specs/085-scheduler-ha.md

Rama main · réplica de lectura del repositorio original.

# Spec: Scheduler High Availability

## Status

Partially implemented. Conformance covers roles, takeover, role reporting,
missed slots, and queued runs. Further sections are added as the behavior they
describe lands.

## Scope

This spec defines how several `dagu scheduler` processes sharing one data
directory choose an active instance, how a standby takes over, and how an
operator tells them apart. It does not define the availability of the shared
storage, the execution of steps, or distributed worker behavior.

## Goal

An operator can run a standby scheduler on a second host and know, without
reading logs, which instance is active and how long a failover takes.

## Behavior

### Roles

- Schedulers that use the same data directory compete for one lock. The
  instance holding it is `active`. Every other instance is `standby`.
- Only the active instance schedules runs and consumes queues.
- A standby keeps waiting for as long as the active instance renews the lock.
- A standby keeps waiting through storage errors once it has started waiting.
  An error on its first attempt at the lock ends startup, so a wrong data
  directory fails at once.
- The active instance renews the lock every `I`, the smaller of 7 seconds and
  a quarter of the stale threshold `T`. It keeps running through renewal errors,
  and stops being active once it has gone `T - I` without a renewal or finds
  the lock is no longer its own. A renewal that blocks on storage counts as
  no renewal.
- An instance past that limit dispatches nothing, even before it stops.
- An instance that stops being active this way exits with a non-zero status,
  so a supervisor can restart it. `dagu start-all` exits too, rather than
  leaving its web UI running without a scheduler. A process that cannot finish
  stopping within 10 seconds is ended.

### Takeover

- The active instance releases the lock when it stops cleanly. A standby then
  becomes active within the lock retry interval plus startup.
- When the active instance dies without releasing the lock, a standby becomes
  active after the lock goes stale. With a stale threshold `T` and a retry
  interval `R`, this happens within `T + 2R` plus startup of the death.

- A standby judges staleness by watching the active instance renew the lock,
  timed on its own clock. Clock differences between hosts do not affect it. A
  standby therefore waits a full `T` even for a lock left behind long ago, so
  a host restarted after a crash starts scheduling `T` after it comes up.

### Run identity

- Each slot of a recurring schedule has one run ID, derived from the data
  directory, the DAG name, and the slot's time, in the form
  `sched-<hash>-<UTC time>`. A live run and a catch-up of the same slot share
  it, so a slot has at most one run, including around a failover. Installations
  with different data directories never share an ID.
- When existing runs cannot be read, a run is not created.

### Runs in progress

- A run keeps going when the scheduler that started it stops or fails, whether
  it was started directly or from a queue.
- The exception is `signal_handling.enable_propagation`. With it, a stopping
  scheduler forwards SIGINT and SIGTERM to the runs it started, which stops
  them.

### Missed slots

- The first tick of a new active scheduler runs the slots of the current
  minute.
- A scheduler that takes over a stale lock also replays earlier slots of
  recurring schedules that fell between the failed scheduler's last tick and
  the takeover, as far back as `scheduler.failover_recovery_window`. A slot the
  failed scheduler planned but did not start is replayed too.
- Replayed runs go through the queue, have the catch-up trigger type, and
  follow the DAG's `overlap_policy`. Nothing is replayed when queues are
  disabled.
- A scheduler that acquires a released lock replays only what the DAG's own
  `catchup_window` asks for.

### Observability

- Every scheduler serves `GET /health` on `scheduler.port`, including a
  standby. The response is `200` with `status` and `role`, where `role` is
  `active` or `standby`.
- `GET /health?role=active` answers `503` on an instance that is not active,
  so a load balancer or probe can select the active instance.
- The server's metrics include `dagu_scheduler_instances{status}`, the number of
  instances with `status` `active` or `standby`.

## Configuration

| Key | Default | Meaning |
| --- | --- | --- |
| `scheduler.lock_stale_threshold` | `30s` | Age at which a lock can be taken over |
| `scheduler.lock_retry_interval` | `5s` | How often a standby tries the lock |
| `scheduler.failover_recovery_window` | `2m` | How far back a takeover replays missed slots; `0s` disables it |
| `scheduler.port` | `8090` | Port of the health endpoint; `0` disables it |

## Storage

All schedulers must use the same `DAGU_HOME`, or the same data and DAG
directories, on storage that every host can reach. The data directory must have
the same path on every host, because it is part of each slot's run ID.

## Not provided

- Availability of the shared storage itself.
- Exactly-once execution of step side effects, or resuming a step after a
  crash. Runs on a failed host are marked failed.

## Conformance

`conformance/spec085_scheduler_ha` runs the real binary. It asserts that one
of two schedulers is active and stays so past the stale threshold, that a
standby takes over within the bound after the active process is killed, that
a clean stop hands over without waiting for staleness, that losing the lock
ends `dagu scheduler` and `dagu start-all` with a failure, that a slot that
passes with no active scheduler runs once after a later takeover, that a slot
runs once when its scheduler is killed as it comes due, that a scheduler that
stalls past the stale threshold exits with a failure when it resumes while the
slot it missed runs once on the new active scheduler, and that a queued run
finishes after the scheduler that started it stops.