New Prometheus exporter for Docker Swarm scheduler visibility. Looking for feedback

Docker Engine Swarm
1

Hi everyone,

I run Docker Swarm in my homelab as a single-node Swarm.
While monitoring it, I realized I was missing clear visibility into what the Swarm scheduler is actually doing.
Especially around task states and global services.

I have been working on a small open source Prometheus exporter called Swarm Scheduler Exporter and I would like some early feedback from other Swarm users.

What it focuses on:

  • Task state visibility per service using a latest-per-slot model.
  • Accurate desired replicas for global services based on eligible nodes only.
    Status, availability, constraints, and platform are considered.
  • Simple service readiness signals for alerting.

Some technical details that may be relevant here:

  • Watches service and node events and polls tasks periodically.
  • Stable labels and controlled cardinality.
  • Runs on a manager node and exposes only /metrics and /healthz.

Example metrics:

  • swarm_service_desired_replicas
  • swarm_task_replicas_state
  • swarm_service_at_desired
  • swarm_cluster_nodes_by_state

The project started as a fork of akerouanton/swarm-tasks-exporter, but it has diverged quite a bit.

Repository and documentation:
https://github.com/leinardi/swarm-scheduler-exporter

I am mainly looking for feedback on:

  • Desired replicas behavior for global services.
  • Missing task or service states.
  • Swarm edge cases I may have overlooked.

This is not an official Docker project, just a tool I built for my own setup and decided to share.

2

Not many Swarm users left, you could post it to /r/DockerSwarm.

3

I guess I’m late to the party :sweat_smile: I’ll try on the subreddit you suggested.

4

This directly addresses the visibility gaps in Swarm scheduling, especially around task states and global services. Feedback from real-world edge cases like node drain, constraints, and updates will help validate the approach.

5

No worries, we still use it for multiple production systems. It just works, way less components and configuration than k8s.