[
https://issues.apache.org/jira/browse/HDFS-16902?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Viraj Jasani updated HDFS-16902:
--------------------------------
Description:
Recently came across an k8s environment where randomly some datanode pods are
not able to stay connected to all namenode pods (e.g. last heartbeat time stays
higher than 2 hr sometimes). When any standby namenode becomes active, any
datanode that is not heartbeating to it for quite sometime would not be able to
send any further block reports, leading to missing replicas immediately after
namenode failover, which could only be resolved with datanode pod restart.
While the issue seems env specific, BPServiceActor's offer service could use
some logging improvements. It is also good to get namenode status exposed with
BPServiceActorInfo to identify any lags from datanode side in recognizing
updated Active namenode status with heartbeats.
was:
Recently came across an k8s environment where randomly some datanode pods are
not able to stay connected to all namenode pods (e.g. last heartbeat time stays
higher than 2 hr sometimes). When new namenode becomes active, any datanode
that is not heartbeating to it would not be able to send any further block
reports, leading to missing replicas sometimes, which would be resolved only
with datanode pod restart.
While the issue seems env specific, BPServiceActor's offer service could use
some logging improvements. It is also good to get namenode status exposed with
BPServiceActorInfo to identify any lags from datanode side in recognizing
updated Active namenode status with heartbeats.
> Add Namenode status to BPServiceActor metrics and improve logging in
> offerservice
> ---------------------------------------------------------------------------------
>
> Key: HDFS-16902
> URL: https://issues.apache.org/jira/browse/HDFS-16902
> Project: Hadoop HDFS
> Issue Type: Task
> Reporter: Viraj Jasani
> Assignee: Viraj Jasani
> Priority: Major
>
> Recently came across an k8s environment where randomly some datanode pods are
> not able to stay connected to all namenode pods (e.g. last heartbeat time
> stays higher than 2 hr sometimes). When any standby namenode becomes active,
> any datanode that is not heartbeating to it for quite sometime would not be
> able to send any further block reports, leading to missing replicas
> immediately after namenode failover, which could only be resolved with
> datanode pod restart.
> While the issue seems env specific, BPServiceActor's offer service could use
> some logging improvements. It is also good to get namenode status exposed
> with BPServiceActorInfo to identify any lags from datanode side in
> recognizing updated Active namenode status with heartbeats.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
