Abhishek Singh Chouhan created ZOOKEEPER-2785:
-------------------------------------------------
Summary: Server inappropriately throttles connections under load
before SASL completes
Key: ZOOKEEPER-2785
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2785
Project: ZooKeeper
Issue Type: Bug
Affects Versions: 3.4.10
Reporter: Abhishek Singh Chouhan
Priority: Critical
Fix For: 3.4.11
When a zk server is running close to its outstanding requests limit, the server
incorrectly throttles the sasl request. This leads to the client waiting for
the final sasl packet (session is already established) and deferring all non
priming packets till then which also includes the ping packets. The client then
waits for the final packet but never gets it and times out saying haven't heard
from server. This is fatal for services such as HBase which retry for finite
attempts and exit post these attempts.
Issue being that in ZooKeeperServer.processPacket(..) incase of sasl we send
the response and incorrectly also call cnxn.incrOutstandingRequests(h), which
throttles the connection if we're running over outstandingrequests limit, which
results in the server not processing the subsequent packet from the client.
Also we donot have any pending request to send for the connection and hence
never call enableRecv(). We should return after sending response to the sasl
request.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
