More Bang for Your Buck: How Yelp Autoscales Mesos + Marathon on AWS Spot Fleet
summary
Yelp runs autoscaling PaaSTA with Mesos + Marathon, operating ~900 Marathon apps, ~5.5k tasks and ~600 agents across data centers and AWS. A service autoscaler (writes to Zookeeper every 10 min) and a cluster autoscaler (runs ~20 min, targets ~80% utilization) adjust instances; host shutdowns use Mesos maintenance primitives, dynamic reservations and Marathon draining. Spot Fleet use (2‑minute termination notices) is diversified via many instance types and Terraform configs; bidding ~2× on-demand and average costs drop to ~47–60% with 95th percentile instance uptime ≈ one day. Substantial savings with notable engineering effort.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 0:48Agenda
- 3:44Yelps Production Clusters
- 5:10Yelp Traffic
- 6:08Service Autoscale
- 9:06Service Metrics
- 10:34Decision Policies
- 15:09Autoscaling Operations
- 17:56Scaling Down
- 22:07AWS Spot Fleet Conditions
- 22:30Why Live With Volatile Capacity
- 23:08Strategy 1 counterintuitive
- 23:46Strategy 2 diversifying
- 29:12Ignoring Hosts
- 29:41Cost Comparison
- 31:10Future Plans
- 32:01Parallel Scaling
- 32:28Deployment to More Services
- 32:51Conclusion
- 34:01Hiring
- 34:19Mitigating risks
related talks
36mHow Yelp.com Runs on Apache Mesos in AWS Spot Fleet for Fun and Profit
30mKubernetes + Operator + PaaSTA = Flink @ Yelp
42mAWS Spot instances: huge savings (or huge risks?)
21mHow you can scale up with Amazon EC2 for up to 90% less
43mSpot Instances
31mKeeping Your Cloud Footprint in Check
38m50 Shades of K8s Autoscaling - Mario Leander Reimer
32mBlending Spot and On-Demand Instances and Optimizing Costs