Replies: 3 comments 3 replies
|
A startup probe adds little. |
|
With the current liveness check a startup probe wouldn't change anything for brokers: liveness doesn't fail just because a broker is slow to start. Here's the detail behind that. What the probes do (1.3.1):
The restarts you saw with ZooKeeper fit the old checks. The table in proposal 046 shows that a ZK-based broker's liveness probe needed a ZooKeeper session until the broker was ready, and the listener on 9091 after that. A broker that couldn't get or keep a ZK session while starting failed liveness and was restarted. That's the case a startup probe is for. Now a broker doing a long log recovery after an unclean shutdown keeps passing liveness, and only readiness stays false until it reaches RUNNING. The operator covers the rest. If a pod doesn't become ready during a rolling update, On the config side, The trade-off: a JVM that's alive but stuck won't be restarted by Kubernetes. It shows up as not ready instead, so alert on readiness or on the broker state rather than counting on liveness. |
|
@brundev I think you would need to be a lot more specific about what issues you are talking about to get a good answer. Without understanding what issue are you talking about, it I hard to say how/if startup probes would be useful. But I'm not aware of any outstanding reasons why to use startup probes in the Kafka nodes. Startup probes are in general good when your application startup takes a long, but clearly defined time. I do not think we have that in Kafka. But there are two main things to keep in mind:
|
Uh oh!
There was an error while loading. Please reload this page.
Hi everyone, I have a quick design question after reading proposal #46 and how the liveness probes are working in the brokers.
Basically I can that the liveness probe does a very simple check on the presence of the JVM.
I was wondering if the use of a startup probe was ever considered, as I have seen some issues in the past with the ZK version of Strimzi (and Kafka itself) that would go into restarts waiting for a good response from the brokers.
So is the startup probe a viable alternative or is the community set on the JVM check for a particular reason?
Thank you!
All reactions