queue

name	queue
description	Message queue and job processing -- Kafka, RabbitMQ, SQS, BullMQ, Celery, Sidekiq.

Activate When

• /godmode:queue, "add background jobs", "set up queue"
• "retry failed jobs", "dead letter queue", "job stuck"
• "rate limit processing", "backpressure"

Workflow

1. Requirements

grep -r "bullmq\|celery\|sidekiq\|kafkajs\|sqs" \
  package.json requirements.txt 2>/dev/null

Use case: job processing | event streaming | pub/sub
Volume: <msg/sec>, Payload: <avg size>
Ordering: strict FIFO | partition | best-effort
Delivery: exactly-once | at-least-once | at-most-once
Latency: <100ms | <1s | <30s
Existing infra: Redis | PostgreSQL | AWS | none

2. Technology Selection

SQS: AWS-native, managed, unlimited throughput
BullMQ: Node.js+Redis, 10-50K/sec, great dashboard
Celery: Python+Redis/RabbitMQ, complex workflows
Kafka: millions/sec, partition-ordered, replayable
RabbitMQ: complex routing, 10-50K/sec
Redis Streams: lightweight, 100K+/sec
PG SKIP LOCKED: no new infra, <1K jobs/sec

IF AWS simple: SQS. IF Node.js+Redis: BullMQ. IF event streaming: Kafka. IF low volume+PG: SKIP LOCKED.

3. Architecture

Producers -> Broker
  -> [high-priority] -> Worker Pool A (concurrency 10)
  -> [default]       -> Worker Pool B (concurrency 20)
  -> [bulk]          -> Worker Pool C (concurrency 5)
  -> [dead-letter]   -> DLQ Processor

4. Retry Strategy & Dead Letters

Retry: 0s -> 1s -> 4s -> 16s -> 60s (cap) -> DLQ
Formula: min(base * 2^attempt + jitter, max_delay)

Retryable: network timeout, 5xx, DB connection, 429
Non-retryable: 4xx, auth, deserialization, biz logic

DLQ: <original>-dlq, retention 30 days, alert >100
Options: replay | replay with fix | skip | escalate

5. Delivery Guarantees & Idempotency

• At-most-once: ack before process (metrics/logs)
• At-least-once: ack after process + DLQ (most tasks)
• Exactly-once: transactional + idempotency keys (payments, financial, orders)

Idempotency: check Redis key, acquire lock (NX+TTL), process, store result (TTL 24h), release lock.

6. Priority & Rate Limiting

P0 critical: password reset, payment (SLA <10s)
P1 high: welcome email, order confirm (SLA <60s)
P2 normal: notifications, image proc (SLA <5m)
P3 low: reports, exports (SLA <1h)
P4 background: cleanup, analytics (SLA <24h)

Rate limit: token bucket or BullMQ limiter { max: 100, duration: 60000 }.

7. Worker Pool & Backpressure

Min workers: 2, Max: 20, Scale on depth > 100
Concurrency per worker: 10 (I/O-bound)

Backpressure:
  Depth > 1K -> WARN + alert
  Depth > 10K -> SCALE workers
  Depth > 100K -> SHED low-priority
  Worker memory > 80% -> PAUSE accepting
  Downstream 5xx > 10% -> CIRCUIT BREAK

Graceful shutdown: SIGTERM -> pause -> finish -> exit.

8. Scheduling

Daily digest: 0 9 * * * (default, 5m timeout)
Cleanup: 0 */6 * * * (bg, 30m timeout)
Weekly reports: 0 2 * * 1 (bulk, 2h timeout)

Distributed lock (one scheduler), all UTC, alert if job misses window.

9. Monitoring

Queue depth waiting (<1K), active (<50), DLQ (<100)
Processing rate (>10/s), Success rate (>95%)
Avg time (<10s), P95 (<30s)
Worker count (>2), Retry rate (<10%)

Hard Rules

1. NEVER >500ms tasks in request handlers.
2. NEVER retry without exponential backoff + jitter.
3. ALWAYS design idempotent handlers.
4. NEVER >1MB payloads in queue (use S3 reference).
5. NEVER share workers across priority levels.
6. NEVER skip graceful shutdown (SIGTERM).
7. ALWAYS configure DLQ for every production queue.
8. ALWAYS monitor DLQ depth and alert.
9. ALWAYS classify retryable vs non-retryable.
10. ALWAYS set job TTL/retention.

TSV Logging

Append .godmode/queue-results.tsv:

timestamp	technology	queues	worker_pools	retry	dlq	idempotency	status

Keep/Discard

KEEP if: depth stable AND error rate < 1%
  AND P95 < target AND DLQ not growing.
DISCARD if: depth growing OR error spike
  OR DLQ growing.

Stop Conditions

STOP when ALL of:
  - All >500ms tasks queued
  - Retry + backoff configured
  - DLQ on all queues
  - Idempotent handlers
  - Graceful shutdown
  - Error rate < 1%

Autonomous Operation

On failure: git reset --hard HEAD~1. Never pause.

Error Recovery

Failure	Action
Jobs stuck active	Check logs, verify timeout
DLQ growing	Inspect patterns, fix root cause
Memory exhaustion	Limit concurrency, check leaks
Duplicates	Verify idempotency key, add dedup
Backlog growing	Scale workers, add backpressure