Fires scheduled work by arming a timer for each deadline, rather than polling for records whose deadline has passed.

AshWorkflow.Scheduler.Oban discovers work by asking the data layer which records match, on a cron interval, so a deadline cannot be finer than one minute. This scheduler is told each deadline as the record reaches it and arms a one-shot timer, so fire_after: {30, :seconds} fires 30 seconds later rather than up to 60 seconds after that.

Selecting it

workflow do
  scheduler AshWorkflow.Scheduler.Precise
end

Then start the process that holds the timers, naming the resources it sweeps:

children = [
  {AshWorkflow.Scheduler.Precise.Timeline, resources: [MyApp.Candidate]}
]

transform/3 adds nothing to the resource. This scheduler needs no triggers, no workers and no attributes, which is why a workflow selecting it carries neither the AshOban DSL nor a deadline table.

Two paths, and why both exist

A timer lives in memory, so a node restart forgets every deadline it held. Precision therefore comes from one path and correctness from another.

AshWorkflow.Scheduler.deadline_changed/2 fires the moment a record enters a step, and arms a timer for the exact instant. That is the precise path.

A look-ahead sweep runs every :look_ahead_ms and arms timers for every deadline falling inside the next :horizon_ms. That is the recovery path: it picks up deadlines armed by a node that has since died, deadlines set while this node was down, and state written by anything that did not run the workflow's own actions. It is the same structure a sequencer uses, where a loose scheduling loop places events on an exact timeline.

Both paths converge on AshWorkflow.Scheduler.execute/3, and both re-read the record and re-check the work's match expression before running it, so a timer armed twice or armed stale does nothing the second time.

Options

• :leader — who decides whether this node arms timers. Defaults to AshWorkflow.Scheduler.Leader.Single. See AshWorkflow.Scheduler.Leader.
• :look_ahead_ms — how often the recovery sweep runs. Defaults to 5000.
• :horizon_ms — how far ahead the sweep arms timers. Defaults to 30000.
• :actor, :authorize?, :tenant — passed to AshWorkflow.Scheduler.execute/3.

What this gives up

Oban's durability. A deadline held only as a timer is lost when the node dies, and recovered by the sweep on whichever node leads next, late by at most :look_ahead_ms. A workflow measured in days should stay on AshWorkflow.Scheduler.Oban, which is the default for that reason.