Hi all, As part of a customer incident report we have found and recently I've filed HIVE-28977 <https://issues.apache.org/jira/browse/HIVE-28977> to optimize the cleanup of the Hive delegation tokens.
Currently we can have 3 types of token stores: memory, DB, ZK. Initially Hive was keeping the DTs only in memory, and the initial design did the cleanup as a local java thread. Since typically we have a standalone metastore, and the tokens are stored centrally (in the HMS database or in ZK), the original design is not efficient as all the HMS and HS2 instances are trying to do the same. The point of the Hive jira ticket is to do the DT cleanup from a single place, from the leader HMS instance. Q1. Per my understanding we can optimize only the DBTokenstore cleanup, but if you agree we could also include the ZKTokenstore too. However I see some risk in including the ZK tokenstore, as one can configure different ZNodes for different (HS2/HMS) instances. Do you think if anyone does that, or can we assume that the same type of tokenstore and same ZNode is used within a single cluster? Q2. As we still need to keep the old behavior with the memory tokenstore, the code should be fairly untouched, we cannot simply move the code only under the metastore maintenance threads. Because of that I am thinking that the solution / new code (as a new HMS maintenance "remote" thread) should just simply "allow" or "disallow" the run of the existing cleanup thread, so no major refactor would be needed. Do you think if that is OK? Thanks for sharing your thoughts on this. Miklos
