Getting the Most Out of Your Mesos
summary
Yelp runs Apache Mesos in production since 2014, hosting about 400 services and ~600 batch jobs and running 100k unit/integration tests hundreds of times a day (one full run ≈15 minutes). Built tools include Pasta, Tron, Jolt and a benchmarking harness that archives experiments into Docker images with Jupyter notebooks. Load tests show Docker launching 10,000 containers averages ~60s (worst ~3m) versus UCR worst ~8s. Cluster Man provides a pluggable autoscaler plus a simulated AWS/Mesos environment for reproducible predictability and what’s called feedbackability.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:34Yelp's Mission Connecting people with great local businesses
- 2:30(A few) Examples of Mesos Applications
- 4:31Mesos Adoption at Yelp
- 7:31Metrics and Monitoring
- 8:04Tracing and Logging
- 8:55How many services can we support?
- 10:30How many frameworks can we run?
- 12:17Load Testing Apache Mesos
- 15:43Docker Containerizer vs UCR
- 21:38Clusterman Example: what happened
- 22:34Clusterman Example: What will happen
- 24:07Feedbackability
related talks
39mReal-Time Event Processing and Handling Stateful Applications on Apache Mesos
42mStreaming Data Pipelines on Apache Mesos: Lessons Learned
24mUncovering Community and Project Insights Through Data Driven Methods
39mThe Sparkling Solution that Scales to IoT
12mEdward Hsu, Mesosphere | DockerCon 2018
33mContainerizing & Orchestrating Hadoop Infrastructure at Scale
29mScale Your Batch / Big Data / AI Workloads Beyond the K8s...
30mKubernetes + Operator + PaaSTA = Flink @ Yelp
