zookeeper-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Flavio Junqueira <...@apache.org>
Subject Re: Working around Leader election Listner thread death
Date Tue, 30 Aug 2016 23:54:05 GMT
I'm not sure what you're suggesting, David. Could you be more specific, please?

-Flavio

> On 30 Aug 2016, at 23:54, David Brower <david.brower@oracle.com> wrote:
> 
> Anything you could do with iptables you can do in the process by having it drop connections
from things not on a whitelist, and not having a thread waiting indefinitely for operations
from any connection.
> 
> -dB
> 
> 
> On 8/30/2016 2:46 PM, Flavio Junqueira wrote:
>> I was trying to write down an analysis and I haven't been able to come up with anything
that is foolproof. Basically, the two main issues are:
>> 
>> - A bad server is able to connect to a good server in the case it has a message outstanding
and is trying to establish a connection to the good server. This happens if the server is
LOOKING or has an outstanding message from the previous round. The converse isn't true, though.
A good server can't start a connection to a bad server because the bad server doesn't have
a listener.
>> - If we bounce servers sequentially, there is a chance that a bad server is elected
more than once along the process, which induces multiple leader election rounds.
>> 
>> Perhaps this is overkill, but I was wondering if it makes sense to filter election
traffic to and from bad servers using, for example, iptables. The idea is to a rule that are
local to each server preventing the server to get connections established for leader election.
For each bad server, we stop it, remove the rule, and bring it back up. We also stop a minority
first before stoping the bad leader.
>> 
>> -Flavio
>> 
>>> On 29 Aug 2016, at 09:29, Guy Laden <guy.laden@gmail.com> wrote:
>>> 
>>> Hi Flavio, Thanks for your reply. The situation is that indeed all the
>>> servers are in a bad state so it looks like we will have to perform a
>>> cluster restart.
>>> 
>>> We played with attempts to optimize the downtime along the lines you
>>> suggested. In testing it we ran into the issue where a server with no
>>> Listener thread can initiate a leader election connection to a
>>> newly-restarted server that does have a Listener. The result is a quorum
>>> that may include 'bad' servers, even a 'bad' leader. So we tried to first
>>> restart the higher-id servers, because lower-id servers will drop their
>>> leader-election connections to higher id servers.
>>> I'm told there are issues with this flow as well but have not yet
>>> investigated the details.
>>> I also worry about the leader-election retries done with exponential
>>> backoff.
>>> 
>>> I guess we will play with things a bit more but at this point I am tending
>>> towards a simple parallel restart of all servers..
>>> 
>>> Once the clusters are healthy again we will do a rolling upgrade to 3.4.8
>>> sometime soon.
>>> 
>>> Thanks again,
>>> Guy
>>> 
>>> 
>>> On Sun, Aug 28, 2016 at 5:52 PM, Flavio Junqueira <fpj@apache.org> wrote:
>>> 
>>>> Hi Guy,
>>>> 
>>>> We don't have a way to restart the listener thread, so you really need to
>>>> bounce the server. I don't think there is a way of doing this without
>>>> forcing a leader election, assuming all your servers are in this bad state.
>>>> To minimize downtime, one thing you can do is to avoid bouncing the current
>>>> leader until it loses quorum support. Once it loses quorum support, you
>>>> have a quorum of healthy servers and they will elect a new, healthy leader.
>>>> At the point, you can bounce all your unhealthy servers.
>>>> 
>>>> You may also want to move to a later 3.4 release.
>>>> 
>>>> -Flavio
>>>> 
>>>>> On 24 Aug 2016, at 23:15, Guy Laden <guy.laden@gmail.com> wrote:
>>>>> 
>>>>> Hi all,
>>>>> 
>>>>> It looks like due to a security scan sending "bad" traffic to the leader
>>>>> election port, we have clusters in which
>>>>> the leader election Listener thread is dead (unchecked exception was
>>>> thrown
>>>>> and thread died - seen in the log).
>>>>> (This seems to be fixed by fixed in
>>>>> https://issues.apache.org/jira/browse/ZOOKEEPER-2186)
>>>>> 
>>>>> In this state, when a healthy server comes up and tries to connecnt to
>>>> the
>>>>> quorum, it gets stuck on
>>>>> the leader election. It establishes TCP connections to the other servers
>>>>> but any traffic it sends seems
>>>>> to get stuck in the receiver's TCP Recv queue (seen with netstat), and
is
>>>>> not read/processed by zk.
>>>>> 
>>>>> Not a good place to be :)
>>>>> 
>>>>> This is with 3.4.6
>>>>> 
>>>>> Is there a way to get such clusters back to a healthy state without loss
>>>> of
>>>>> quorum / client impact?
>>>>> Some way of re-starting the listener thread? or restarting the servers
>>>> in a
>>>>> certain order?
>>>>> e.g. If I restart a minority, say the ones with lower server id's - is
>>>>> there a way to get the majority servers
>>>>> to re-initiate leader election connections with them so as to connect
>>>> them
>>>>> to the quorum? (and to do this without
>>>>> the majority losing quorum).
>>>>> 
>>>>> Thanks,
>>>>> Guy
>>>> 
> 


Mime
View raw message