Docker Swarm Auto Scale

[SCBionicle]

I didn't see a mention for Docker Swarm auto scaling support, so I am assuming it is not supported. Since there is service discovery support and I typically deploy in a Docker Swarm Stack, is it possible to implement auto scaling for Docker Swarm?

[itzg]

Yes, should be possible to implement

[itzg]

Auto-Scaling for Docker Swarm Complexity Assessment

Current State

The Docker Swarm watcher already has stubs in place (lines 45-63 in docker_swarm.go) that call logrus.Fatal(). The infrastructure is ready; only implementation is needed.

Implementation Complexity: 🟡 Moderate

What Exists

• ✅ Service discovery for Docker Swarm services (parsing labels, VirtualIPs)
• ✅ Interface definitions for waker/sleeper functions
• ✅ Working reference implementations from Docker containers & Kubernetes

Key Components

Component	Difficulty	Notes
Scale-Up	🟡 Moderate	Scale replicas via Service Update API (service.Spec.Mode.Replicated.Replicas)
Scale-Down	🟡 Moderate	Same mechanism (set replicas to 0)
Reachability Check	🟢 Easy	Reuse TCP dial pattern from other backends
Error Handling	🟡 Moderate	Watch task replicas, handle convergence delays

Key Challenges

1. Replica Convergence: Unlike containers (instant) or K8s StatefulSets (managed), Docker Swarm may take time to schedule/start task replicas. Need to poll for actual tasks.
2. No Task Guarantee: Can scale replicas but Swarm doesn't guarantee tasks will run if resources/constraints aren't met. Must check task status.
3. Global State: Swarm tracks replicas globally; need to query actual running tasks via task status checks.

Effort Estimate

• Core implementation: ~2-4 hours (90 lines of code)
• Testing: +1-2 hours
• Edge cases (failed tasks, slow convergence): +1-2 hours
• Total: ~5-7 hours for production-ready code

Implementation Approach

1. Update service replicas (0 → 1 for wake, 1 → 0 for sleep)
2. Poll until actual running tasks appear (for wake)
3. TCP dial to service VirtualIP:port to verify reachability
4. Handle convergence delays and failed tasks gracefully

This is a well-scoped feature that follows the established patterns from Docker containers and Kubernetes implementations.