You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@zookeeper.apache.org by Flavio Junqueira <fp...@yahoo-inc.com> on 2010/11/24 09:40:00 UTC
ZOOKEEPER-702
Hi, ZOOKEEPER-702 changes the way we do failure detection in
ZooKeeper, and it is an important change. I'd like to see it committed
soon (it has been sitting there for a while), but at the same time I'd
like to make sure that no one has an issue with it. If this is
interesting to you or affects you in any way, please consider having a
look at it soon.
Thanks,
-Flavio
Re: ZOOKEEPER-702
Posted by Gustavo Niemeyer <gu...@niemeyer.net>.
Hi Abmar,
> My implementation for the Phi-Accrual is based on the Cassandra's
> failure detector, which in turn is based on the ED FD, that is an
> extension of the original Phi-Accrual paper. The ED FD states that
> an exponential distribution
In the comments you said that it'd be possible to select which
distribution function to use. Is that still the case?
> fits the ping inter-arrival sampling window better than a normal
> one. Indeed, I have done some experiments that confirms that changes
> on the Phi parameter make a greater impact on the average detection
> time when using the exponential distribution.
That's interesting. Given that exponential distribution considers a
significantly larger portion of the data points to be under a single
standard deviation, I guess it's not too surprising it'll reduce the
detection time for failures. The question then, obviously, is how
often will it get it wrong.
The Phi paper presents some data regarding tests they've done with
a live link. Are you aware of any similar papers presenting the
alternative model? I'd certainly appreciate reading about it.
> Regarding the C client, I have already opened a jira
> (https://issues.apache.org/jira/browse/ZOOKEEPER-848) for that. I plan to
> start working on this asap.
That's exciting news. Thanks a lot!
--
Gustavo Niemeyer
http://niemeyer.net
http://niemeyer.net/blog
http://niemeyer.net/twitter
Re: ZOOKEEPER-702
Posted by Abmar Barros <ab...@gmail.com>.
Hi Gustavo,
My implementation for the Phi-Accrual is based on the Cassandra's failure
detector, which in turn is based on the ED FD, that is an extension of the
original Phi-Accrual paper. The ED FD states that an exponential
distribution fits the ping inter-arrival sampling window better than a
normal one. Indeed, I have done some experiments that confirms that changes
on the Phi parameter make a greater impact on the average detection time
when using the exponential distribution.
Regarding the C client, I have already opened a jira (
https://issues.apache.org/jira/browse/ZOOKEEPER-848) for that. I plan to
start working on this asap.
On Wed, Nov 24, 2010 at 6:43 AM, Flavio Junqueira <fp...@yahoo-inc.com> wrote:
> True, the jira refers to Cassandra. I think Abmar has just experimented
> with it, but it is certainly best if he clarifies it.
>
> On the C client, that's a very good point. In one of my reviews I pointed
> it out, but I couldn't find another comment talking about it. I'm not aware
> of anyone working on it. We should create a sub-task for it and Abmar could
> perhaps clarify if he is planning on producing a patch for the C client.
>
> Please feel free to add to the jira so that we keep it documented, Gustavo.
>
> -Flavio
>
> On Nov 24, 2010, at 10:26 AM, Gustavo Niemeyer wrote:
>
> Hi Gustavo, I don't know if Abmar was using Cassandra as reference
>
> (it doesn't sound like from the jira comments), but I followed the paper.
>
> Feel free to double-check it if you have a chance.
>
>
> The description references Cassandra explicitly, which is why I
> wondered. But then, his comments down the JIRA mention normal vs.
> exponential distribution, which means he must certainly be aware of
> the subtleties.
>
> Also, note that it won't replace the current mechanism with phi accrual,
> but it will enable the use of phi accrual.
>
>
> I'd certainly like to experiment with it at some point. Phi enables
> the distinction between "slowing down" and "most likely broken", which
> is an interesting feature. It isn't clear from the JIRA if someone is
> working on the C client, though. Do you know if there's any
> conversation regarding this?
>
> --
> Gustavo Niemeyer
> http://niemeyer.net
> http://niemeyer.net/blog
> http://niemeyer.net/twitter
>
>
> *flavio*
> *junqueira*
>
> research scientist
>
> fpj@yahoo-inc.com
> direct +34 93-183-8828
>
> avinguda diagonal 177, 8th floor, barcelona, 08018, es
> phone (408) 349 3300 fax (408) 349 3301
>
>
>
--
Abmar Barros
MSc candidate on Computer Science at Federal University of Campina Grande -
www.ufcg.edu.br
OurGrid Team Member - www.ourgrid.org
Paraíba - Brazil
Re: ZOOKEEPER-702
Posted by Flavio Junqueira <fp...@yahoo-inc.com>.
True, the jira refers to Cassandra. I think Abmar has just
experimented with it, but it is certainly best if he clarifies it.
On the C client, that's a very good point. In one of my reviews I
pointed it out, but I couldn't find another comment talking about it.
I'm not aware of anyone working on it. We should create a sub-task for
it and Abmar could perhaps clarify if he is planning on producing a
patch for the C client.
Please feel free to add to the jira so that we keep it documented,
Gustavo.
-Flavio
On Nov 24, 2010, at 10:26 AM, Gustavo Niemeyer wrote:
>> Hi Gustavo, I don't know if Abmar was using Cassandra as reference
>> (it doesn't sound like from the jira comments), but I followed the
>> paper.
>> Feel free to double-check it if you have a chance.
>
> The description references Cassandra explicitly, which is why I
> wondered. But then, his comments down the JIRA mention normal vs.
> exponential distribution, which means he must certainly be aware of
> the subtleties.
>
>> Also, note that it won't replace the current mechanism with phi
>> accrual, but it will enable the use of phi accrual.
>
> I'd certainly like to experiment with it at some point. Phi enables
> the distinction between "slowing down" and "most likely broken", which
> is an interesting feature. It isn't clear from the JIRA if someone is
> working on the C client, though. Do you know if there's any
> conversation regarding this?
>
> --
> Gustavo Niemeyer
> http://niemeyer.net
> http://niemeyer.net/blog
> http://niemeyer.net/twitter
flavio
junqueira
research scientist
fpj@yahoo-inc.com
direct +34 93-183-8828
avinguda diagonal 177, 8th floor, barcelona, 08018, es
phone (408) 349 3300 fax (408) 349 3301
Re: ZOOKEEPER-702
Posted by Gustavo Niemeyer <gu...@niemeyer.net>.
> Hi Gustavo, I don't know if Abmar was using Cassandra as reference
> (it doesn't sound like from the jira comments), but I followed the paper.
> Feel free to double-check it if you have a chance.
The description references Cassandra explicitly, which is why I
wondered. But then, his comments down the JIRA mention normal vs.
exponential distribution, which means he must certainly be aware of
the subtleties.
> Also, note that it won't replace the current mechanism with phi accrual, but it will enable the use of phi accrual.
I'd certainly like to experiment with it at some point. Phi enables
the distinction between "slowing down" and "most likely broken", which
is an interesting feature. It isn't clear from the JIRA if someone is
working on the C client, though. Do you know if there's any
conversation regarding this?
--
Gustavo Niemeyer
http://niemeyer.net
http://niemeyer.net/blog
http://niemeyer.net/twitter
Re: ZOOKEEPER-702
Posted by Flavio Junqueira <fp...@yahoo-inc.com>.
Hi Gustavo, I don't know if Abmar was using Cassandra as reference (it
doesn't sound like from the jira comments), but I followed the paper.
Feel free to double-check it if you have a chance.
Also, note that it won't replace the current mechanism with phi
accrual, but it will enable the use of phi accrual.
-Flavio
On Nov 24, 2010, at 10:06 AM, Gustavo Niemeyer wrote:
> Hi Flavio,
>
>> Hi, ZOOKEEPER-702 changes the way we do failure detection in
>> ZooKeeper, and
>> it is an important change. I'd like to see it committed soon (it
>> has been
>> sitting there for a while), but at the same time I'd like to make
>> sure that
>> no one has an issue with it. If this is interesting to you or
>> affects you in
>> any way, please consider having a look at it soon.
>
> As a minor comment, even though Cassandra claims to implement the Phi
> accrual failure detector, last time I looked they didn't really
> implement the algorithm which is in the paper. If someone is going to
> implement something similar, I suggest looking at the paper rather
> than Cassandra's code.
>
> --
> Gustavo Niemeyer
> http://niemeyer.net
> http://niemeyer.net/blog
> http://niemeyer.net/twitter
flavio
junqueira
research scientist
fpj@yahoo-inc.com
direct +34 93-183-8828
avinguda diagonal 177, 8th floor, barcelona, 08018, es
phone (408) 349 3300 fax (408) 349 3301
Re: ZOOKEEPER-702
Posted by Gustavo Niemeyer <gu...@niemeyer.net>.
> As a minor comment, even though Cassandra claims to implement the Phi
> accrual failure detector, last time I looked they didn't really
> implement the algorithm which is in the paper. If someone is going to
> implement something similar, I suggest looking at the paper rather
> than Cassandra's code.
Sorry, spoke too soon. This comment was based on the bug description
I got via email, but there are some good comments in the page which
indicate people are already aware of this distinction.
--
Gustavo Niemeyer
http://niemeyer.net
http://niemeyer.net/blog
http://niemeyer.net/twitter
Re: ZOOKEEPER-702
Posted by Gustavo Niemeyer <gu...@niemeyer.net>.
Hi Flavio,
> Hi, ZOOKEEPER-702 changes the way we do failure detection in ZooKeeper, and
> it is an important change. I'd like to see it committed soon (it has been
> sitting there for a while), but at the same time I'd like to make sure that
> no one has an issue with it. If this is interesting to you or affects you in
> any way, please consider having a look at it soon.
As a minor comment, even though Cassandra claims to implement the Phi
accrual failure detector, last time I looked they didn't really
implement the algorithm which is in the paper. If someone is going to
implement something similar, I suggest looking at the paper rather
than Cassandra's code.
--
Gustavo Niemeyer
http://niemeyer.net
http://niemeyer.net/blog
http://niemeyer.net/twitter