dlespiau/balance

Client side load balancing for Kubernetes clusters

balance's Issues

Live view of pending requests

It'd be nice to have a live view to debug/understand things.

A few notes:

• live may be not very interesting for low trafic services
• View of the requests in the last n minutes (30s, 1m, 5m)
• A small websocket protocol to update values live

Decide what to do with logging

We want support to support logging, but as balance is a library, we should allow users to provide their own logging object we can use.

Add coverage report

Maybe using coveralls.

Add travis CI & gometalinter

Write a stats Algorithm object

It's be nice to be able to gather statistics to observe decisions made by the load balancing algorithms. This could be done by introducing objects implementing the LoadBlancer and AffinityLoadBlancer interfaces and that would gather stats and call the real algorithm implementation.

Basic e2e tests

We need some start on e2e tests. A basic one could be:

1. deploy proxy + service on a minikube cluster
2. send some traffic
3. collect metrics and check things look right.

Consistent hashing implementation doesn't take into account hash collisions

With a 32-bit hash and lots of v-nodes, it doesn't take much to have collisions somewhat likely. For 32 endpoints, 256 v-nodes: 2**32/256/32 = 524288.

We should be able to detect collisions and have a strategy to solve them (eg. hash of the hash?).

Provide a load balancer that uses the Kubernetes service DNS/VIP

We could provide a LoadBalancer object that just uses the Kubernetes DNS/VIP so we can write generic code that just get an endpoint whether we use a fancy algorithm or just the default Service mechanism.

That's also the first step to be able to switch algorithms live, but that's another story.