Description
Recently we purge the complicated RecoverableZooKeeper dependency in hbase-client and use curator to get data from zookeeper.
But when debugging HBASE-19266, we found that if we get data immediately after the creation of CuratorFramework, the request will always cost 100x ms to complete.
After digging, we found that there is a sleep in CuratorFrameworkImpl if the connection is not established yet
This is really a bad news for us. We decide to make the async hbase client fully asynchronous. You can see this example, where we execute the request to hbase directly in the netty event loop thread.
So if there is a sleep in foreground then the event loop will be stuck for 1 second which will cause very bad performance impact...
Attachments
Issue Links
- breaks
-
HBASE-19312 Find out why sometimes we need to spend more than one second to get the cluster id
- Resolved
- links to