Distributed Scheduler Hell

microXchg · 2017 · Berlin · 2017-02-17
kubernetesmesosnomaddockerdevops video
Watch on YouTube Slides (PDF) ← All speaking

microXchg 2017 · Berlin

A practical tour of distributed container schedulers drawn from running DigitalOcean’s monitoring stack at scale.

What the talk covers

  • What a distributed scheduler actually provides vs the Linux kernel (deploy, quotas, networking, scaling)
  • Vulcan — DigitalOcean’s Prometheus-compatible timeseries system — as a real workload: ~3 Gbit/s network, 100k writes/s, sub-100ms reads, tens of TB storage
  • Deploying that workload on Mesos + Marathon, custom Kafka schedulers, Cassandra pinning
  • Failure modes (network partitions) and a counter-example with Nomad
  • Moving to Kubernetes, internal deploy tooling (Docc / kube-diff style workflows), Flannel load balancing, metrics and logging

Takeaways

How to choose an abstraction for middle-size apps growing into large multi-service fleets — and when not to reach for a full scheduler.