Kubernetes Deployment

Deployment Manifest

apiVersion: apps/v1
kind: Deployment
metadata:
  name: celery-outbox-relay
spec:
  replicas: 2
  selector:
    matchLabels:
      app: celery-outbox-relay
  template:
    metadata:
      labels:
        app: celery-outbox-relay
    spec:
      containers:
        - name: relay
          image: myapp:latest
          command:
            - python
            - manage.py
            - celery_outbox_relay
            - --batch-size=100
            - --idle-time=1.0
            - --liveness-file=/tmp/alive
          env:
            - name: DATABASE_URL
              valueFrom:
                secretKeyRef:
                  name: myapp-secrets
                  key: database-url
            - name: CELERY_BROKER_URL
              valueFrom:
                secretKeyRef:
                  name: myapp-secrets
                  key: broker-url
          livenessProbe:
            exec:
              command:
                - python
                - -c
                - |
                  import os
                  import sys
                  import time

                  path = "/tmp/alive"
                  max_age_seconds = 90

                  try:
                      stale_for = time.time() - os.path.getmtime(path)
                  except FileNotFoundError:
                      sys.exit(1)

                  sys.exit(0 if stale_for < max_age_seconds else 1)
            initialDelaySeconds: 10
            periodSeconds: 30
          resources:
            requests:
              memory: "128Mi"
              cpu: "100m"
            limits:
              memory: "256Mi"
              cpu: "500m"

Scaling

The relay scales horizontally. Each instance locks different rows via SELECT FOR UPDATE SKIP LOCKED.

Recommended: 2-3 replicas for high availability.

Graceful Shutdown

The relay handles SIGTERM gracefully. Kubernetes sends SIGTERM during pod termination.

Set terminationGracePeriodSeconds to allow batch completion:

spec:
  terminationGracePeriodSeconds: 120  # >= shutdown_timeout + send_timeout + margin

Size the grace period to cover your configured --shutdown-timeout, --send-timeout, and some scheduling margin. A shorter window can force SIGKILL before the relay finishes draining.

Resource Requirements

Workload	Memory	CPU
Low volume	128Mi	100m
Medium volume	256Mi	250m
High volume	512Mi	500m

Relay is I/O bound (database + broker), not CPU bound.