12 Commits

Author	SHA1	Message	Date
Niklas Ye	60ebb75cd2	Show notes from similar earlier incidents ... Each incident gets a signature: the alert name plus the group labels that say what is broken, minus the ones that only say where it ran (instance, pod, container, ...). GET /api/incidents/{id}/similar returns resolved incidents in the same team with the same signature that have notes. Notes can be marked as the resolution note, "what fixed it", either with a resolution field on resolve or pinned on a note. Those lead the similar list, show on the incident page as "Seen before", and the triggered notification carries the latest one. Claude-Session: https://claude.ai/code/session_01MMados3BD1oSjevHxbmVqU	2026-09-25 15:42:25 +02:00
Niklas Ye	3183e7e5c5	Page the next person when nobody answers ... Closes #6, and closes the thing this whole line of work was opened for. Until now an unacknowledged incident re-paged the same topic every notify_repeat forever, which is a louder version of the same silence: if the person on call is asleep, out of signal or has left the company, nothing else happened. A team can now configure an ordered ladder. Each level has a timeout and a set of targets; a target is a named person or whoever the team's rota says is on call today. That second kind is the one that keeps working when the rota changes and nobody remembers to edit the policy. When a level's timeout passes with the incident still triggered, the next level is paged; off the end the chain repeats repeat_count times and then the team's fallback topic is paged once. The incident stays open throughout, because running out of people to wake is not somebody answering. Escalation rides the notifier's existing 30-second tick and its outbox rather than adding a second scheduler, and runs before delivery so a level that comes due on a tick is paged on that tick. Each target gets its own outbox row and therefore its own Acknowledge token: the button in a notification must acknowledge as the person holding the phone, not as whoever was paged first. Acknowledging or resolving takes the incident off the ladder. Snoozing pauses it -- a deliberate "not now" holds the ladder where it is and it resumes when the snooze runs out, rather than carrying on without the person who asked for quiet. Reminders and escalation never both run. A team with a ladder gets escalation; a team without keeps today's behaviour exactly. Both would mean two pages for one silence, which is how a tool gets muted. A level whose targets cannot be reached -- no topic, a disabled account, an empty rota -- is entered anyway, recorded as "nobody reachable", and the ladder moves on. Stalling on a rung that cannot ring would be the failure this feature exists to prevent, wearing the feature's clothes. A policy with such a level cannot be created, but an older row could hold one. The API replaces the ladder wholesale rather than patching a rung, because the levels are an order: editing one has to answer what happens to the numbering of the others, and a whole-ladder PUT makes that the client's decision and the edit atomic. Verified against a live server as well as in tests: alice paged, nobody answers, bob paged, nobody answers, the fallback topic paged once and the timeline reading "level 2: bob" then "escalation exhausted: paged terdut-oncall-all" -- and a second incident acknowledged before its timeout, which woke nobody else. No UI yet. The team-settings screens for escalation, integrations and dead man's switches are all still missing, and they are one piece of work rather than three. Claude-Session: https://claude.ai/code/session_01RHPj4ggeFdEjKKfm4SHbD7	2026-09-20 18:37:54 +02:00
Niklas Ye	7c87ae2af8	Remove the unauthenticated webhook and the SQLite migration script ... Both existed to carry an upgrade across, and both upgrades are done. /api/alertmanager/webhook took no credential at all: anything able to reach the port could open an incident for anybody. v0.12.0 kept it, deprecated, so the teams release did not stop delivery while the Alertmanager config was edited, and logged a line per payload asking to be moved. The cluster's Alertmanager now posts on an integration key -- verified in the log, every two minutes, with no deprecation line since the rollout -- so the door can be shut rather than left ajar until somebody remembers. A sender still posting there gets the JSON 404 every unknown /api path gets. The tests move with it, which they should have done anyway: the harness mints an integration key for the default team and posts on that, so they exercise the path production uses rather than one only they still used. scripts/sqlite-to-postgres.go goes the same way. It was written to be temporary, it was the last thing needing modernc.org/sqlite, and this install migrated on 2026-09-20. `go mod tidy` drops the driver and its six transitive dependencies with it; the module graph is now chi, pgx, pgerrcode and x/crypto. Anyone still on v0.10.x can take the script out of the v0.12.0 tag, which the README now says. Claude-Session: https://claude.ai/code/session_01RHPj4ggeFdEjKKfm4SHbD7	2026-09-20 18:11:30 +02:00
Niklas Ye	74359c72ab	Give each team its own dead man's switches, and the UI a team to show ... The rest of #4. Two halves that belong together because they are the same sentence from opposite ends: a team decides which of its alerts are heartbeats, and the UI has to be able to say which team it is talking about. Switches were three environment variables, which made them one setting for the whole install. That was the last piece of the alerting path a team could not control: it could take its own alerts on its own key and still not say which of them were heartbeats, or how long a silence had to last. They are a row per team now, edited by an owner through PUT /api/teams/{teamID}/deadman, and the sweeper runs each team against its own matchers, timeout and severity. The environment variables become the starting point rather than the setting. Every team without a configuration is seeded from them at startup, so an upgrade keeps watching exactly what it was watching, and SeedDeadmanConfigs never overwrites -- a redeploy must not put the environment's value back over an owner's edit. A team created later watches nothing until somebody says otherwise: inheriting an install-wide heartbeat would page a new team about a source it has never heard of, and a switch nobody chose is the kind that gets muted rather than fixed. A matcher string with no alertname in it is refused at the door instead of stored. Storing it would produce a switch that watches nothing silently, which is the exact failure the feature exists to prevent. NewRouter and Sweep lose their DeadmanConfig parameter -- there is no longer one answer to hand them. The type stays, because parsing a matcher string is still parsing a matcher string. The UI side: rows in the queue carry a team badge, the filter row gains a team chip per team, and "on call now" shows one card per team. All three appear only when the viewer is in more than one team -- otherwise they are the same word repeated down a list, which is noise rather than information, and the single-team install reads exactly as it did before teams existed. Verified against a live two-team server as well as in tests: the combined queue labelled by team, the team_id filter, a heartbeat that is a heartbeat in one team and an ordinary alert in another, and a new team's switches starting empty while the upgraded team keeps the environment's. Claude-Session: https://claude.ai/code/session_01RHPj4ggeFdEjKKfm4SHbD7	2026-09-20 15:18:22 +02:00
Niklas Ye	a4fbd60441	Scope everything to a team, and route alerts by integration key ... The core of #4, and what #1 is for: terdut stops being one shared space. A team owns its incidents, alerts, schedule and integrations; a user sees exactly the teams they are in. Everything that existed moves into one Default team and every existing user becomes an owner of it, so the upgrade is a no-op for the people using it. Ingestion is the load-bearing half. An alert arrives on a team's integration key, and the key is both the credential and the routing: it says that the sender may post, and which team the alerts belong to. That also closes the unauthenticated webhook -- the old path stays for one release, deprecated and routed to the oldest team, so an upgrade does not stop delivering while somebody edits the Alertmanager config. Scoping is enforced in as few places as possible, because the failure mode is silent. serveAs loads the caller's memberships once; list queries carry `team_id = ANY(...)`; and every incident route goes through incidentIDParam, which now parses the id AND checks the team in the same call, so a new handler cannot remember the first half and forget the second. Anything in another team is 404, never 403: whether an incident exists is that team's business. Two bugs this found, both of which would have been silent: * upsertAlerts decided "is this a new occurrence" by looking up the fingerprint alone. Across teams that made team B's first alert look like a re-send of team A's, so it opened no incident at all. The lookups are keyed on (team_id, fingerprint) now, as the index is. * Every uniqueness rule was written for one tenant. Two teams watching two clusters legitimately see the same fingerprint, the same groupKey, and want somebody on call on the same day; all three constraints move to include team_id. Roles inside a team are separate from the system administrator flag: an owner configures the team, a member works its incidents, and an admin is NOT implicitly in every team -- administration is about accounts, not about reading other people's incidents. An admin can still repair a team whose owner has left, which is why requireTeamOwner lets them through. A shift can only be given to somebody in the team. Paging a person who cannot open the incident is worse than paging nobody. The UI is updated only as far as keeping it working: it loads the viewer's teams with the session and uses the first one, since nobody has a second yet. "On call now" shows every team the viewer is in, named only when there is more than one, so the common case reads exactly as before. The team switcher, badges and per-team settings pages are the next step. Breaking for API clients: the schedule endpoints moved under the team, and /api/schedule/current returns an array rather than an object or a 404. terdut-tui will need a version for that. Per-team dead-man configuration is deliberately not here. A heartbeat's incident already opens in the team whose key received it, which is the part that matters for isolation; moving the matchers out of env into per-team rows is a change to how deadman.go is configured rather than to who sees what. Claude-Session: https://claude.ai/code/session_01RHPj4ggeFdEjKKfm4SHbD7	2026-09-20 13:36:24 +02:00
Niklas Ye	dc39e3a5d3	Move the database to Postgres, before teams need the schema ... First step of #1, and it goes first for one reason: #4 adds a team_id to nearly every table, and doing that twice -- once for SQLite, once for Postgres -- is work nobody gets paid for. The teams migrations now only have to be written against one database. The ten SQLite migrations are replaced by a single Postgres baseline rather than ported one by one. They were incremental in a way that has no value on a fresh install: 004 adds columns 008 drops again, and 008's backfill rewrites data a Postgres database never had. The history stays in git; the schema they add up to is now 001_baseline.sql. Timestamps stay BIGINT unix seconds and are NOT converted to timestamptz. Everything in Go already speaks epochs, so converting would have been a second, larger change riding along inside this one. It is worth doing on its own. The JSON columns did move to jsonb, because #4 will want to filter and index on labels. Most of the port is mechanical -- 170 placeholders from ? to $1 -- but four things needed more than a search and replace: * Dynamically built WHERE clauses cannot keep their numbering straight by hand, so they hand out placeholders through sqlArgs instead. A filter can now be added or reordered without renumbering anything. * SUM(resolved_at IS NULL) was SQLite counting a boolean as 0 or 1. Postgres has no sum(boolean), and this was breaking every dead man's switch -- silently, since the sweeper only logs. Now COUNT(*) FILTER. * unixepoch() became FLOOR(EXTRACT(EPOCH FROM now()))::bigint. The FLOOR is load-bearing: a bare cast rounds half up, so a row written at .6 of a second claimed a timestamp a second in the future and disagreed with the time.Now().Unix() the Go side stamps. * The unique-violation check matched SQLite's error text. It matches SQLSTATE 23505 now, so a renamed constraint cannot turn a 409 back into a 500. Tests need a real Postgres, because there is no in-memory Postgres the way there was an in-memory SQLite. Each test gets its own schema on a shared server -- cheaper than a database each, and still isolated. TERDUT_TEST_DSN says where it is; `make test-db` starts one locally and ci.yaml runs one as a service container. An unset DSN fails the suite rather than skipping it: a run that quietly tests nothing is worse than one that does not run. TestMigration_BackfillCarriesAckAndComments is deleted along with the migrations it replayed. What it protected -- an upgrade not losing acknowledgements and comments -- now belongs to scripts/sqlite-to-postgres.go, which is build-tagged so the SQLite driver stays out of the server binary. Both are meant to be deleted once this install has migrated. The chart loses the PVC, the data volume and the python backup sidecar, and requires database.dsnSecret.name: it provisions no database and cannot guess where the credentials live, so a render without it is meant to fail. Backups move to where Postgres actually runs. The other half of that -- the postgresql CR, the k8up pg_dump annotation and the network policy -- is a change to the wrapper chart in Ryuvia/charts and is not in here. Verified rather than assumed: the gate is green with -race against Postgres 17, govulncheck and gitleaks are clean, and the migration script was run end to end against a SQLite database built at the old schema and seeded in every table. Ids survive, so incidents keep their numbers and every foreign key still points where it did; the identity sequences are moved past the copied ids, and a webhook after the migration opened incident 12 rather than colliding at 1.	2026-09-20 10:44:12 +02:00
Niklas Ye	14c24f8fda	Notice when the Watchdog alert stops arriving ... Everything this server does assumes alerts arrive. If Prometheus stops evaluating, or Alertmanager cannot reach us, nothing arrives — and silence is indistinguishable from everything being fine. The cluster has shipped the alert for exactly this case all along: Watchdog is expr: vector(1), so it fires permanently and is re-sent forever, and it is worth nothing unless something downstream notices it stop. Nothing did. It arrived, opened no incident because a repeat_interval re-send is not a new occurrence, and when the monitoring stack died the sweeper quietly expired it and paged nobody. So the handling is inverted for a configurable set of alerts: receiving one opens no incident, and the absence of one does. TERDUT_DEADMAN_MATCHERS selects them as label matchers, defaulting to alertname=Watchdog. The unit of monitoring is the fingerprint rather than the alert name. Two clusters sending the same Watchdog are two independent switches, so a healthy one can never mask a dead one. Every matcher must name an alertname, which keeps the sweeper's candidate query on alerts_name_idx instead of JSON-extracting labels from every row, and leaves matching with a single implementation. A switch is dormant until its first heartbeat: a matcher nothing has ever sent opens nothing, so a fresh deploy or a restored database does not page. Resolving the incident by hand sticks, exactly as it does for an alert-backed one, so a decommissioned source is a one-time page rather than a nag; the switch re-arms only when the heartbeat comes back, and dying again is a new incident. The incident has no member alerts on purpose. Linking the heartbeat would have the settled-incident cascade close it on the very sweep that opened it, and there is no alert describing the problem anyway — the problem is that no alert arrived. What happened is on the timeline instead, and recovery is the only automatic way out. One narrow exemption in the ingest guard makes recovery possible at all. A heartbeat we declared dead is marked resolved, and the one that proves us wrong carries the unchanged startsAt of an alert that never stopped firing — so "resolution is terminal within an instance" would discard it forever and a switch could die exactly once. The exemption is scoped to resolution_source = 'deadman', which is the only resolution this server infers from silence on a timeout of its own, so nothing another writer set can be undone by a stale retry. Matched alerts are also held back from the generic staleness expiry, which would otherwise resolve a heartbeat as 'expiry' long before its own tighter deadline. The timeout points the opposite way to TERDUT_STALE_AFTER: staleness is a generous grace period around a repeat_interval you do not control, while this is a deadline you set deliberately and configure the heartbeat's route to beat. Inheriting a 4h or 12h repeat_interval gives a dead man's switch with a twelve hour fuse, so the README spells out the route the heartbeat needs.	2026-08-08 21:28:55 +02:00
Niklas Ye	bc285799d1	Page the on-call person when an incident opens ... An incident opened, got assigned to whoever held today's schedule entry, and then sat there silently until somebody thought to look. The schedule and the incident model were both built; nothing reached the person holding the pager. Notifications go out through ntfy, over plain HTTP with no new dependencies. Delivery is an outbox rather than an inline call: the pool is limited to a single connection, so a POST made while holding the webhook's transaction would stall every other request behind it. The webhook inserts a row and a notifier goroutine sends it within a tick, retrying with exponential backoff. Only opening an incident has to resolve a topic from scratch. Reminders and all-clears reuse whatever that first notification chose, which keeps configuration out of resolveIfSettled and gives the right rule for free: you only hear that something resolved if you were told it started. Each push carries an Acknowledge button, because the useful thing to do at 3am is stop the pager without unlocking anything. It POSTs to an unauthenticated /api/notify/ack/{token} — a notification body lives on the ntfy server and in the device cache, so a real API key must never appear in one. The token is minted per delivery, scoped to one incident and one action, and expires in a day. Reminders repeat until the incident stops being untouched. The stop conditions are the states that already mean somebody has it: acknowledged, snoozed, resolved, archived. Snooze is the mute button, so there is no separate reminder cap. Notifications sent to the fallback topic carry no Acknowledge button. The topic is shared, and a button on it would let any subscriber acknowledge as somebody else.	2026-08-07 08:51:38 +02:00
Niklas Ye	279ef6cf8b	Turn incoming alerts into incidents ... The alerts row was both Alertmanager's record and the human work queue, and the two have different owners. The webhook upsert rewrites that row on every notification; acknowledgement, comments and archiving were columns on it that the upsert happened not to touch. So an alert that resolved and re-fired days later still read as acknowledged by whoever acked the first occurrence — the ack outlived the thing it referred to. Nothing recorded transitions either: rows are mutated in place, so there was no timeline and no way to compute how long anything took. Alerts are now read-only signal records with two states, and incidents are the work item: triggered, acknowledged or resolved, with an assignee, a snooze, notes and an append-only timeline. Many alerts map to one incident, and a new occurrence opens a new incident, which is what makes a stale ack impossible rather than merely unlikely. Correlation uses Alertmanager's own groupKey. It already grouped the alerts according to the group_by routing tree the operator configured and sends the result on every webhook, where it was being discarded; adopting it means changing group_by in alertmanager.yml changes correlation here, with no second grouping scheme to configure and keep in sync. An incident opens only when an alert transitions into firing — an unseen fingerprint, a newer startsAt, or a resolved alert starting again. The unchanged notifications Alertmanager re-sends every repeat_interval are none of those. That rule is what lets manual resolution be terminal: without it, closing an incident by hand would be undone by the next re-send of an alert that never stopped firing, and the button would be a lie. Snooze covers the "not now" case instead. Incidents otherwise resolve by cascade, once every alert under them has stopped firing, whether by webhook or by expiry. New incidents are assigned to whoever holds today's schedule entry. The schedule table has existed since the first release with nothing reading it. Also here, following from the split: - Incident severity is a high-water mark over its alerts, never lowered. An incident that hit critical was a critical incident, and downgrading a live one would demote it in the queue while the work is still open. - /api/stats/incidents reports MTTA and MTTR, null rather than zero until there is something to average. Neither was computable before. - Alert archiving becomes sweeper-only housekeeping; the archive people interact with is the incident's. Breaking: the alert acknowledge, archive and comment endpoints are gone, and the alert object drops the acknowledgement fields and gains incident_id. The README maps each removed endpoint to its replacement. Migration 008 backfills an incident per existing alert, archived ones included so no comment is orphaned, carrying acknowledgements across and turning comments into timeline notes. Both documented alert contracts are untouched: received_at still advances on every accepted payload, re-sends included, and resolution_source still says how much to trust ends_at. The upsert is byte-for-byte what it was, now running inside the ingest transaction.	2026-07-30 17:02:13 +02:00
Niklas Ye	a602ff3efc	Document received_at and resolution_source as public contract ... The API reference listed endpoints but never the alert object's fields, so two of them were load-bearing for clients while being described nowhere. received_at appeared only in passing, as a stats filter; resolution_source only inside the stale-expiry prose. Both carry meaning a client cannot derive on its own. starts_at comes from Prometheus and never changes for an alert instance, so received_at is the only signal that a firing alert is still being refreshed — it advances on every accepted webhook, including the unchanged notifications Alertmanager re-sends every repeat_interval. resolution_source then says how much to trust ends_at: under 'alertmanager' it is an end time somebody reported, but under 'expiry' nothing ever reported one, so it is either a stale watermark or the sweep timestamp, and only an upper bound. README gains an alert object field table plus a contract section for each, including the nullability rules and the advice to tolerate unrecognised resolution_source values. The field comments in models.Alert now say these are public API rather than ingest details, and the upsert carries a note at the received_at line, which is where a regression would be introduced. Three tests lock the newly documented behaviour, none of which was covered before — the whole suite passed with the received_at bump deleted from the upsert, because the expiry tests only ever set that column via SQL: - a re-send advances received_at and leaves starts_at alone - a discarded out-of-order retry does not count as a heartbeat - an expiry resolve preserves a reported ends_at watermark and stamps sweep time only when none was known	2026-07-30 09:02:44 +02:00
Niklas Ye	42e846f876	Expire stale firing alerts ... A resolved webhook was the only path out of the firing state, so a notification that was dropped, silenced, or lost to a restart pinned an alert as firing forever — Prometheus showed it resolved while terdut-server kept listing it. The archiver only ever touched resolved alerts, and both the list and stats queries compared status with plain equality, so a stale row was indistinguishable from a live one. A sweeper pass now resolves firing alerts on either of two signals: the ends_at watermark Alertmanager sets on outgoing firing notifications has passed (plus a grace period for clock skew), or no webhook has refreshed the alert within TERDUT_STALE_AFTER (default 6h, above Alertmanager's 4h repeat_interval). Such alerts get resolution_source = 'expiry', distinguishing them from a real 'alertmanager' resolve. Two related webhook bugs fixed alongside: - The upsert had no ordering guard, so a retried firing notification arriving after the resolved one resurrected the alert. Payloads for an older alert instance are now discarded: a stale retry carries the same startsAt, a genuine re-fire a newer one. - archived_at was never cleared on re-fire, leaving a re-fired alert archived and invisible in the default list. Stats now exclude archived alerts to match the default list view; this lowers historical firing/resolved totals. The chart exposes both sweeper durations via sweeper.staleAfter and sweeper.archiveAfter.	2026-07-28 11:49:39 +02:00
Niklas Ye	9b4ca1482f	Stage 3: Alertmanager webhook ingestion and alert query API ... - Migration 003: alerts table with fingerprint UNIQUE, JSON label/annotation columns, nullable ends_at, and indexed status/name/received_at - POST /api/alertmanager/webhook — upserts each alert by fingerprint; zero endsAt ("0001-01-01") stored as NULL (still firing) - GET /api/alerts — filtered list (?status, ?name, ?from, ?to, ?limit) - GET /api/alerts/{id} — single alert lookup	2026-05-20 21:54:17 +02:00