[HBASE-16713] Bring back connection caching as a client API

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Fix Version/s: 3.0.0-beta-2
• Component/s: Client, spark
• Assignee: Unassigned
• Reporter: Enis Soztutar
• Created: 26/Sep/16 23:43
• Updated: 16/Dec/23 09:45

Description

Connection.getConnection() is removed in master for good reasons. The connection lifecycle should always be explicit. We have replaced some of the functionality with ConnectionCache for rest and thrift servers internally, but it is not exposed to clients.

Turns out our friends doing the hbase-spark connector work needs a similar connection caching behavior that we have in rest and thrift server. At a higher level we want:

• Spark executors should be able to run short living hbase tasks with low latency
• Short living tasks should be able to share the same connection, and should not pay the price of instantiating the cluster connection (which means zk connection, meta cache, 200+ threads, etc)
• Connections to the cluster should be closed if it is not used for some time. Spark executors are used for other tasks as well.
• Spark jobs may be launched with different configuration objects, possibly connecting to different clusters between different jobs.
• Although not a direct requirement for spark, different users should not share the same connection object.

Looking at the old code that we have in branch-1 for ConnectionManager, managed connections and the code in ConnectionCache, I think we should do a first-class client level API called ConnectionCache which will be a hybrid between ConnectionCache and old ConnectionManager. The lifecycle of the ConnectionCache is still explicit, so I think API-design-wise, this will fit into the current model.

Issue Links

• is related to: HBASE-16638 Reduce the number of Connection's created in classes of hbase-spark module
• is related to: HBASE-17009 Revisiting the removement of managed connection and connection caching