I would like to speak about what I am actually doing at InfluxData. Sharing with you some ideas about how an orchestrator should work. We will start from a bit of history about distributed system, containers, runtime and so on. Hoping to have a good chat about the future of scheduling and orchestrator.
20. Advanced networking ideas about v2
Switch from Security groups isolation to
Security Groups plus subnet
Master Master
Master
us-east-1
sg-21412
E
L
B
Web
21. Interface between Master and Slave ideas about v2
Master
CRI from k8s can be a good gRPC
interface between Master and Worker
● Worker will be simpler (it will get ONLY
what to do directly from the master)
● We can reuse cri-o, docker-cri,
contaierd-cri what ever! (runtime are
bored!
22. ¨ Join worker and master under the
same repository to make it easy to test
and improve.
¨ First step for a lot of things like
¨ Advantage container networking.
¨ Scheduling based on cluster
topology and stull file labels.
¨ If you don’t need Raft you are
wrong!
Raft and membership ideas about v2
image credits Martin Kleppmann
23. I have no idea yet. That’s why I am here!
¨ Efficient and easy to manage in-cluster
networking.
¨ More secure with something like BPF
scanning or Cilium.
Advanced container networking ideas about v2
24. Stability and HealthCheck ideas about v2
Gossip protocol based on SWIM paper.
¨ Fast propagation
¨ Small and constant network load
¨ We need to check instances and
containers from outside and inside
https://youtu.be/CDQaqiRhTtk?t=12m15s
25. ¨ Raft should allow us to distribute the requests to the right datacenter (the one where
the cluster is).
¨ Will improve isolation and security.
¨ It will make the code a lot easy.
Master Master
us-east-1 eu-west-2
Master
Master Master
Master
Improve multi datacenters distribution ideas about v2
26. ¨ Recover EC2 from failure
¨ Recover Containers
¨ Secure and Stable update/termination from S1 to S2
Achieve and maintain desired state ideas about v2
28. ¨ In InfluxData we think that distributed
tracing can be the right decision
¨ We can trace our a requests moves
between:
¨ Frontend
¨ Queue System
¨ Master and Slave
¨ AWS
¨ Runtime
¨ We would like to validate if the tracing
model can work with InfluxDB and if it can
be really helpful. We are the guinea pig.
Distributed Tracing ideas about v2