Chris Thunes commented on ZOOKEEPER-2985:

The ephemeral nodes do eventually get removed once the new ZK leader marks the 
session as expired and performs the associated session tear down.

One fix may be to have the server close the client connection, _without_ 
sending the Expired event, if it finds the session is in the closing state with 
an uncommitted closeSession entry. Alternatively, session re-validation could 
be blocked for "closing" sessions until their corresponding closeSession entry 
is committed.

> Expired session may unexpired after leader failover
> ---------------------------------------------------
>                 Key: ZOOKEEPER-2985
>                 URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2985
>             Project: ZooKeeper
>          Issue Type: Bug
>    Affects Versions: 3.5.3, 3.4.11
>            Reporter: Chris Thunes
>            Priority: Major
> We recently observed an inconsistency in our Kafka cluster which we tracked 
> down to ZooKeeper sessions expiring and then re-appearing after a ZooKeeper 
> leadership failover. The Kafka nodes received session "Expired" events, 
> leading to them starting new sessions and attempting to re-create some 
> ephemeral nodes (broker ID nodes in kafka/brokers/ids specifically). However, 
> between receiving the session Expired event and establishing a new session a 
> leadership failover occurred within the ZooKeeper cluster which resulted in 
> the expired session re-appearing. When Kafka attempted to re-create the 
> ephemeral nodes mentioned above it (unexpectedly) received NODEEXISTS errors.
> This behavior is a result of how session expiration is handled by the leader. 
> Specifically, the expired session is marked as "closing" immediately upon 
> expiration (in SessionTrackerImpl) and _before_ the corresponding 
> "closeSession" entry is committed. A client can therefore receive a session 
> Expired event before its session is fully closed. A leadership failover which 
> results in the loss of the (uncommitted) closeSession entry thus leads to the 
> sessions' ephemeral nodes "re-appearing" until another expiration of the old 
> session on the new leader takes place.
> I'm not certain if this should be considered a bug or an edge case that 
> client are expected to handle. If it is the latter then I think it would be 
> good to include this in the Programmer's Guide in the documentation.
> If it's helpful I have code to reproduce this on an in-process cluster 
> running 3.4.11 or 3.5.3-beta.

This message was sent by Atlassian JIRA

Reply via email to