manna

Synchronise multiple containers/nodes competing for limited cloud quotas with a single point of failure.

Proposed usage

Create a Deployment with a single replica and a service that exposes it via the Kubernetes DNS addon:

kubectl create --save-config -f k8s-dep.yaml
kubectl create --save-config -f k8s-svc.yaml

The per-second growth-rate for manna is defined as environment variables in the deployment, along with the maximum attainable value for manna.

Usage in apps running across multiple pods:

bqManna := manna.NewBar("gcloud-bigquery").WithPort(9999) // Which manna service do we connect to?
bqManna.Timeout(60) // In seconds - give up if no available manna for a minute in this case. 0 = infinite
bqManna.Protect(func() {
    // This block is synchronous - i.e., returns only when protected func returns
})

Functional Specifications/Behaviour

In case of available manna

Require a point of manna and execute block. Open finishing the block, release manna back.

In case of no manna available

Block and poll with a one second sleep until manna is available. Respect the timeout.

In case of connection failure to the service

Treat this as no manna available.

In case of pod crashing

On every restart, manna is to grow at its per-second growth rate. Thus, a crash is treated as a reset of manna to zero. In the future, we can consider exploiting the SIGTERM issued by Kubernetes to our pod to save our count-state and elegantly restore it upon pod restart.

Misc

All clients talk to the master manna note (the central point-of-failure) via gRPC. The manna.Bar type is a struct (passed by value) and only holds a string identifying the service to attempt connecting to. The actual connection is made (sort of lazily) when Protect is called.

angadn/manna