Some broker pods get stuck caused by a terminated container when there is no healthy broker in the kafka cluster. #802
Conversation
bartam1
changed the title
pkg/resources/kafka/kafka.go
Outdated
| if err = r.reconcilePerBrokerDynamicConfig(broker.Id, brokerConfig, configMap, log); err != nil { | ||
| return err | ||
| perBrokerDynamicConfigCombineError = errors.Combine(perBrokerDynamicConfigCombineError, err) |
There was a problem hiding this comment.
We could add some details for the error like broker.Id, brokerConfig, configMap if it's not already included.
pkg/resources/kafka/kafka.go
Outdated
| } | ||
| } | ||
|
|
||
| if perBrokerDynamicConfigCombineError != nil { | ||
| return perBrokerDynamicConfigCombineError |
There was a problem hiding this comment.
If any of the other operations failed in the loop, e.g.
koperator/pkg/resources/kafka/kafka.go
Line 247 in f0fc247
koperator/pkg/resources/kafka/kafka.go
Line 255 in f0fc247
koperator/pkg/resources/kafka/kafka.go
Line 262 in f0fc247
And perBrokerDynamicConfigCombineError is not nil because there was some problems reconciling the broker dynamic configs for the previous brokers, then this perBrokerDynamicConfigCombineError is not going to be returned, which doesn't seem to be ideal to me...
|
I have disabled the cyclo check for the reconcile function temporarily. |
What's in this PR?
If reconcilePerBrokerDynamicConfig gets an error then I let the loop continue to the next broker. A combined error will be returned after the broker loop so the operator will try again later.
Why?
When in a kafka cluster there is not any working kafka broker present, but there are multiple broker pods with terminated container, only one broker pod will be restarted. Other brokers with terminated containers will not until at least one healthy broker will be present in the kafka cluster. The root cause of this that brokers are unreachable and reconcilePerBrokerDynamicConfig function can't set dynamic configs and It will be ended with an error.
Checklist