Laravel queue design

Laravel Queue Design

A queued job will run twice, run late, or run after the world changed under it. Design every job for those three facts and production gets boring — in the good way.

When to use

• Creating or reviewing queued jobs, listeners, or notifications
• "Jobs are stuck / duplicated / silently failing"
• Moving slow work (mail, exports, API calls) out of the request
• Setting up or tuning Horizon / queue workers

Principles

• At-least-once means duplicates. Every handler must tolerate running twice — idempotency is the entry fee, not a nice-to-have.
• Small payload, fresh state. Pass IDs, re-fetch inside the job. The world at dispatch time is not the world at run time.
• Failure is a feature. Decide per job what retry, backoff, and final failure mean — the defaults are guesses.
• One job, one outcome. A job that does five things fails five ways and retries all of them.

Process

1. Design the job

class ProcessRefund implements ShouldQueue
{
    public function __construct(public int $orderId) {}   // ID, not model state you'll trust later

    public $tries = 3;
    public $backoff = [10, 60, 300];          // seconds; spread retries out
    public $timeout = 120;                     // must be < retry_after in queue config!

    public function handle(): void
    {
        $order = Order::findOrFail($this->orderId);   // fresh state
        if ($order->refunded) return;                  // idempotency check
        // ...
    }
}

• Eloquent models in constructors are fine (SerializesModels re-fetches), but decisions made on dispatch-time state are stale — re-check conditions in handle()
• $timeout < retry_after (queue config) or the job runs twice concurrently — the classic Laravel queue foot-gun

2. Make it idempotent

Pick the mechanism that fits:

• State check: if ($order->refunded) return; — re-read, then act
• Unique jobs: ShouldBeUnique to prevent duplicate dispatch
• DB constraint: unique index + firstOrCreate/upsert absorbs the duplicate
• External calls: pass an idempotency key to the payment/API provider

3. Decide the failure story

• failed(Throwable $e) method: notify, mark the domain object, compensate
• Don't retry the unretryable: validation-type failures → $this->fail($e) immediately
• Poison-pill protection: $tries/$maxExceptions so one bad payload doesn't loop forever
• Monitor failed_jobs — a table nobody looks at is a black hole. Alert on growth; queue:retry after fixing the cause

4. Compose bigger work

Need	Tool
Steps in order, abort on failure	Job chain (Bus::chain)
Fan-out N items + completion hook	Batch (Bus::batch(...)->then(...)->catch(...))
Per-tenant/user serialization	WithoutOverlapping middleware
External API rate limits	RateLimited middleware, dedicated queue

For batches: items must be independently retryable; allowFailures() if one bad item shouldn't sink the batch.

5. Separate queues by SLA

• high (user-facing: mail verification, receipts), default, low (exports, sync)
• Workers: --queue=high,default,low priority order, or dedicated workers per queue
• One slow queue must not starve fast ones — that's the whole point of separation

6. Operate it

• Horizon: balance/processes per queue, alert on wait time, watch failed rate
• Deploys: queue:restart (workers hold old code in memory until restarted)
• Track queue wait time (latency from dispatch to start) — it's the user-facing number

Output format

## Queue review: <job/feature>

| Job | Idempotent via | tries/backoff | timeout<retry_after | On final failure |
|-----|----------------|---------------|---------------------|------------------|
| …   | state check    | 3 / 10-60-300 | ✅ 120<300          | notify + mark    |

### Gaps (ranked)
1. <job>: <risk> → <fix>

### Queue topology
high: …, default: …, low: … — workers: …

Anti-patterns

• ❌ Serializing a model and trusting its attributes at run time (stale decisions)
• ❌ $timeout ≥ retry_after — concurrent double execution by design
• ❌ Default retries on a non-idempotent job touching money
• ❌ One giant job processing 10 000 rows instead of a batch of small ones
• ❌ failed_jobs as a write-only table with no alerting
• ❌ Forgetting queue:restart on deploy, then debugging "old code" for an hour
• ❌ Everything on the default queue — exports starving password-reset mails