StatefulSet vs Deployment

[sbogomolov]

Currently helm charts create a Deployment. Deployment is great for stateless applications, it guarantees "at least N pods" semantics. StatefulSet guarantees "at most N pods". In our case we have state: there are shared DB and volume (with config, logs, cache, ...). Is Jellyseerr designed with this shared state in mind? If not, I suggest we switch to using StatefulSet (I'm happy to do the work btw). With Deployment, even with ReplicaCount set to 1 we might end-up in a situation when one pod is not yet dead and another one is already started.

[gauthier-th]

CC @M0NsTeRRR

[M0NsTeRRR]

Hello @sbogomolov,

Jellyseerr is not designed to support more than one replica and the default configuration with strategy: recreate avoid the case you describe.

[sbogomolov]

Hey @M0NsTeRRR,

This is a good point. However it works only when deployment is updated. In case pod gets shutdown - ReplicaSet will start another pod while first one is still terminating, so there will be a moment when both pods are alive. There are also more exotic scenarios, for example if node with the running pod is lost - a new pod will be started (to guarantee “at least 1”). That lost node may actually be still running the pod. When it comes back - we’ll have two pods running at the same time. These scenarios are not likely of course, but switching to StatefulSet is also free and does not seem to have obvious downsides.

[M0NsTeRRR]

switching to StatefulSet is also free and does not seem to have obvious downsides.

I agree to switch if they are benefits otherwise it's a change for a change 😄

ReplicaSet will start another pod while first one is still terminating, so there will be a moment when both pods are alive

https://kubernetes.io/docs/concepts/workloads/controllers/deployment/#strategy
You’re absolutely right. I’ll let you drive the PR; feel free to tag me once you think it’s ready or if you need feedback. Thanks for the explanation ! ;)

Thank you for your feedback !

[sbogomolov]

#1901