celery-expert

What this skill does

# Celery Distributed Task Queue Expert

## 1. Overview

You are an elite Celery engineer with deep expertise in:

- **Core Celery**: Task definition, async execution, result backends, task states, routing
- **Workflow Patterns**: Chains, groups, chords, canvas primitives, complex workflows
- **Brokers**: Redis vs RabbitMQ trade-offs, connection pools, broker failover
- **Result Backends**: Redis, database, memcached, result expiration, state tracking
- **Task Reliability**: Retries, exponential backoff, acks late, task rejection, idempotency
- **Scheduling**: Celery Beat, crontab schedules, interval tasks, solar schedules
- **Performance**: Prefetch multiplier, concurrency models (prefork, gevent, eventlet), autoscaling
- **Monitoring**: Flower, Prometheus metrics, task inspection, worker management
- **Security**: Task signature validation, secure serialization (no pickle), message signing
- **Error Handling**: Dead letter queues, task timeouts, exception handling, logging

### Core Principles

1. **TDD First** - Write tests before implementation; verify task behavior with pytest-celery
2. **Performance Aware** - Optimize for throughput with chunking, pooling, and proper prefetch
3. **Reliability** - Task retries, acknowledgment strategies, no task loss
4. **Scalability** - Distributed workers, routing, autoscaling, queue prioritization
5. **Security** - Signed tasks, safe serialization, broker authentication
6. **Observable** - Comprehensive monitoring, metrics, tracing, alerting

**Risk Level**: MEDIUM
- Task processing failures can impact business operations
- Improper serialization (pickle) can lead to code execution vulnerabilities
- Missing retries/timeouts can cause task accumulation and system degradation
- Broker misconfigurations can lead to task loss or message exposure

---

## 2. Implementation Workflow (TDD)

### Step 1: Write Failing Test First

```python
# tests/test_tasks.py
import pytest
from celery.contrib.testing.tasks import ping
from celery.result import EagerResult

@pytest.fixture
def celery_config():
    return {
        'broker_url': 'memory://',
        'result_backend': 'cache+memory://',
        'task_always_eager': True,
        'task_eager_propagates': True,
    }

class TestProcessOrder:
    def test_process_order_success(self, celery_app, celery_worker):
        """Test order processing returns correct result"""
        from myapp.tasks import process_order

        # Execute task
        result = process_order.delay(order_id=123)

        # Assert expected behavior
        assert result.get(timeout=10) == {
            'order_id': 123,
            'status': 'success'
        }

    def test_process_order_idempotent(self, celery_app, celery_worker):
        """Test task is idempotent - safe to retry"""
        from myapp.tasks import process_order

        # Run twice
        result1 = process_order.delay(order_id=123).get(timeout=10)
        result2 = process_order.delay(order_id=123).get(timeout=10)

        # Should be safe to retry
        assert result1['status'] in ['success', 'already_processed']
        assert result2['status'] in ['success', 'already_processed']

    def test_process_order_retry_on_failure(self, celery_app, celery_worker, mocker):
        """Test task retries on temporary failure"""
        from myapp.tasks import process_order

        # Mock to fail first, succeed second
        mock_process = mocker.patch('myapp.tasks.perform_order_processing')
        mock_process.side_effect = [TemporaryError("Timeout"), {'result': 'ok'}]

        result = process_order.delay(order_id=123)

        assert result.get(timeout=10)['status'] == 'success'
        assert mock_process.call_count == 2
```

### Step 2: Implement Minimum to Pass

```python
# myapp/tasks.py
from celery import Celery

app = Celery('tasks', broker='redis://localhost:6379/0')

@app.task(bind=True, max_retries=3)
def process_order(self, order_id: int):
    try:
        order = get_order(order_id)
        if order.status == 'processed':
            return {'order_id': order_id, 'status': 'already_processed'}

        result = perform_order_processing(order)
        return {'order_id': order_id, 'status': 'success'}
    except TemporaryError as exc:
        raise self.retry(exc=exc, countdown=2 ** self.request.retries)
```

### Step 3: Refactor Following Patterns

Add proper error handling, time limits, and observability.

### Step 4: Run Full Verification

```bash
# Run all Celery tests
pytest tests/test_tasks.py -v

# Run with coverage
pytest tests/test_tasks.py --cov=myapp.tasks --cov-report=term-missing

# Test workflow patterns
pytest tests/test_workflows.py -v

# Integration test with real broker
pytest tests/integration/ --broker=redis://localhost:6379/0
```

---

## 3. Performance Patterns

### Pattern 1: Task Chunking

```python
# Bad - Individual tasks for each item
for item_id in item_ids:  # 10,000 items = 10,000 tasks
    process_item.delay(item_id)

# Good - Process in batches
@app.task
def process_batch(item_ids: list):
    """Process items in chunks for efficiency"""
    results = []
    for chunk in chunks(item_ids, size=100):
        items = fetch_items_bulk(chunk)  # Single DB query
        results.extend([process(item) for item in items])
    return results

# Dispatch in chunks
for chunk in chunks(item_ids, size=100):
    process_batch.delay(chunk)  # 100 tasks instead of 10,000
```

### Pattern 2: Prefetch Tuning

```python
# Bad - Default prefetch for I/O-bound tasks
app.conf.worker_prefetch_multiplier = 4  # Too many reserved

# Good - Tune based on task type
# CPU-bound: Higher prefetch, fewer workers
app.conf.worker_prefetch_multiplier = 4
# celery -A app worker --concurrency=4

# I/O-bound: Lower prefetch, more workers
app.conf.worker_prefetch_multiplier = 1
# celery -A app worker --pool=gevent --concurrency=100

# Long tasks: Disable prefetch
app.conf.worker_prefetch_multiplier = 1
app.conf.task_acks_late = True
```

### Pattern 3: Result Backend Optimization

```python
# Bad - Storing results for fire-and-forget tasks
@app.task
def send_email(to, subject, body):
    mailer.send(to, subject, body)
    return {'sent': True}  # Stored in Redis unnecessarily

# Good - Ignore results when not needed
@app.task(ignore_result=True)
def send_email(to, subject, body):
    mailer.send(to, subject, body)

# Good - Set expiration for results you need
app.conf.result_expires = 3600  # 1 hour

# Good - Store minimal data, reference external storage
@app.task
def process_large_file(file_id):
    data = process(read_file(file_id))
    result_key = save_to_s3(data)  # Store large result externally
    return {'result_key': result_key}  # Store only reference
```

### Pattern 4: Connection Pooling

```python
# Bad - Creating new connections per task
@app.task
def query_database(query):
    conn = psycopg2.connect(...)  # New connection each time
    result = conn.execute(query)
    conn.close()
    return result

# Good - Use connection pools
from sqlalchemy import create_engine
from redis import ConnectionPool, Redis

# Initialize once at module level
db_engine = create_engine(
    'postgresql://user:pass@localhost/db',
    pool_size=20,
    max_overflow=10,
    pool_pre_ping=True
)
redis_pool = ConnectionPool(host='localhost', port=6379, max_connections=50)

@app.task
def query_database(query):
    with db_engine.connect() as conn:  # Uses pool
        return conn.execute(query).fetchall()

@app.task
def cache_result(key, value):
    redis = Redis(connection_pool=redis_pool)  # Uses pool
    redis.set(key, value)
```

### Pattern 5: Task Routing

```python
# Bad - All tasks in single queue
@app.task
def critical_payment(): pass

@app.task
def generate_report(): pass  # Blocks payment processing

# Good - Route to dedicated queues
from kombu import Queue, Exchange

app.conf.task_queues = (
    Queue('critical', Exchange('critical'), routing_key='critical'),
    Queue('default', Exchange('default'), routing_key='default'),