeers, under mu, lists the peers whose session went down and that
 	// the next PeersDown takes. peerDownRunning, under mu, says whether
 	// peerDown is running. peerDowns counts the peerDown goroutines that have
 	// not returned.
 	downPeers       []string
 	peerDownRunning bool
 	peerDowns       sync.WaitGroup
 	// tablePeersDown is the table's PeersDown. Only peerDown calls it, so a
 	// test may replace it before it hands the handler a message.
 	tablePeersDown func(peers []string)
 }
 // seenPrefix is what the batch keeps of an update: the prefix it names and its
@@ -199,11 +150,8 @@ func newPrefixHandler(
 		batchFullCh:          make(chan struct{}, 1),
 		stopCh:               make(chan context.Context, 1),
 		done:                 make(chan struct{}),
 		heldBack:             make(map[string][]*ristypes.RISMessage),
 		tablePeersDown:       table.PeersDown,
 	}
 	h.batchTaken.L = &h.mu
 	h.heldTaken.L = &h.mu
 	// Start the flush timer goroutine
 	go flushLoop(batchTimeout, h.batchFullCh, h.stopCh, h.done, h.flushBatch)
@@ -231,100 +179,11 @@ func (h *PrefixHandler) QueueCapacity() int {
 	return prefixHandlerQueueSize
 }
-// HandleMessage processes a message to track prefix information. A message
+// HandleMessage processes a message to track prefix information
 // from a peer whose session went down is held back until peerDown applies it.
 func (h *PrefixHandler) HandleMessage(msg *ristypes.RISMessage) {
 	h.mu.Lock()
 	defer h.mu.Unlock()
 	// While maxHeldBackMessages messages are held back, wait until one of them
 	// is taken out to be applied. Meanwhile the handler takes nothing off its
 	// queue, as at batchMaxSize.
 	for h.heldCount >= maxHeldBackMessages {
 		h.heldTaken.Wait()
 	}
 	held, goingDown := h.heldBack[msg.Peer]
 	if goingDown {
 		h.heldBack[msg.Peer] = append(held, msg)
 		h.heldCount++
 		return
 	}
 	h.applyMessage(msg)
 }
 // HeldBackCount returns how many messages the handler holds back from peers
 // whose session went down. The status page shows it.
 func (h *PrefixHandler) HeldBackCount() int {
 	h.mu.Lock()
 	defer h.mu.Unlock()
 	return h.heldCount
 }
 // Stop waits for every PeersDown the handler started and for the messages held
 // back meanwhile to be applied. It gives up this wait once ctx ends. Then Stop
 // stops the flush goroutine and waits for it until ctx ends. The flush
 // goroutine makes one last flush before it returns. That flush writes at most
 // tableChangesPerWrite of the routing table's changes, and the prefixes of the
 // updates the handler holds. Once ctx has ended the flush goroutine skips that
 // last flush, and a write still running is lost when the process exits. A
 // failed write of the table's changes is not retried either. Stop logs how
 // many updates and table changes are not written. It also logs how many
 // held-back messages are not applied. The caller calls Stop once its last
 // HandleMessage has returned.
 func (h *PrefixHandler) Stop(ctx context.Context) {
 	peerDownsDone := make(chan struct{})
 	go func() {
 		h.peerDowns.Wait()
 		close(peerDownsDone)
 	}()
 	select {
 	case <-peerDownsDone:
 	case <-ctx.Done():
 	}
 	h.stopCh <- ctx
 	select {
 	case <-h.done:
 	case <-ctx.Done():
 	}
 	h.mu.Lock()
 	unwritten := h.writing + len(h.batch)
 	tableChanges := h.writingTableChanges + h.table.ChangeCount()
 	heldMessages := h.heldCount
 	h.mu.Unlock()
 	if unwritten == 0 && tableChanges == 0 && heldMessages == 0 {
 		return
 	}
 	if ctx.Err() != nil {
 		h.logger.Error("Stop deadline passed before the prefix handler wrote "+
 			"its updates", "updates", unwritten, "table_changes", tableChanges,
 			"held_messages", heldMessages)
 		return
 	}
 	h.logger.Error("The prefix handler stopped before it wrote everything",
 		"updates", unwritten, "table_changes", tableChanges,
 		"held_messages", heldMessages)
 }
 // applyMessage applies msg. The caller holds mu. A peer's session going down
 // has the peer leave every prefix. An update is applied to the routing table,
 // and its prefixes join the batch.
 func (h *PrefixHandler) applyMessage(msg *ristypes.RISMessage) {
 	if msg.Type == ristypes.MessageTypeState {
 		if msg.State == ristypes.StateDown {
-			h.startPeerDown(msg.Peer)
+			h.table.PeerDown(msg.Peer)
 		}
 		return
@@ -339,6 +198,9 @@ func (h *PrefixHandler) applyMessage(msg *ristypes.RISMessage) {
 	// the path after the rest of it.
 	path := slices.Concat(msg.Path.Sequence, msg.Path.Set)
 	h.mu.Lock()
 	defer h.mu.Unlock()
 	// While the batch is at its maximum size, wait until the flush goroutine
 	// takes it. Meanwhile the handler takes nothing off its queue, which
 	// fills, and the streamer drops messages for it.
@@ -399,88 +261,39 @@ func (h *PrefixHandler) applyMessage(msg *ristypes.RISMessage) {
 	}
 }
-// startPeerDown adds peer to downPeers. It starts peerDown on a new goroutine
+// Stop stops the flush goroutine, which makes one last flush before it returns,
-// unless peerDown is running. The caller holds mu. The peer's messages are
+// and waits for it until ctx ends. That flush writes at most
-// held back until peerDown has applied them. Messages held back for an earlier
+// tableChangesPerWrite of the routing table's changes, and the prefixes of
-// session of the peer going down, and not yet applied, stay held back.
+// the updates the handler holds. Once ctx has ended the flush goroutine skips
-func (h *PrefixHandler) startPeerDown(peer string) {
+// that last flush, and a write still running is lost when the process exits. A
-	_, goingDown := h.heldBack[peer]
+// failed write of the table's changes is not retried either. Stop logs how many
-	if !goingDown {
+// updates and how many of the table's changes are not written.
-		h.heldBack[peer] = []*ristypes.RISMessage{}
+func (h *PrefixHandler) Stop(ctx context.Context) {
 	h.stopCh <- ctx
 	select {
 	case <-h.done:
 	case <-ctx.Done():
 	}
-	h.downPeers = append(h.downPeers, peer)
+	h.mu.Lock()
 	unwritten := h.writing + len(h.batch)
 	tableChanges := h.writingTableChanges + h.table.ChangeCount()
 	h.mu.Unlock()
-	if h.peerDownRunning {
+	if unwritten == 0 && tableChanges == 0 {
 		return
 	}
-	h.peerDownRunning = true
+	if ctx.Err() != nil {
-	h.peerDowns.Add(1)
+		h.logger.Error("Stop deadline passed before the prefix handler wrote "+
 			"its updates", "updates", unwritten, "table_changes", tableChanges)
-	go h.peerDown()
+		return
 }
 // peerDown takes every peer in downPeers and runs the table's PeersDown for
 // them. Then it applies the messages held back from each of those peers. It
 // repeats this until downPeers is empty.
 func (h *PrefixHandler) peerDown() {
 	defer h.peerDowns.Done()
 	for {
 		h.mu.Lock()
 		peers := h.downPeers
 		h.downPeers = nil
 		if len(peers) == 0 {
 			h.peerDownRunning = false
 			h.mu.Unlock()
 			return
 		}
 		h.mu.Unlock()
 		h.tablePeersDown(peers)
 		for _, peer := range peers {
 			h.applyHeldBack(peer)
 		}
 	}
 }
-// applyHeldBack applies the messages held back from peer, in order. It takes
+	h.logger.Error("The prefix handler stopped before it wrote everything",
-// mu for one message at a time, so HandleMessage takes new messages in
+		"updates", unwritten, "table_changes", tableChanges)
 // between. The peer's new messages join the end of the held-back ones. A
 // held-back message that says the peer's session went down again adds the
 // peer to downPeers. The rest stay held back until the next PeersDown has
 // returned.
 func (h *PrefixHandler) applyHeldBack(peer string) {
 	for {
 		h.mu.Lock()
 		held := h.heldBack[peer]
 		if len(held) == 0 {
 			delete(h.heldBack, peer)
 			h.mu.Unlock()
 			return
 		}
 		msg := held[0]
 		held[0] = nil // so that the message is freed once applied
 		h.heldBack[peer] = held[1:]
 		h.heldCount--
 		h.heldTaken.Broadcast()
 		h.applyMessage(msg)
 		h.mu.Unlock()
 		if msg.Type == ristypes.MessageTypeState &&
 			msg.State == ristypes.StateDown {
 			return
 		}
 	}
 }
 // takeBatch returns the pending updates and leaves an empty batch in their
@@ -237,83 +237,6 @@ func TestPrefixHandlerStopLogsNothingWhenItWroteEverything(t *testing.T) {
 	}
 }
 // TestPrefixHandlerStopLogsTheMessagesHeldBack stops a prefix handler with a
 // deadline that has passed, while a peer's PeersDown is held up and one message
 // of the peer is held back. Stop must log that message as not applied.
 func TestPrefixHandlerStopLogsTheMessagesHeldBack(t *testing.T) {
 	t.Parallel()
 	var logs bytes.Buffer
 	h := newPrefixHandler(newMockStore(), newErrorLogger(&logs),
 		routetable.New(routetable.Params{}), longerThanAnyTest,
 		maxTableChangesPerWrite)
 	release := holdUpPeersDown(h)
 	defer close(release)
 	h.HandleMessage(peerAnnouncement(testPeerIP, testV4Prefix))
 	h.HandleMessage(sessionDown())
 	h.HandleMessage(peerAnnouncement(testPeerIP, testV6Prefix))
 	ctx, cancel := context.WithCancel(t.Context())
 	cancel()
 	h.Stop(ctx)
 	logged := lastLogged(t, &logs)
 	if logged.HeldMessages != 1 {
 		t.Errorf("Stop logged %d held messages not applied, want 1",
 			logged.HeldMessages)
 	}
 }
 // TestPrefixHandlerStopWaitsForTheMessagesHeldBack stops a prefix handler
 // while a peer's PeersDown is held up and an announcement of the peer is held
 // back. Stop must wait. Once the PeersDown is let go and Stop returns, the
 // announcement must be in the database.
 func TestPrefixHandlerStopWaitsForTheMessagesHeldBack(t *testing.T) {
 	t.Parallel()
 	db := newTestDatabase(t)
 	h := newPrefixHandler(db, logger.New(), routetable.New(routetable.Params{}),
 		longerThanAnyTest, maxTableChangesPerWrite)
 	release := holdUpPeersDown(h)
 	h.HandleMessage(sessionDown())
 	h.HandleMessage(peerAnnouncement(testPeerIP, testV4Prefix))
 	stopped := make(chan struct{})
 	go func() {
 		h.Stop(t.Context())
 		close(stopped)
 	}()
 	select {
 	case <-stopped:
 		t.Error("Stop returned while a PeersDown was held up")
 	case <-time.After(stillWaitingAfter):
 	}
 	close(release)
 	select {
 	case <-stopped:
 	case <-time.After(stepTimeout):
 		t.Fatalf("Stop did not return within %s of the PeersDown's release",
 			stepTimeout)
 	}
 	read, err := db.ReadRouteTable(t.Context())
 	if err != nil {
 		t.Fatalf("ReadRouteTable: %v", err)
 	}
 	checkTableOrigins(t, read, map[string]map[int][]string{
 		testV4Prefix: {testASNB: {testPeerIP}},
 	})
 }
 // longerThanAnyTest is a timer interval that no test lasts. A test gives it to
 // newPrefixHandler, newASHandler, newPeerHandler or newPeeringHandler when the
 // handler must write only when its batch is full, when the test says, or on
@@ -324,32 +247,15 @@ const longerThanAnyTest = time.Hour
 // applies the updates to an empty routing table. It is built without
 // NewPrefixHandler's flush goroutine, so the test flushes itself.
 func prefixHandlerWithoutTimer(store database.Store) *PrefixHandler {
 	table := routetable.New(routetable.Params{})
 	return &PrefixHandler{
 		db:                   store,
 		logger:               logger.New(),
-		table:                table,
+		table:                routetable.New(routetable.Params{}),
 		batchMaxSize:         prefixBatchMaxSize,
 		tableChangesPerWrite: maxTableChangesPerWrite,
 		heldBack:             make(map[string][]*ristypes.RISMessage),
 		tablePeersDown:       table.PeersDown,
 	}
 }
 // holdUpPeersDown makes each PeersDown of h wait until the returned channel is
 // closed, as another peer's long PeersDown would hold it up. The test calls it
 // before it hands h a message.
 func holdUpPeersDown(h *PrefixHandler) chan struct{} {
 	release := make(chan struct{})
 	h.tablePeersDown = func(peers []string) {
 		<-release
 		h.table.PeersDown(peers)
 	}
 	return release
 }
 // newTestDatabase opens a database in a new temporary directory. It is closed
 // when the test ends.
 func newTestDatabase(t *testing.T) *database.Database {
@@ -7,10 +7,8 @@ import (
 	"log/slog"
 	"net/netip"
 	"reflect"
 	"runtime"
 	"slices"
 	"strings"
 	"sync"
 	"testing"
 	"time"
@@ -103,359 +101,6 @@ func TestPrefixHandlerAppliesUpdatesToTheRoutingTable(t *testing.T) {
 	})
 }
 // TestPrefixHandlerTakesMessagesWhileAPeerGoesDown holds up the PeersDown of a
 // peer whose session went down, as another peer's long PeersDown would. The
 // handler must go on taking messages. It applies another peer's announcement
 // at once. It holds back these messages of the down peer, in this order:
 //   - an announcement;
 //   - its session going down again;
 //   - a second announcement.
 //
 // HeldBackCount must count them. Once the PeersDown calls have returned, the
 // down peer must carry only the prefix of the second announcement.
 func TestPrefixHandlerTakesMessagesWhileAPeerGoesDown(t *testing.T) {
 	t.Parallel()
 	const heldMessages = 3
 	h := prefixHandlerWithoutTimer(newMockStore())
 	release := holdUpPeersDown(h)
 	h.HandleMessage(peerAnnouncement(testPeerIP, testV4Prefix))
 	handled := returnsInTime(func() {
 		for _, msg := range []*ristypes.RISMessage{
 			sessionDown(),
 			peerAnnouncement(secondPeerIP, otherV4Prefix),
 			peerAnnouncement(testPeerIP, testV6Prefix),
 			sessionDown(),
 			peerAnnouncement(testPeerIP, setV4Prefix),
 		} {
 			h.HandleMessage(msg)
 		}
 	})
 	if !handled {
 		t.Fatalf("the handler did not take the messages after the session "+
 			"went down within %s", stepTimeout)
 	}
 	checkTableOrigins(t, h.table, map[string]map[int][]string{
 		testV4Prefix:  {testASNB: {testPeerIP}},
 		otherV4Prefix: {testASNB: {secondPeerIP}},
 		testV6Prefix:  {},
 		setV4Prefix:   {},
 	})
 	if h.HeldBackCount() != heldMessages {
 		t.Errorf("HeldBackCount() = %d, want %d", h.HeldBackCount(),
 			heldMessages)
 	}
 	close(release)
 	if !returnsInTime(h.peerDowns.Wait) {
 		t.Fatalf("the PeersDown calls did not return within %s", stepTimeout)
 	}
 	checkTableOrigins(t, h.table, map[string]map[int][]string{
 		testV4Prefix:  {},
 		otherV4Prefix: {testASNB: {secondPeerIP}},
 		testV6Prefix:  {},
 		setV4Prefix:   {testASNB: {testPeerIP}},
 	})
 }
 // TestPrefixHandlerTakesPeersDownTogether has four peers announce testV4Prefix.
 // It holds up the PeersDown of the first peer, whose session goes down.
 // Meanwhile the sessions of the three others go down, and each of them
 // announces otherV4Prefix. Once the first PeersDown is let go, the next one
 // must take the three others together. No peer may then carry testV4Prefix,
 // and the three others must carry otherV4Prefix.
 func TestPrefixHandlerTakesPeersDownTogether(t *testing.T) {
 	t.Parallel()
 	others := []string{secondPeerIP, thirdPeerIP, fourthPeerIP}
 	h := prefixHandlerWithoutTimer(newMockStore())
 	// calls receives the peers of each PeersDown, with room for one call
 	// for each peer.
 	calls := make(chan []string, len(others)+1)
 	release := make(chan struct{})
 	h.tablePeersDown = func(peers []string) {
 		calls <- slices.Clone(peers)
 		<-release
 		h.table.PeersDown(peers)
 	}
 	for _, peer := range append([]string{testPeerIP}, others...) {
 		h.HandleMessage(peerAnnouncement(peer, testV4Prefix))
 	}
 	h.HandleMessage(sessionDown())
 	select {
 	case <-calls:
 	case <-time.After(stepTimeout):
 		t.Fatalf("PeersDown did not begin within %s", stepTimeout)
 	}
 	for _, peer := range others {
 		h.HandleMessage(sessionDownOf(peer))
 		h.HandleMessage(peerAnnouncement(peer, otherV4Prefix))
 	}
 	close(release)
 	if !returnsInTime(h.peerDowns.Wait) {
 		t.Fatalf("the PeersDown calls did not return within %s", stepTimeout)
 	}
 	close(calls)
 	later := make([][]string, 0, len(calls))
 	for peers := range calls {
 		later = append(later, peers)
 	}
 	if !reflect.DeepEqual(later, [][]string{others}) {
 		t.Errorf("the PeersDown calls after the first took %v, want %v "+
 			"together", later, others)
 	}
 	checkTableOrigins(t, h.table, map[string]map[int][]string{
 		testV4Prefix:  {},
 		otherV4Prefix: {testASNB: others},
 	})
 }
 // TestPrefixHandlerAppliesMessagesWhilePeersDownRuns gives a peer 50,000
 // prefixes and hands the prefix handler that peer's session going down. The
 // table's PeersDown then visits the prefixes in five batches. Once PeersDown
 // has begun, the handler is handed 1,000 announcements of another peer. The
 // handler must apply them all before PeersDown ends, while the down peer still
 // carries some of its prefixes.
 func TestPrefixHandlerAppliesMessagesWhilePeersDownRuns(t *testing.T) {
 	t.Parallel()
 	const (
 		// downPeerPrefixes is five times the 10,000 prefixes PeersDown visits
 		// in one hold of the table's lock.
 		downPeerPrefixes = 50000
 		// otherPeerMessages is far more messages than the handler applies
 		// while PeersDown runs, if PeersDown takes the lock back at once after
 		// each batch. The handler then applies only a few for each batch.
 		otherPeerMessages = 1000
 		// firstOtherAddr is the address of the other peer's first prefix.
 		firstOtherAddr = "2001:db8:2::"
 	)
 	h := prefixHandlerWithoutTimer(newMockStore())
 	announceManyPrefixes(h.table, downPeerPrefixes)
 	h.HandleMessage(sessionDown())
 	// Wait until PeersDown has begun: the down peer has left some of its
 	// prefixes. The down peer's prefixes are IPv4 and the other peer's are
 	// IPv6, so the counts tell them apart.
 	for h.table.Counts().IPv4Routes == downPeerPrefixes {
 		runtime.Gosched()
 	}
 	// Each announcement names the /128 of the next address.
 	addr := netip.MustParseAddr(firstOtherAddr)
 	for range otherPeerMessages {
 		h.HandleMessage(peerAnnouncement(secondPeerIP,
 			netip.PrefixFrom(addr, addr.BitLen()).String()))
 		addr = addr.Next()
 	}
 	counts := h.table.Counts()
 	if counts.IPv6Routes != otherPeerMessages {
 		t.Errorf("the other peer carries %d prefixes, want %d",
 			counts.IPv6Routes, otherPeerMessages)
 	}
 	if counts.IPv4Routes == 0 {
 		t.Errorf("PeersDown ended before the handler applied the other peer's "+
 			"%d announcements", otherPeerMessages)
 	}
 	if !returnsInTime(h.peerDowns.Wait) {
 		t.Fatalf("the PeersDown did not return within %s", stepTimeout)
 	}
 }
 // TestPrefixHandlerTakesMessagesWhileAReturningPeerSendsItsTable gives a peer
 // 50,000 prefixes and hands the prefix handler that peer's session going down.
 // Once PeersDown has begun, the peer comes back and sends more announcements
 // than the handler's queue holds, as a peer that sends its whole table again
 // does. The handler must hold them back and go on taking messages. Another
 // peer's announcement, handed next, must be applied before PeersDown ends.
 func TestPrefixHandlerTakesMessagesWhileAReturningPeerSendsItsTable(t *testing.T) {
 	t.Parallel()
 	const (
 		// downPeerPrefixes is five times the 10,000 prefixes PeersDown visits
 		// in one hold of the table's lock.
 		downPeerPrefixes  = 50000
 		returningMessages = prefixHandlerQueueSize + 1
 	)
 	table := routetable.New(routetable.Params{})
 	announceManyPrefixes(table, downPeerPrefixes)
 	h := newPrefixHandler(newMockStore(), logger.New(), table,
 		longerThanAnyTest, maxTableChangesPerWrite)
 	defer stopInTime(t, h)
 	h.HandleMessage(sessionDown())
 	// Wait until PeersDown has begun: the down peer has left some of its
 	// prefixes, which are all IPv4.
 	for h.table.Counts().IPv4Routes == downPeerPrefixes {
 		runtime.Gosched()
 	}
 	returning := peerAnnouncement(testPeerIP, testV6Prefix)
 	if !returnsInTime(func() {
 		for range returningMessages {
 			h.HandleMessage(returning)
 		}
 		h.HandleMessage(peerAnnouncement(secondPeerIP, testV6Prefix))
 	}) {
 		t.Fatalf("the handler did not take the messages within %s", stepTimeout)
 	}
 	if h.table.Counts().IPv4Routes == 0 {
 		t.Error("PeersDown ended before the handler took the other peer's " +
 			"announcement")
 	}
 	checkTableOrigins(t, h.table, map[string]map[int][]string{
 		testV6Prefix: {testASNB: {secondPeerIP}},
 	})
 	if !returnsInTime(h.peerDowns.Wait) {
 		t.Fatalf("the PeersDown did not return within %s", stepTimeout)
 	}
 	checkTableOrigins(t, h.table, map[string]map[int][]string{
 		testV6Prefix: {testASNB: {testPeerIP, secondPeerIP}},
 	})
 }
 // TestPrefixHandlerTakesMessagesWhileItAppliesHeldBackOnes holds back many
 // messages of a peer whose session went down. Each names a prefix that does
 // not parse, so applying it logs an error. Each log write takes a
 // millisecond. While the handler applies the held-back messages, it must take
 // another peer's announcement.
 func TestPrefixHandlerTakesMessagesWhileItAppliesHeldBackOnes(t *testing.T) {
 	t.Parallel()
 	const (
 		// heldMessages take the handler at least a second to apply.
 		heldMessages = 1000
 		badPrefix    = "not a prefix"
 	)
 	logs := &slowWriter{firstWrite: make(chan struct{})}
 	h := newPrefixHandler(newMockStore(), newErrorLogger(logs),
 		routetable.New(routetable.Params{}), longerThanAnyTest,
 		maxTableChangesPerWrite)
 	defer stopInTime(t, h)
 	release := holdUpPeersDown(h)
 	h.HandleMessage(sessionDown())
 	for range heldMessages {
 		h.HandleMessage(peerAnnouncement(testPeerIP, badPrefix))
 	}
 	close(release)
 	select {
 	case <-logs.firstWrite:
 	case <-time.After(stepTimeout):
 		t.Fatal("the handler did not start to apply the held-back messages")
 	}
 	if !returnsInTime(func() {
 		h.HandleMessage(peerAnnouncement(secondPeerIP, otherV4Prefix))
 	}) {
 		t.Fatalf("HandleMessage did not return within %s", stepTimeout)
 	}
 	h.mu.Lock()
 	left := len(h.heldBack[testPeerIP])
 	h.mu.Unlock()
 	if left == 0 {
 		t.Error("the handler took another peer's announcement only once it " +
 			"had applied every held-back message")
 	}
 }
 // TestPrefixHandlerWaitsAtTheMaximumHeldBack holds up the PeersDown of a peer
 // whose session went down. It hands the handler maxHeldBackMessages messages
 // of that peer. HandleMessage must then wait, even for another peer's
 // message, until the PeersDown returns and a held-back message is applied.
 func TestPrefixHandlerWaitsAtTheMaximumHeldBack(t *testing.T) {
 	t.Parallel()
 	h := newPrefixHandler(newMockStore(), logger.New(),
 		routetable.New(routetable.Params{}), longerThanAnyTest,
 		maxTableChangesPerWrite)
 	defer stopInTime(t, h)
 	release := holdUpPeersDown(h)
 	h.HandleMessage(sessionDown())
 	held := peerAnnouncement(testPeerIP, testV6Prefix)
 	for range maxHeldBackMessages {
 		h.HandleMessage(held)
 	}
 	returned := make(chan struct{})
 	go func() {
 		h.HandleMessage(peerAnnouncement(secondPeerIP, otherV4Prefix))
 		close(returned)
 	}()
 	select {
 	case <-returned:
 		t.Error("HandleMessage returned while maxHeldBackMessages messages " +
 			"were held back")
 	case <-time.After(stillWaitingAfter):
 	}
 	close(release)
 	select {
 	case <-returned:
 	case <-time.After(stepTimeout):
 		t.Fatal("HandleMessage did not return after the PeersDown returned")
 	}
 }
 // slowWriter takes a millisecond over each write and drops what it is given.
 // It closes firstWrite at the first write.
 type slowWriter struct {
 	firstWrite chan struct{}
 	once       sync.Once
 }
 func (w *slowWriter) Write(p []byte) (int, error) {
 	w.once.Do(func() { close(w.firstWrite) })
 	time.Sleep(time.Millisecond)
 	return len(p), nil
 }
 // TestPrefixHandlerAppliesAPathWithoutAnOriginToTheRoutingTable hands the
 // prefix handler announcements whose AS path gives no origin. One whose path
 // is empty or holds AS 0 removes the peer's route, as a withdrawal does; one
@@ -554,9 +199,8 @@ func TestPrefixHandlerKeepsThePathOfAPathEndingInASet(t *testing.T) {
 	}
 }
-// checkTableCases hands the messages of each case to a new prefix handler. It
+// checkTableCases hands the messages of each case to a new prefix handler and
-// checks the handler's routing table once the handler's PeersDown calls have
+// checks its routing table right after, with no flush.
 // returned, with no flush.
 func checkTableCases(t *testing.T, tests []tableCase) {
 	t.Helper()
@@ -569,28 +213,17 @@ func checkTableCases(t *testing.T, tests []tableCase) {
 				h.HandleMessage(msg)
 			}
-			h.peerDowns.Wait()
+			for prefix, want := range tt.want {
-			checkTableOrigins(t, h.table, tt.want)
+				got := tableOrigins(h.table, prefix)
 				if !reflect.DeepEqual(got, want) {
 					t.Errorf("peers of %s by origin AS = %v, want %v",
 						prefix, got, want)
 				}
 			}
 		})
 	}
 }
 // checkTableOrigins checks, for each prefix in want, the peers that carry it
 // in table with each origin AS.
 func checkTableOrigins(
 	t *testing.T, table *routetable.Table, want map[string]map[int][]string,
 ) {
 	t.Helper()
 	for prefix, wantOrigins := range want {
 		got := tableOrigins(table, prefix)
 		if !reflect.DeepEqual(got, wantOrigins) {
 			t.Errorf("peers of %s by origin AS = %v, want %v",
 				prefix, got, wantOrigins)
 		}
 	}
 }
 // tableWritesStore is a real database that hands each write of the routing
 // table's changes to written once it is made.
 type tableWritesStore struct {
@@ -865,7 +498,6 @@ func TestARestartReadsTheRoutingTableBack(t *testing.T) {
 		h.HandleMessage(msg)
 	}
 	h.peerDowns.Wait()
 	h.flushBatch()
 	var logs bytes.Buffer
@@ -904,7 +536,6 @@ func TestARestartReadsTheRoutingTableBack(t *testing.T) {
 		restarted.HandleMessage(msg)
 	}
 	h.peerDowns.Wait()
 	restarted.Stop(t.Context())
 	read, err := db.ReadRouteTable(t.Context())
@@ -486,7 +486,6 @@ type HandlerStatsInfo struct {
 	QueueHighWaterMark int     `json:"queue_high_water_mark"`
 	ProcessedCount     uint64  `json:"processed_count"`
 	DroppedCount       uint64  `json:"dropped_count"`
 	HeldBackCount      int     `json:"held_back_count"`
 	AvgProcessTimeMs   float64 `json:"avg_process_time_ms"`
 	MinProcessTimeMs   float64 `json:"min_process_time_ms"`
 	MaxProcessTimeMs   float64 `json:"max_process_time_ms"`
@@ -571,7 +570,6 @@ func handlerStatsInfo(handlerStats []streamer.HandlerStats) []HandlerStatsInfo {
 			QueueHighWaterMark: hs.QueueHighWaterMark,
 			ProcessedCount:     hs.ProcessedCount,
 			DroppedCount:       hs.DroppedCount,
 			HeldBackCount:      hs.HeldBackCount,
 			AvgProcessTimeMs: float64(hs.AvgProcessTime.Microseconds()) /
 				microsecondsPerMillisecond,
 			MinProcessTimeMs: float64(hs.MinProcessTime.Microseconds()) /
@@ -295,43 +295,6 @@ func TestStatsAnswerCarriesChurnPerMinute(t *testing.T) {
 	}
 }
 // heldBackMessages is how many messages heldBackHandler holds back.
 const heldBackMessages = 7
 // heldBackHandler is a stream handler that wants no message and holds back
 // heldBackMessages messages.
 type heldBackHandler struct{}
 func (heldBackHandler) WantsMessage(string) bool           { return false }
 func (heldBackHandler) HandleMessage(*ristypes.RISMessage) {}
 func (heldBackHandler) QueueCapacity() int                 { return 1 }
 func (heldBackHandler) HeldBackCount() int                 { return heldBackMessages }
 // TestStatsAnswerGivesHowManyMessagesAHandlerHoldsBack registers a handler
 // that holds back messages. Its figures in /api/v1/stats, which the status
 // page shows, must give how many.
 func TestStatsAnswerGivesHowManyMessagesAHandlerHoldsBack(t *testing.T) {
 	t.Parallel()
 	s := newServerWithTestRoutes(t)
 	s.streamer.RegisterHandler(heldBackHandler{})
 	var body struct {
 		Data StatsResponse `json:"data"`
 	}
 	err := json.Unmarshal(get(t, s, statsPath).Body.Bytes(), &body)
 	if err != nil {
 		t.Fatalf("decoding the answer: %v", err)
 	}
 	stats := body.Data.HandlerStats
 	if len(stats) != 1 || stats[0].HeldBackCount != heldBackMessages {
 		t.Errorf("handler figures %+v, want one handler holding back %d "+
 			"messages", stats, heldBackMessages)
 	}
 }
 // fixedIntervalASNFetcher is a WHOIS fetcher that does nothing and reports
 // interval as its wait between fetches.
 type fixedIntervalASNFetcher struct {
@@ -261,7 +261,7 @@ func checkAnswerCarriesMeta(
 ) {
 	t.Helper()
-	// The values README "JSON API" gives the two fields that never change.
+	// The values README "API v1" gives the two fields that never change.
 	const (
 		wantTimeZone   = "UTC"
 		wantAPIVersion = 1
@@ -12,6 +12,7 @@ import (
 	"fmt"
 	"io"
 	"math"
 	"math/rand"
 	"net/http"
 	"sync"
 	"sync/atomic"
@@ -57,6 +58,10 @@ const (
 	// Well above an extended BGP message (RFC 8654) in hex plus its prefixes
 	maxLineSize           = bytesPerMB
 	maxConcurrentHandlers = 800 // Maximum number of concurrent message handlers
 	// Backpressure constants
 	backpressureThreshold = 0.5 // Start dropping at 50% queue utilization
 	backpressureSlope     = 2.0 // Slope for linear drop probability increase
 )
 // messageTypeNotification is the type of a BGP NOTIFICATION message, which ends
@@ -88,8 +93,9 @@ type MessageHandler interface {
 	// QueueCapacity returns the desired queue capacity for this handler.
 	// Handlers that process quickly can have larger queues to buffer bursts.
-	// While the queue is full, the streamer drops the handler's messages,
+	// When the queue fills up, messages will be dropped according to the
-	// except STATE messages, which are held until there is room.
+	// backpressure algorithm, except STATE messages, which are held until
 	// there is room.
 	QueueCapacity() int
 }
@@ -120,8 +126,8 @@ type handlerInfo struct {
 // Streamer manages a connection to the RIPE RIS Live streaming API for receiving
 // real-time BGP UPDATE messages. It handles automatic reconnection with exponential
-// backoff and dispatches messages to registered handlers via per-handler queues.
+// backoff, dispatches messages to registered handlers via per-handler queues, and
-// Its read loop never waits for room on a queue (see queueMessage).
+// implements backpressure to prevent queue overflow during high traffic periods.
 type Streamer struct {
 	logger *logger.Logger
 	client *http.Client
@@ -138,6 +144,7 @@ type Streamer struct {
 	workers      sync.WaitGroup // the handler workers; Stop waits for them
 	metrics      *metrics.Tracker
 	totalDropped atomic.Uint64   // Total dropped messages across all handlers
 	random       *rand.Rand      // Random number generator for backpressure drops
 	bgpPeers     map[string]bool // Track active BGP peers by peer IP
 	bgpPeersMu   sync.RWMutex    // Protects bgpPeers map
 }
@@ -161,6 +168,8 @@ func New(logger *logger.Logger, metrics *metrics.Tracker) *Streamer {
 		},
 		handlers: make([]*handlerInfo, 0),
 		metrics:  metrics,
 		//nolint:gosec // Non-cryptographic randomness is fine for backpressure
 		random:   rand.New(rand.NewSource(time.Now().UnixNano())),
 		bgpPeers: make(map[string]bool),
 	}
 }
@@ -342,10 +351,6 @@ type HandlerStats struct {
 	AvgProcessTime     time.Duration
 	MinProcessTime     time.Duration
 	MaxProcessTime     time.Duration
 	// HeldBackCount is how many messages the handler holds back to apply
 	// later: what its HeldBackCount method returns, or 0 for a handler
 	// without one.
 	HeldBackCount int
 }
 // GetHandlerStats returns a snapshot of performance statistics for all registered
@@ -387,11 +392,6 @@ func (s *Streamer) GetHandlerStats() []HandlerStats {
 		info.metrics.mu.Unlock()
 		holder, holdsBack := info.handler.(interface{ HeldBackCount() int })
 		if holdsBack {
 			hs.HeldBackCount = holder.HeldBackCount()
 		}
 		stats = append(stats, hs)
 	}
@@ -399,7 +399,7 @@ func (s *Streamer) GetHandlerStats() []HandlerStats {
 }
 // GetDroppedMessages returns the total number of messages dropped across all
-// handlers because their queue was full. This counter is monotonically
+// handlers due to queue overflow or backpressure. This counter is monotonically
 // increasing and is safe to call concurrently.
 func (s *Streamer) GetDroppedMessages() uint64 {
 	return s.totalDropped.Load()
@@ -882,13 +882,13 @@ func (s *Streamer) dispatchMessage(
 // msg instead. It never waits for room, so a full queue does not hold up the
 // read loop.
 //
-// A message that finds the queue full is dropped, except a STATE message. A
+// A message is dropped while the queue is filling up (shouldDropForBackpressure)
-// STATE message whose state is down is the only notice that all of a peer's
+// or full, except a STATE message: one whose state is down is the only notice
-// routes are gone. A STATE message that finds the queue full is held in
+// that all of a peer's routes are gone. A STATE message skips the random drop,
-// info.heldState instead. Held messages go into the queue, oldest first, when a
+// and one that finds the queue full is held in info.heldState. Held messages
-// later message for the handler finds room, or when Stop is called. While any
+// go into the queue, oldest first, before anything else for the handler, as
-// message is held, the handler's other messages are dropped, so none of them
+// soon as there is room, or when Stop is called; while any is still held, the
-// overtakes a held STATE message.
+// handler's other messages are dropped, so none of them overtakes a held down.
 //
 // The held list stays small: a STATE message is a few hundred bytes and the
 // feed sends about ten a second, so a queue full for a minute holds about 600.
@@ -907,7 +907,7 @@ func (s *Streamer) queueMessage(
 		return true
 	}
-	if len(info.heldState) > 0 {
+	if len(info.heldState) > 0 || s.shouldDropForBackpressure(info) {
 		return false
 	}
@@ -923,3 +923,26 @@ func trySend(queue chan *ristypes.RISMessage, msg *ristypes.RISMessage) bool {
 		return false
 	}
 }
 // shouldDropForBackpressure determines if a message should be dropped based on
 // queue utilization
 func (s *Streamer) shouldDropForBackpressure(info *handlerInfo) bool {
 	// Calculate queue utilization
 	queueLen := len(info.queue)
 	queueCap := cap(info.queue)
 	utilization := float64(queueLen) / float64(queueCap)
 	// No drops below threshold
 	if utilization < backpressureThreshold {
 		return false
 	}
 	// Calculate drop probability (0.0 at threshold, 1.0 at 100% full)
 	dropProbability := (utilization - backpressureThreshold) * backpressureSlope
 	if dropProbability > 1.0 {
 		dropProbability = 1.0
 	}
 	// Random drop based on probability
 	return s.random.Float64() < dropProbability
 }
@@ -5,7 +5,6 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
 	"log/slog"
 	"net/http"
@@ -259,65 +258,6 @@ func TestHeldStateMessagesKeepTheirPlaceInTheQueue(t *testing.T) {
 	}
 }
 // burstQueueCapacity is the size of a burstHandler's queue.
 const burstQueueCapacity = 1000
 // burstHandler wants UPDATE messages and has a queue of burstQueueCapacity.
 // Tests read what it is handed from its queue, so HandleMessage does nothing.
 type burstHandler struct{}
 func (burstHandler) WantsMessage(messageType string) bool {
 	return messageType == ristypes.MessageTypeUpdate
 }
 func (burstHandler) QueueCapacity() int { return burstQueueCapacity }
 func (burstHandler) HandleMessage(*ristypes.RISMessage) {}
 // TestUpdatesDroppedOnlyWhenTheQueueIsFull has the server send one UPDATE more
 // than the handler's queue holds, each with its own timestamp. Nothing takes
 // messages off the queue meanwhile. The streamer must queue the updates in
 // feed order until the queue is full, and drop only the last one.
 func TestUpdatesDroppedOnlyWhenTheQueueIsFull(t *testing.T) {
 	t.Parallel()
 	const (
 		firstTimestamp = 1700000000
 		lineFormat     = `{"type":"ris_message","data":{"type":"UPDATE",` +
 			`"peer":"192.0.2.1","peer_asn":"64496","timestamp":%d}}` + "\n"
 	)
 	var lines strings.Builder
 	for i := range burstQueueCapacity + 1 {
 		fmt.Fprintf(&lines, lineFormat, firstTimestamp+i)
 	}
 	s := New(logger.New(), metrics.New())
 	s.RegisterHandler(burstHandler{})
 	streamLines(t, s, lines.String())
 	dropped := s.GetHandlerStats()[0].DroppedCount
 	if dropped != 1 {
 		t.Errorf("%d updates dropped, want only the one that found the queue "+
 			"full", dropped)
 	}
 	queue := s.handlers[0].queue
 	if len(queue) != burstQueueCapacity {
 		t.Fatalf("%d updates queued, want %d", len(queue), burstQueueCapacity)
 	}
 	for i := range burstQueueCapacity {
 		msg := <-queue
 		if msg.Timestamp != float64(firstTimestamp+i) {
 			t.Fatalf("update %d on the queue has timestamp %.0f, want %d",
 				i, msg.Timestamp, firstTimestamp+i)
 		}
 	}
 }
 // takeAll takes the messages off queue, as a handler's worker would, and sends
 // them on the returned channel once queue is closed.
 func takeAll(queue chan *ristypes.RISMessage) chan []*ristypes.RISMessage {
@@ -541,10 +541,6 @@
                        <span class="metric-label">Dropped</span>
                        <span class="metric-value ${handler.dropped_count > 0 ? 'disconnected' : ''}">${formatNumber(handler.dropped_count)}</span>
                    </div>
                    <div class="metric">
                        <span class="metric-label">Held Back</span>
                        <span class="metric-value">${formatNumber(handler.held_back_count)}</span>
                    </div>
                    <div class="metric">
                        <span class="metric-label">Avg Time</span>
                        <span class="metric-value">${formatProcessingTime(handler.avg_process_time_ms)}</span>