Ok, great. Good to see you guys are on top of this one!


Henry Robinson wrote:
The patch is simple, but a test is harder because it's a race condition. I'm
working on it.


On Wed, Nov 11, 2009 at 12:06 AM, Flavio Junqueira <f...@yahoo-inc.com>wrote:

Henry already opened one: ZOOKEEPER-569.


On Nov 11, 2009, at 7:03 AM, Patrick Hunt wrote:

 Closing the loop - what's the status on this? Can one of you open a
JIRA and provide a patch for this?



Flavio Junqueira wrote:

Hi Henry, Apologies for the the delay. Your observation sounds right to
me. Here is how I'm reading it; let me know if it makes sense.

If everyone votes for 3 in the second round and 3 has crashed, then in
countVotes we will remove all votes to 3 and there will be no vote left.
In such a case, there will be no winner as a result of the call to
countVotes and lookForLeader won't change the current vote
(LeaderElection.java:201). This is a situation in which we are stuck.

Does it sound reasonable to add an "else" to the "if" statement of
LeaderElection.java:201 to reset the vote? This modification would
implementing resetting the vote when countVotes returns no winner, which
should happen only when the replica itself votes for a dead leader.


On Oct 28, 2009, at 7:44 AM, Henry Robinson wrote:

 [ Sending this direct since the Apache mailserver is rejecting my
e-mails at the moment ]

As I understand it, 1 and 2 receive a vote for 3 in the first round,
which causes them to vote for 3 in the second round. So in the second
round, all votes cast are for 3. But 3 has died, so all votes for it
are discounted. 1 and 2 continue to vote for 3 ad infinitum, never
resetting their vote.

Does this sound plausible, or am I missing something?


On Tue, Oct 27, 2009 at 3:48 PM, Flavio Junqueira <f...@yahoo-inc.com>
Hi Henry, I don't understand how 1 and 2 do not end up electing 2 in
your situation. If they exclude 3 in countVotes, then countVotes will
end up returning 2 and not 3, assuming there is a vote for 2. What am
I missing?

The problem with QuorumPeer you're pointing at was also an issue with
the FLE tests, and I couldn't see an easy way around it other than
timing out and restarting leader election.


On Oct 27, 2009, at 6:35 AM, Henry Robinson wrote:

I've been working on adding a TCPResponderThread to the leader election
process so that if a deployment needs to be TCP only, it can be and
use all election types. Testing this has exposed what might be a race
condition in the leader election code that prevents a leader from being

Here's the behaviour I see in LETest occasionally. With three nodes
from 30 for ease of debugging), node 3 gets elected before either node
1 or
node 2 finish their election (there is one round where each node that
3 has
the highest id, and then 3 completes its second round by receiving
votes for
itself from 1 and 2, but 1 and 2 do not receive votes from 3).

Now 3 is killed by the test harness. 1 and 2 are still voting for it,
every time they try, the vote tally excludes 3 since it hasn't been
from. They then spin round the voting process, unable to reset their
vote. I
expect that the heartbeat mechanism in a running QuorumPeer takes care
this when the leader is lost, but the associated QuorumPeers aren't

If this is the case, then there is a simple fix to reset the nodes
vote to
themselves if they are voting for a node that hasn't been heard from. I
don't know why using TCP instead of UDP for the responder thread is
exacerbating this (and we can't rule out my introducing a bug :)); but
it's a race condition the different timings associated with waiting on
socket might just be enough to expose the issue.

Can someone verify this might be possible / figure out what I missed?


Reply via email to