Lifecycle hooks¶
Lifecycle hooks let application code react to the points in a job's life — when it starts, finishes, retries, parks on an external callback, and so on — without putting that logic inside the job handler itself. Typical uses are metrics, alerting, and job-type-specific cleanup on permanent failure.
Hooks are registered on the ClientBuilder and keyed by job kind. They work for both the typed (register::<T, _>()) and raw (register_worker(...)) worker paths.
Registering a hook¶
use awa::{Client, JobEvent, QueueConfig};
let client = Client::builder(pool)
.queue("email", QueueConfig::default())
.register::<SendEmail, _, _>(|args, _ctx| async move { /* ... */ Ok(awa::JobResult::Completed) })
// Typed hook: args are deserialized from the committed job snapshot.
.on_event::<SendEmail, _, _>(|event| async move {
match event {
JobEvent::Started { job, .. } => metrics::started(&job.queue),
JobEvent::Completed { duration, .. } => metrics::completed(duration),
JobEvent::Exhausted { error, .. } => alert::job_failed(&error),
_ => {}
}
})
.build()?;
on_event::<T>() gives you typed args; on_event_kind("send_email", ...) gives an untyped handler keyed by the kind string. Multiple hooks may be registered for the same kind; they stack and run sequentially.
Events¶
| Event | Fires when |
|---|---|
Started |
the claim transaction commits, just before the handler runs (job.state == running) |
WaitingForCallback |
the handler returned WaitForCallback and the job parked (job.state == waiting_external; job.callback_id is set) |
Completed |
the job finished successfully |
Retried |
the attempt failed (or RetryAfter) and the job will run again |
Exhausted |
the job exhausted its retries, or failed terminally, and moved to failed |
Cancelled |
the job was cancelled |
Rescued |
maintenance rescued the job (expired callback, stale heartbeat, or exceeded deadline) — reason: RescueReason carries which |
How JobResult / outcomes map to events:
| Outcome | Event |
|---|---|
| claim commits | Started |
Ok(Completed) |
Completed |
Ok(RetryAfter) / retryable Err (retries left) |
Retried |
retries exhausted (final attempt returned RetryAfter or a retryable Err), or Err(Terminal) |
Exhausted |
Ok(Cancel) |
Cancelled |
Ok(Snooze) |
none — an internal reschedule, not an outcome |
Ok(WaitForCallback) |
WaitingForCallback |
Callbacks¶
A job that returns WaitForCallback parks in waiting_external and emits WaitingForCallback. The terminal event fires later, when the callback is resolved — but only if you resolve it through the worker Client, which owns the hook registry:
// In your callback endpoint (same process as the worker Client):
let outcome = client.resolve_callback(callback_id, payload, default_action, None).await?;
// or the lower-level forms:
client.complete_external(callback_id, payload, None).await?; // -> Completed
client.fail_external(callback_id, error, None).await?; // -> Exhausted
client.retry_external(callback_id, None).await?; // -> Retried
Resolving through the bare awa_model::admin::* functions (or the CLI, or a process that has no worker Client) still transitions the job correctly but fires no hook — hooks are an in-process, worker-side facility. Dispatch happens once, in the process that performs the resolution, mirroring how an inline outcome fires once from the worker that claimed the job.
For a callback-driven Completed, duration is Duration::ZERO (no handler ran during the parked phase).
A callback that resumes the job to running emits no event itself; the job re-enters execution and emits Started plus a terminal event as usual.
Semantics¶
Hooks are best-effort, in-process notifications — not a durable workflow mechanism:
- They observe committed state. An outcome hook carries the post-transition
JobRow; a stale/rejected finalization emits no outcome event (though aStartedmay already have fired for that attempt). - They do not delay a job's state transition, and
shutdown()does not wait for them. They can be lost on process crash or shutdown. - Each attempt's hooks hold one of the worker's finalizer slots until they return, the same bounded slots that wait for completions to commit. Hooks slower than job throughput therefore slow claiming instead of piling up in memory. Keep hooks fast, and hand slow work to a job or follow-up instead.
- Maintenance dispatches
Rescuedevents, andCancelledfor duplicates it supersedes, on detached tasks with at most 1,024 outstanding per worker. Events beyond that are dropped with a warning rather than holding up rescue. While it drains a backlog, a queue-storage rescue pass keeps rescuing until it reachesrescue_pass_max_rows(5,000 by default) and then finishes its current batch, so a pass can exceed that by nearly one batch, andRescuedhooks slower than the drain lose events. - Dispatch is not atomic with the state transition. The hook runs after the job's transition commits, as a separate in-process step — it is not a transactional outbox. If the process dies between the commit and the dispatch, the job state is durable but the event is lost. The same applies to callback resolution:
complete_external/resolve_callbackcommit the outcome first, then dispatch. Startedruns alongside the handler, so a very fast job's state transition may commit before itsStartedhook completes. For outcomes the worker finalizes itself, the outcome hook runs only after theStartedhook has returned.- A panicking hook is logged; later hooks for the same kind still run.
If a side effect must fire exactly once or survive a crash, do not drive it from a hook at all — the dispatch may never run, and enqueueing the follow-up work from the hook just inherits that same unreliability. See the next section for the durable mechanism.
Durable follow-up jobs (on_*_enqueue)¶
For side effects that must survive a process crash — sending a welcome email after signup, kicking off downstream work, persisting an audit row — use the transactional follow-up API instead of an on_event hook. For worker-driven outcomes (the trigger handler returned Ok/Err) and for callback resolution via the worker Client — complete_external, fail_external, retry_external, and resolve_callback — the follow-up INSERTs in the same database transaction as the triggering state transition: either both commit or both roll back, with no in-between. Maintenance rescue dispatches follow-ups in a separate transaction (best-effort) — see the atomicity table below before designing a workflow that depends on rescue-driven follow-ups landing.
use awa::{Client, EnqueueRequest, QueueConfig};
let client = Client::builder(pool)
.queue("signup", QueueConfig::default())
.register::<Signup, _, _>(|_, _| async move { Ok(awa::JobResult::Completed) })
// When a Signup completes, enqueue a WelcomeEmail in the *same* tx.
// Closure receives the trigger args and the post-completion JobRow.
.on_completed_enqueue::<Signup, WelcomeEmail, _>(|args, job| WelcomeEmail {
user_id: args.user_id,
signup_job_id: job.id,
})
.build()?;
There is one builder per outcome, plus a _with variant whose closure returns Option<EnqueueRequest<F>>. Return Some to override InsertOpts on the follow-up (queue, priority, max_attempts, metadata, tags, unique, run_at, deadline_duration, ordering_key), or None to enqueue nothing (Conditional follow-ups):
| Builder | Fires when | Closure signature |
|---|---|---|
on_completed_enqueue |
JobResult::Completed |
(args, &job) |
on_cancelled_enqueue |
JobResult::Cancel(reason) |
(args, &job, &reason) |
on_retried_enqueue |
retryable err (retries left) or RetryAfter |
(args, &job, &error, attempt, next_run_at) |
on_exhausted_enqueue |
retries exhausted, or Err(Terminal) |
(args, &job, &error, attempt) |
on_waiting_for_callback_enqueue |
JobResult::WaitForCallback |
(args, &job) |
on_rescued_enqueue |
maintenance rescue (expired callback / stale heartbeat / deadline) | (args, &job, RescueReason) |
An on_*_enqueue_with closure returns Option<EnqueueRequest<F>>:
.on_completed_enqueue_with::<Signup, WelcomeEmail, _>(|args, job| {
Some(
EnqueueRequest::new(WelcomeEmail { user_id: args.user_id })
.queue("email")
.priority(1),
)
})
Conditional follow-ups¶
Return None from a _with closure to enqueue nothing for that trigger. The trigger's transition commits either way:
.on_exhausted_enqueue_with::<Charge, NotifyFinance, _>(|args, _job, _error, _attempt| {
(args.amount_cents >= 100_000).then(|| EnqueueRequest::new(NotifyFinance { charge_id: args.id }))
})
The closure decides from the trigger's args, its job row and the outcome's context. It runs wherever the follow-up insert runs, so this works for every outcome, including callback resolutions (complete_external, resolve_callback) where no handler runs. For worker-driven outcomes and callback resolutions that is inside the transition's transaction. For maintenance rescue it is the separate transaction opened after the rescue commits, so a conditional on_rescued_enqueue_with is no more atomic than any rescue follow-up (see Atomicity guarantees). A job whose kind has a registered Completed spec still takes the per-job completion transaction when the closure returns None. When the decision depends on something the handler computed, stage the follow-up from the handler instead (see Handler-staged follow-ups).
Deduplicating follow-ups¶
Give the follow-up unique options to collapse duplicates. When several triggers produce a follow-up with the same unique key, the first insert wins and the later ones are skipped: each trigger's transition still commits, and exactly one follow-up job exists. With UniqueOpts::default() the key covers the follow-up's kind and args, so put only the deduplication key in the follow-up's args. Custom options can add the queue or a time period, or set by_args: false so the key leaves args out; then follow-ups of the same kind collapse whatever their args. UniqueOpts::states decides how long the deduplication lasts: the default includes completed, so a finished follow-up keeps blocking duplicates until its terminal row is pruned. Drop completed from the states to block duplicates only while the existing follow-up is still scheduled, available, running or retryable.
.on_completed_enqueue_with::<StageA, StageB, _>(|args, _job| {
Some(EnqueueRequest::new(StageB { item_id: args.item_id }).unique(UniqueOpts::default()))
})
.unique() and .metadata() set one option each. with_opts replaces every option, including an earlier .queue() or .priority().
Handler-staged follow-ups¶
Registered specs see only the trigger's args and job row. When the follow-up needs a value the handler computes, such as the ID of the item it resolved, the handler stages the follow-up itself (ADR-046):
.register::<StageA, _, _>(|args, ctx| {
// A cloneable handle: the returned future cannot borrow `ctx`.
let follow_ups = ctx.follow_ups();
async move {
let item_id = resolve_item(&args).await?;
follow_ups.enqueue_on_completion(
EnqueueRequest::new(StageB { item_id }).unique(UniqueOpts::default()),
)?;
Ok(JobResult::Completed)
}
})
A Worker impl can call ctx.enqueue_on_completion(..) directly. In Python, the job passed to the handler has the same method, with the keyword arguments of Transaction.insert:
@client.task(StageA, queue="stage_a")
async def resolve(job):
item_id = await resolve_item(job.args)
job.enqueue_on_completion(StageB(item_id=item_id), unique_opts={})
Staged follow-ups are inserted in the same transaction as the attempt's completion, after any registered Completed specs, with the same deduplication. The rules:
- Only
Completedinserts them.RetryAfter,Snooze,Cancel, an error, a panic, a rescued or superseded attempt, andClient::complete_externalall discard them. A retried attempt starts empty and stages again when its handler re-runs.WaitForCallbackdiscards them with a warning and theawa.job.staged_follow_ups_discardedcounter; to build a follow-up from a callback payload, wait withctx.wait_for_callbackinside the handler, then stage. - A failed stage fails the attempt. Staging serializes the args, checks the limits, and rejects anything the insert would reject (priority outside 1–4,
max_attemptsoutside 1–1000, kind or queue names over 200 characters, more than 20 tags, NUL bytes, arun_atoutside the years 1–9999), returningStageError(Python:awa.StagingError, aTerminalError, also raised when args or options do not convert). If the handler ignores the error and returnsCompleted, the attempt fails terminally instead of completing without its follow-up. - Staging closes when the handler returns. A task the handler spawned and did not await gets
StageError::Sealed. - Limits per attempt: 16 follow-ups and 64 KiB of serialized requests, counting kind, queue, args, metadata, tags and ordering key (
MAX_STAGED_FOLLOW_UPS,MAX_STAGED_BYTES). Larger fan-out belongs in a dedicated fan-out job. - Cost. A job that stages takes a per-job completion transaction instead of the batched path; jobs that stage nothing are unaffected.
- Metrics.
awa.job.staged_follow_ups_committedcounts the follow-ups of committed completions byawa.follow_up.result:inserted, orduplicatewhen a job with the same unique key already existed. The Grafana dashboard's "Staged Follow-ups" panel shows both beside the discards. - Deploy order. A worker that does not know the follow-up's kind fails it terminally. Deploy the workers that handle the follow-up first, or put it on a queue only they poll.
The test harnesses behave the same way: awa_testing::TestClient::work_one and Python's AwaTestClient.work_one insert staged follow-ups when the handler completes, and awa.testing.HandlerJob records them in job.staged for unit tests without a database.
A handler that also wants to leave a record of what it did for people to inspect can record a job output beside the follow-up. The follow-up still gets its value in its args: a job must not read another job's output, and registered specs receive the row without it. Completed events carry the output.
Atomicity guarantees¶
| Emission site | Atomic with state commit? |
|---|---|
Worker-driven outcomes (handler returns Ok/Err) on either storage engine |
yes — one transaction, one commit |
Callback resolution on Client (complete_external, fail_external, retry_external, resolve_callback) on either storage engine |
yes — one transaction, one commit. A spec INSERT failure (or a panic in the user-supplied closure) rolls the callback transition back together with the follow-up; Client::*_external returns Err and the external sender can retry. |
| Maintenance rescue (stale heartbeat / deadline / expired callback) | no — best-effort. The rescue commits first, then specs dispatch in a separate tx. A spec failure leaves the rescue applied and is logged. |
The trigger transition and the follow-up INSERT either both commit or neither does — exactly-once enqueue per committed worker outcome or callback resolution. The one exception is a follow-up with unique options whose key already exists: that insert is skipped and the transition commits on its own (see Deduplicating follow-ups). The follow-up itself, once enqueued, is delivered with Awa's usual at-least-once semantics, so the follow-up handler still needs to be safe under re-execution.
Rescue stays best-effort by design: making it atomic would couple rescue liveness to the correctness of user follow-up code (a panic in on_rescued_enqueue would roll back a stale-heartbeat rescue and leave the job stuck in running). Zero-loss rescue notifications belong in an outbox/sweeper, not inlined into the rescue tx.
When to choose which API¶
- Observation (metrics, traces, alerts) →
on_event— cheap, in-process, acceptable to drop a sample on a crash. - Side effect that must survive a crash (send email, persist record, kick off downstream work) →
on_*_enqueue— slightly more expensive (it's an Awa job) but durable, retried, DLQ'd, and visible in admin. - Follow-up built from a value the handler computes → stage it with
enqueue_on_completion. It has the same atomicity ason_completed_enqueue, only forCompleted.
If you can't tell which you want, use on_*_enqueue. The cost of an unnecessary follow-up is bounded; the cost of a lost reliable side effect is not.
See ADR-015 for hook rationale, ADR-029 for the follow-up enqueue design and atomicity matrix, and ADR-046 for handler-staged follow-ups.