Distributed scheduler hell (MicroXChg 2017 Berlin)
0 likes222 views
The document discusses the transition from virtual machines to containers, focusing on distributed schedulers like Kubernetes, Nomad, and Mesos, which manage deployments and resource allocation for microservices. It highlights the challenges and requirements for deploying a timeseries database called Vulcan, including high network traffic and storage needs. The document also emphasizes the use of tools like Mesos and Kubernetes for scaling, load balancing, and managing services effectively.
![Custom Json Language
{
"application": {
"name": "timeseries-ingestor",
"scale": 1,
"ingresses": {
"timeseries-ingestor-health": {
"scheme": "http",
"container_port": 8001
}
},
"containers": {
"timeseries-ingestor": {
"image": "docker.com/timeseries/ingestor",
"image_tag": "fcb24ca",
"ports": [
8001,
9090
],
"env": {
"CASSANDRA_ADDRESS_LIST": ""
},
"resources": {
"cpu": "4",
"memory": "4000"
}
}
},
"metrics": {
"port": 8001,
"path": "/debug/metrics"
}
},
"maintainer": "dummy@digitalocean.com"
}](https://image.slidesharecdn.com/distributedtracinghell-170224120629/85/Distributed-scheduler-hell-MicroXChg-2017-Berlin-25-320.jpg)
Distributed scheduler hell (MicroXChg 2017 Berlin)
- 2. Story of .. How we moved 100(s) of virtual machines, onto containers
- 3. What is a distributed scheduler Your cloud provider: Digital Ocean, AWS, Google. For containers: Kubernetes, Nomad, Mesos
- 4. Why do we care? Microservices...
- 5. Vulcan Distributed timeseries database. open source https://github.com/digitalocean/vulcan
- 6. Requirements 4 3 Gbits a second of Network traffic 4 20 TB Storage 4 Sub 100ms Read times 4 100k writes a second
- 10. Kernel Provides 4 Virtual Memory 4 Process Isolation 4 Disk storage 4 Networking 4 CPU scheduling
- 11. Distributed Scheduler Provides 4 Container Deploy 4 Virtual Machine Deploy 4 Memory Quota 4 Disk storage 4 Networking 4 CPU scheduling 4 Scaling Instances
- 12. Mesos
- 15. Deploying our App On Mesos
- 19. Cassandra Pinning to specific Mesos nodes
- 20. Mesos Failure Modes Network Partition FAIL
- 23. Kubernetes
- 25. Custom Json Language { "application": { "name": "timeseries-ingestor", "scale": 1, "ingresses": { "timeseries-ingestor-health": { "scheme": "http", "container_port": 8001 } }, "containers": { "timeseries-ingestor": { "image": "docker.com/timeseries/ingestor", "image_tag": "fcb24ca", "ports": [ 8001, 9090 ], "env": { "CASSANDRA_ADDRESS_LIST": "" }, "resources": { "cpu": "4", "memory": "4000" } } }, "metrics": { "port": 8001, "path": "/debug/metrics" } }, "maintainer": "dummy@digitalocean.com" }
- 26. Command Line Deploys docc --contexts dev_env deploy timeseries-microservice1.json Docc is our internal Kubernetes tool
- 27. Deployment tool with diffs to Kubernetes "env": { +++ "CASSANDRA_CQL_VERSION": "3.1.7", --- "CASSANDRA_KEYSPACE": "staging_vulcan", +++ "CASSANDRA_KEYSPACE": "staging_vulcan_123", Simliar to KubeDiff
- 30. Load balancing Pushing 3 Gbs to kubernetes using Flannel
- 31. Metrics
- 33. Logging
- 35. Upsides to Distributed Schedulers
- 36. How to choice your abstractation
- 37. Alles hat ein Ende nur die Wurst hat zwei ('everything has one end, only the sausage has two')