Hi ZooKeeper users,

I am playing currently with zookeeper and testing what happens if the
leader of an ensemble goes down.
I know that during the leader election zookeeper server won't reply to
any requests and if leader election takes a long time than existing
sessions might expire.
What I see in my tests that each server reads the last snapshot file
to get last zxid for leader election and when the leader is elected
than the leader reads the snapshot again before it syncs the
followers.

This means that the more data we store in zookeeper the longer it
takes to elect a new leader. This is also means as load of the
ensemble increases clients need bigger session timeout to "survive"
the loss of the leader.

Is it possible to do anything about this and have a fast leader
election even if the snapshot is big?

Regards,
Botond Hejj

Existing sessions will not expire from the server side during election.
Your client code may choose to close them on its end if you sit in a
DISCONNECTED state for too long, but nothing should be expiring the
sessions while quorum is not available.

C

On Mon, Dec 12, 2011 at 12:06 PM, Botond Hejj <[EMAIL PROTECTED]
> wrote:

> Hi ZooKeeper users,
>
> I am playing currently with zookeeper and testing what happens if the
> leader of an ensemble goes down.
> I know that during the leader election zookeeper server won't reply to
> any requests and if leader election takes a long time than existing
> sessions might expire.
> What I see in my tests that each server reads the last snapshot file
> to get last zxid for leader election and when the leader is elected
> than the leader reads the snapshot again before it syncs the
> followers.
>
> This means that the more data we store in zookeeper the longer it
> takes to elect a new leader. This is also means as load of the
> ensemble increases clients need bigger session timeout to "survive"
> the loss of the leader.
>
> Is it possible to do anything about this and have a fast leader
> election even if the snapshot is big?
>
> Regards,
> Botond Hejj
>

Thanks Camille,

Sorry, my assumption was wrong. I've made now a test and indeed the
session doesn't expire in this case.
This means than the only problem is that the service is down for an
increasing amount of time if snapshot is increasing. We use SAN to
store the snapshot and reading it back twice during leader election
can be a little slow (~30 secs). It is not bad at all but still have
anybody tried making the leader election faster in this case?

Regards,
Botond

On Mon, Dec 12, 2011 at 18:35, Camille Fournier <[EMAIL PROTECTED]> wrote:
> Existing sessions will not expire from the server side during election.
> Your client code may choose to close them on its end if you sit in a
> DISCONNECTED state for too long, but nothing should be expiring the
> sessions while quorum is not available.
>
> C
>
> On Mon, Dec 12, 2011 at 12:06 PM, Botond Hejj <[EMAIL PROTECTED]
>> wrote:
>
>> Hi ZooKeeper users,
>>
>> I am playing currently with zookeeper and testing what happens if the
>> leader of an ensemble goes down.
>> I know that during the leader election zookeeper server won't reply to
>> any requests and if leader election takes a long time than existing
>> sessions might expire.
>> What I see in my tests that each server reads the last snapshot file
>> to get last zxid for leader election and when the leader is elected
>> than the leader reads the snapshot again before it syncs the
>> followers.
>>
>> This means that the more data we store in zookeeper the longer it
>> takes to elect a new leader. This is also means as load of the
>> ensemble increases clients need bigger session timeout to "survive"
>> the loss of the leader.
>>
>> Is it possible to do anything about this and have a fast leader
>> election even if the snapshot is big?
>>
>> Regards,
>> Botond Hejj
>>

--
Botond Hejj
Morgan Stanley | Technology
Lechner Odon fasor 8 | Floor 07
Budapest, 1095
Phone: +36 1 881-3962
[EMAIL PROTECTED]

Hi Botond, I'm under the impression that the leader only loads the  
database twice when it is bootstrapping. In QuorumPeer.start(), we  
call loadDataBase(). The second time corresponds to the call to  
zk.loadData in Leader.lead(). There could be a third time if  
zkDb.isInitialized() is false thought the call path lookForLeader()->  
getInitLastLoggedZxid() -> getLastLoggedZxid(). I don't see how it  
could be false, though.

After we get into the main QuorumPeer loop (in run()), we only call  
loadDataBase() through lookForLeader() upon calling  
getInitLastLoggedZxid() -> getLastLoggedZxid(). loadDataBase() is only  
executed if zkDb.isInitialized() is false, and I actually don't see a  
case in which it would be true once for the call coming from  
lookForLeader().

The part I can't remember is why we need to call loadData() in  
Leader.lead() in the case that the data tree has been initialized  
already. If we are somehow skipping it when it has been initialized  
already, then I missed it. I'll keep looking into it...

-Flavio

On Dec 13, 2011, at 11:55 AM, Botond Hejj wrote:

> Thanks Camille,
>
> Sorry, my assumption was wrong. I've made now a test and indeed the
> session doesn't expire in this case.
> This means than the only problem is that the service is down for an
> increasing amount of time if snapshot is increasing. We use SAN to
> store the snapshot and reading it back twice during leader election
> can be a little slow (~30 secs). It is not bad at all but still have
> anybody tried making the leader election faster in this case?
>
> Regards,
> Botond
>
> On Mon, Dec 12, 2011 at 18:35, Camille Fournier <[EMAIL PROTECTED]>  
> wrote:
>> Existing sessions will not expire from the server side during  
>> election.
>> Your client code may choose to close them on its end if you sit in a
>> DISCONNECTED state for too long, but nothing should be expiring the
>> sessions while quorum is not available.
>>
>> C
>>
>> On Mon, Dec 12, 2011 at 12:06 PM, Botond Hejj <[EMAIL PROTECTED]
>>> wrote:
>>
>>> Hi ZooKeeper users,
>>>
>>> I am playing currently with zookeeper and testing what happens if  
>>> the
>>> leader of an ensemble goes down.
>>> I know that during the leader election zookeeper server won't  
>>> reply to
>>> any requests and if leader election takes a long time than existing
>>> sessions might expire.
>>> What I see in my tests that each server reads the last snapshot file
>>> to get last zxid for leader election and when the leader is elected
>>> than the leader reads the snapshot again before it syncs the
>>> followers.
>>>
>>> This means that the more data we store in zookeeper the longer it
>>> takes to elect a new leader. This is also means as load of the
>>> ensemble increases clients need bigger session timeout to "survive"
>>> the loss of the leader.
>>>
>>> Is it possible to do anything about this and have a fast leader
>>> election even if the snapshot is big?
>>>
>>> Regards,
>>> Botond Hejj
>>>
>
>
>
> --
> Botond Hejj
> Morgan Stanley | Technology
> Lechner Odon fasor 8 | Floor 07
> Budapest, 1095
> Phone: +36 1 881-3962
> [EMAIL PROTECTED]

flavio
junqueira

research scientist

[EMAIL PROTECTED]
direct +34 93-183-8828

avinguda diagonal 177, 8th floor, barcelona, 08018, es
phone (408) 349 3300    fax (408) 349 3301

Thanks Flavio for looking into this.
If the loadData can be removed from lead() that could be improve
leaderElection a lot on slow disk with high amount of data.

Botond

On Tue, Dec 13, 2011 at 14:02, Flavio Junqueira <[EMAIL PROTECTED]> wrote:
> Hi Botond, I'm under the impression that the leader only loads the database
> twice when it is bootstrapping. In QuorumPeer.start(), we call
> loadDataBase(). The second time corresponds to the call to zk.loadData in
> Leader.lead(). There could be a third time if zkDb.isInitialized() is false
> thought the call path lookForLeader()-> getInitLastLoggedZxid() ->
> getLastLoggedZxid(). I don't see how it could be false, though.
>
> After we get into the main QuorumPeer loop (in run()), we only call
> loadDataBase() through lookForLeader() upon calling getInitLastLoggedZxid()
> -> getLastLoggedZxid(). loadDataBase() is only executed if
> zkDb.isInitialized() is false, and I actually don't see a case in which it
> would be true once for the call coming from lookForLeader().
>
> The part I can't remember is why we need to call loadData() in Leader.lead()
> in the case that the data tree has been initialized already. If we are
> somehow skipping it when it has been initialized already, then I missed it.
> I'll keep looking into it...
>
> -Flavio
>
>
> On Dec 13, 2011, at 11:55 AM, Botond Hejj wrote:
>
>> Thanks Camille,
>>
>> Sorry, my assumption was wrong. I've made now a test and indeed the
>> session doesn't expire in this case.
>> This means than the only problem is that the service is down for an
>> increasing amount of time if snapshot is increasing. We use SAN to
>> store the snapshot and reading it back twice during leader election
>> can be a little slow (~30 secs). It is not bad at all but still have
>> anybody tried making the leader election faster in this case?
>>
>> Regards,
>> Botond
>>
>> On Mon, Dec 12, 2011 at 18:35, Camille Fournier <[EMAIL PROTECTED]>
>> wrote:
>>>
>>> Existing sessions will not expire from the server side during election.
>>> Your client code may choose to close them on its end if you sit in a
>>> DISCONNECTED state for too long, but nothing should be expiring the
>>> sessions while quorum is not available.
>>>
>>> C
>>>
>>> On Mon, Dec 12, 2011 at 12:06 PM, Botond Hejj
>>> <[EMAIL PROTECTED]
>>>>
>>>> wrote:
>>>
>>>
>>>> Hi ZooKeeper users,
>>>>
>>>> I am playing currently with zookeeper and testing what happens if the
>>>> leader of an ensemble goes down.
>>>> I know that during the leader election zookeeper server won't reply to
>>>> any requests and if leader election takes a long time than existing
>>>> sessions might expire.
>>>> What I see in my tests that each server reads the last snapshot file
>>>> to get last zxid for leader election and when the leader is elected
>>>> than the leader reads the snapshot again before it syncs the
>>>> followers.
>>>>
>>>> This means that the more data we store in zookeeper the longer it
>>>> takes to elect a new leader. This is also means as load of the
>>>> ensemble increases clients need bigger session timeout to "survive"
>>>> the loss of the leader.
>>>>
>>>> Is it possible to do anything about this and have a fast leader
>>>> election even if the snapshot is big?
>>>>
>>>> Regards,
>>>> Botond Hejj
>>>>
>>
>>
>>
>> --
>> Botond Hejj
>> Morgan Stanley | Technology
>> Lechner Odon fasor 8 | Floor 07
>> Budapest, 1095
>> Phone: +36 1 881-3962
>> [EMAIL PROTECTED]
>
>
> flavio
> junqueira
>
> research scientist
>
> [EMAIL PROTECTED]
> direct +34 93-183-8828
>
> avinguda diagonal 177, 8th floor, barcelona, 08018, es
> phone (408) 349 3300    fax (408) 349 3301
>

--
Botond Hejj
Morgan Stanley | Technology
Lechner Odon fasor 8 | Floor 07
Budapest, 1095
Phone: +36 1 881-3962
[EMAIL PROTECTED]