How to connect to read-write pod with RedisReplication

[renchap]

Thanks a lot for this project, this is exactly what I need!

I see that you recently released the RedisReplication, which is perfect for us as our client does not support multiple Redis servers (yet).

The issue I am facing is that there seem to be no service targetting the current primary server, so if you connect to either of the created services, you can end up on a replica and get an error when you try to write a key.

What is the expected way for applications to have a read-write endpoint? And what is the difference between the 2 existing ones? (<name> and <name>-additional).

I would have expected the first one to point to the read-write pod, but it has 3 endpoints.

Thanks!

[shubham-cmyk]

@renchap Right now you must setup sentinel mode along the replication to perform the write on master and read on slave

[shubham-cmyk]

I would roll out this feature in next release which is near.

[renchap]

My issue is that some software does not support Sentinel, this is where deploying a RedisReplication setup can really help.

At the moment, does the operator support failing over the read-only replica? If yes, then as it makes the decision to fail over, then it should also update the service.

StatefulSets have a special label on each pod: statefulset.kubernetes.io/pod-name= which allows a service to point to a specific pod.
The operator could use this to create and manage a service that points to the write-enabled pod in the statefulset. If it does not handle failing over and enabling writes on the replica then this can be statically set to the first pod (write-enabled), if you handle promoting a read-replica to primary with write, it can update the service when you do the promotion.

[shubham-cmyk]

Managing the service could solve the problem we could create 1 additional service something like
replication-master-service. We could use the label to detect the master pod statefulset.kubernetes.io/pod-name=

We would definitely introduce the failover without the sentinel mode also. So the selector on the replication-master-service would keep getting updated. If the failed to PING the redis-server.

This approach could really solve the problem.

I would push this change on v0.15.1

[renchap]

Yes, you probably want 3 services:

• one targeting all pods, when you want to read on any pod
• one targeting only the primary, when you want to write
• maybe a 3rd service targeting all read-only pods, when you want to have heavy read operations and dont want to impact the primary performance

[shubham-cmyk]

Sounds Good! 3 services could be created here.

[sdwilsh]

@shubham-cmyk, is this work tracked in an issue somewhere?

[pierreozoux]

I believe, from what I understand, sentinel cluster has 2 jobs:

• promote a replica to primary
• tell clients which one is the primary or the replica

For the first question, there is an issue about that topic here:
#775

For the second question, which is exactly the question of OP, I created this issue:
#1341