multiple relay support #

hydrant supports connecting to multiple relays simultaneously for firehose ingestion. when RELAY_HOSTS is configured with multiple URLs:

• one independent firehose stream loop is spawned per relay
• each relay maintains its own firehose cursor state
• all ingestion loops share the same worker pool and database

commit events are de-duplicated according to the repo rev. account / identity events are de-duplicated using the time field.

a relay build also seeds sources from SEED_HOSTS (https://bsky.network by default, see firehose seeding), which adds a firehose per PDS on top of RELAY_HOSTS. to follow only the relays you listed, set HYDRANT_SEED_HOSTS= (empty).

direct PDS connections #

a firehose source can also be a direct connection to a PDS rather than a relay. prefix the URL with pds:: to mark it as such:

HYDRANT_RELAY_HOSTS=wss://bsky.network,pds::wss://pds.example.com

only when a source is marked as a direct PDS (is_pds: true), hydrant enforces host authority. relays (is_pds: false, the default) are exempt from this check, since they forward commits from many PDSes by design.

firehose seeding #

in relay mode, RELAY_HOSTS defaults to empty. set SEED_HOSTS to one or more relay base URLs and hydrant will call com.atproto.sync.listHosts on each at startup and every 30 minutes thereafter, adding every returned PDS as a firehose source:

HYDRANT_SEED_HOSTS=https://bsky.network

seeding runs as a background task so the main firehose loop is not blocked. seed URLs are fetched concurrently (up to four at a time) and the full listHosts pagination is consumed for each. if a request fails partway through, the hosts collected so far are still added and the failure is logged.

each discovered host is added as a persistent PDS firehose source (is_pds: true), equivalent to calling POST /firehose/sources.

banned hosts (status: "banned") are skipped. all other statuses are included since the firehose ingestor retries on disconnect and transiently-unavailable hosts will reconnect on their own.

a host we have no cursor for starts from the seq the seed relay reported for it, or from the live head when it reported none, so its retained history is not replayed. hosts that already have a cursor keep it.

in indexer mode that live tail would miss repos that already lived on a new PDS, so a host that is new to the network also gets one listRepos walk, its repos going through the same signal check as crawler sources. a seed's hosts only count as new once its listHosts has been read to the end once, so a fresh node doesn't walk every PDS it seeds. a PDS admitted through com.atproto.sync.requestCrawl always gets the walk. it runs while the crawler is enabled, resumes from its last page after a restart, and is dropped if the host is banned or removed as a source.

seeding runs from latest cursor on restart so new PDS' added to the upstream relay since the last start are picked up automatically (if they haven't through firehose). sources that are already running are detected and skipped, so re-seeding is idempotent.

requestCrawl #

com.atproto.sync.requestCrawl asks hydrant to follow a host. the hostname is a bare hostname, since one with a port is refused, so hydrant only follows and walks a host on the default port (wss for the firehose, https for listRepos). what it does depends on what hydrant already knows about it:

• a banned host gets HostBanned.
• a host hydrant doesn't follow is added as a PDS firehose source, the same as POST /firehose/sources. it counts against NEW_HOST_LIMIT, and once that's used up for the UTC day the request gets RateLimitExceeded. in indexer mode it also gets the listRepos walk above.
• a host hydrant follows whose firehose is connected or still retrying is left as it is.
• a host whose firehose hydrant gave up on is restarted right away, after the usual startup jitter, with its failure count reset, instead of waiting for OFFLINE_HOST_RETRY_INTERVAL. a PDS that was down for an upgrade can ask to be picked up again this way.

in indexer mode a host hydrant follows also gets walked again when it asks, connected or not. a rewalk adds the repos the host lists that we've never seen, like the first walk. it also resyncs the ones whose doc points at the host and that we hold desynchronized or with a failed backfill waiting on its retry, and the ones the host lists at a newer rev than ours. the spec doesn't define that last one (the IETF drafts don't cover crawling, and the lexicon only says to "persistently crawl hosted repos"). it's how hydrant lets a host without a firehose say it changed. a newer rev costs the repo one of its refetches, see below. the stuck ones are the same resync POST /repos/resync does, so a repo keeps its retry count. a host's rewalks may read REWALK_MB (8 by default) of listRepos responses per REWALK_WINDOW (30 minutes by default), counted from the first page one reads and kept across restarts, so a host can ask as often as it likes while it stays inside that. a walk charges each page's size once it has read it, so it can overshoot by a page, and stops before the next one when the bytes are spent. it keeps its cursor and goes on when the window ends, also after a restart. a request while its walk is queued, running or waiting, or once the bytes are spent, gets the usual 200 and does nothing. every listRepos walk, a new host's included, refuses a page over 512 KiB or with more repos than the 1000 we asked for, and ends with an error when a page returns a cursor an earlier page of the walk already returned. a rewalk doesn't count against NEW_HOST_LIMIT, isn't given to a banned host, and leaves the host's backoff alone: it's paced and throttled like any other walk.

refetch budgets #

a full refetch is the cheapest thing a flaky or hostile peer can ask of hydrant: a #sync at a fresh rev, a chain break, a listRepos listing at a newer rev, a crawl's resync. each one costs the repo a refetch out of REPO_REFETCH_LIMIT per REPO_REFETCH_WINDOW (6 an hour), counted from the repo's first refetch in the window. the repos of one host also share HOST_REFETCH_LIMIT (60) in the same window. hosts on the built-in trusted tier are exempt from that one, because an operator who trusts a host already accepts its volume, but their repos keep their own limit.

a repo that's out of refetches isn't refetched. it goes throttled (an inactive account keeps its status) with a resync entry of kind ratelimited whose next_retry is the end of the window, and it's refetched then. commits for it that arrive meanwhile are dropped because the refetch replaces what they change. a #sync that doesn't move a repo's rev forward is ignored, and a listing dated more than REV_CLOCK_SKEW past hydrant's clock isn't taken for news. retries of a refetch that was admitted and failed don't cost another, and neither does POST /repos/resync, because the repo's backoff or an operator bounds those. /stats counts the refetches denied since start as counts.refetch_denied.

crawler sources #

the crawler is configured separately from the firehose via CRAWLER_URLS. each source is a [mode::]url entry where the mode prefix is optional and defaults to by_collection in filter mode or list_repos in full-network mode.

• list_repos: enumerates the relay via com.atproto.sync.listRepos. relay account status is reconciled for repos already known to hydrant, but a relay listing an active repo as inactive only queues a fetch from its pds (once per listed status), and the account goes inactive only if the pds refuses it, so a relay-level takedown of a repo its pds still serves leaves it synced. the same goes the other way: a relay listing an inactive repo as active at the revision hydrant has only queues a fetch, and the account comes back only if the pds serves it; filtered indexes check unknown repos with describeRepo before enqueueing them. a listing whose revision is newer than the one hydrant stored queues a backfill only once hydrant's own firehose of that relay has caught up to when the page was read (the slowest relay firehose for a host hydrant doesn't subscribe to), since a commit still on its way through hydrant would look the same. the crawl keeps reading pages meanwhile, with up to 32 of them waiting, and a page's cursor only moves once it and every page before it are in. if the firehose doesn't get there within 10 minutes, those listings are skipped until the next pass, with one log line.
• relay_first: an explicit coverage-over-moderation opt-in. it completes the relay's authoritative listRepos pass first, then calls the same relay's listHosts and crawls listRepos on each safe, non-banned PDS. relay-listed inactive repos are retained as status tombstones, so deep pages can only add DIDs the relay did not list and never update relay-authoritative status or revision state. this may discover inactive or other repositories omitted by the relay, which also means it deliberately bypasses relay-level repository moderation for those unknown DIDs. local PDS bans and filter excludes still win.
• by_collection: queries com.atproto.sync.listReposByCollection for each configured signal. more efficient for filtered indexing since it only surfaces repos that have matching records. cursors are stored per collection. note that it won't crawl anything if no signals are specified.

CRAWLER_URLS=by_collection::https://lightrail.microcosm.blue,list_repos::wss://bsky.network

for deep discovery in a full-network indexer:

CRAWLER_URLS=relay_first::https://relay.fire.hose.cam

relay_first processes at most four PDSes concurrently. each PDS request uses the shared per-PDS concurrency gate and rate-tier pacing, holds its permit through the bounded response read, and feeds 429/hard failures back into the normal PDS throttler. unsafe/private DNS targets, locally or upstream-banned hosts, offline hosts, and PDSes without listRepos support are skipped. relay-pass, listHosts, and per-PDS cursors are durable; a restart resumes the last committed page, and /crawler/cursors clears all of them for the source.

each source maintains its own cursor so restarts resume mid-pass.

sources can also be added and removed at runtime via the /crawler/sources API (see crawler management). dynamically added sources are persisted to the database and survive restarts. CRAWLER_URLS sources are startup-only: they are not written to the database and will always reappear after a restart regardless of runtime changes (unless you change the config of course).