When an Elasticsearch
Pod is terminated, its
Endpoint is removed from the
Service and the Elasticsearch process is terminated. As these two operations happen in parallel, a race condition exists. If the Elasticsearch process is already shut down, but the
Endpoint is still a part of the
Service, any new connection might fail. For more information, see Termination of pods.
Moreover, kube-proxy resynchronizes its rules every 30 seconds by default. During that 30 second time window, the terminating Pod IP may still be used when targeting the service. Please note the resync operation itself may take some time, especially if kube-proxy is configured to use iptables with a lot of services and rules to apply.
To address this issue and minimise unavailability, ECK relies on a PreStop lifecycle hook. This waits to terminate the Elasticsearch process until the
Service DNS record does not contain the IP of the
First, the PreStop lifecycle hook will keep querying DNS for
PRE_STOP_MAX_WAIT_SECONDS (defaulting to 20) until the Pod IP is not referenced anymore.
Then, it waits for an additional
PRE_STOP_ADDITIONAL_WAIT_SECONDS (defaulting to 30). Additional wait is used to:
- give time to in-flight requests to be completed
- give clients time to use the terminating Pod IP resolved just before DNS record was updated
- give kube-proxy time to refresh ipvs or iptables rules on all nodes, depending on its sync period setting
The exact behavior is configurable using environment variables, for example:
spec: version: 8.12.2 nodeSets: - name: default count: 1 podTemplate: spec: containers: - name: elasticsearch env: - name: PRE_STOP_MAX_WAIT_SECONDS value: "10" - name: PRE_STOP_ADDITIONAL_WAIT_SECONDS value: "5"