helix-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From kishore g <g.kish...@gmail.com>
Subject Re: Listen for ZkStateChanges
Date Thu, 08 Aug 2013 22:41:11 GMT
Hi Kevin,

Thanks for the interest.

Helix can automatically detect flapping and disable the node if the the
number of disconnects exceeds a threshold within a time period. Will this
be sufficient ?

Temporary disconnection vs permanent one is interesting, this can be due to
GC/Network partition and detecting this reliably is challenging. We have a
jira to address this https://issues.apache.org/jira/browse/HELIX-26. It
will help if you add your requirements to this.

Even though its possible to expose ZkStateChangeListener, we want to avoid
exposing dependency on the library. We can definitely add a
NodeStateChangeListener and provide the callback when the state of the node
changes. We don't have this feature at this time, please file a jira and we
should be able to add it easily.

I am interested in what you plan to do when you detect the state changes.
The reason i ask this is, it is really tricky to handle zookeeper callbacks
in a reliable way and if possible Helix should deal with it properly and
invoke the appropriate set of transitions.

thanks,
Kishore G




On Thu, Aug 8, 2013 at 7:26 AM, Kevin Gao <kgao@box.com> wrote:

> Hi everyone,
>
> First off, I'm new to the mailing list. My name is Kevin, and I'm working
> over at Box. I'm excited to get to work with all of you!
>
> In the mean time, I do have one question regarding detection of ZooKeeper
> connection state changes. For our purposes, we need to know the state of
> the connection between Helix and ZooKeeper at all times to ensure proper
> handling of some failure scenarios. Currently, the model we use for
> detecting connectivity is a long poll to the HelixManager.isConnected()
> method; however, we would like to ideally be notified of any connection
> state changes.
>
> Essentially, I would like to be able to add a ZkStateChangeListener to the
> underlying ZkClient, or at least be able to subscribe to such event changes
> at the HelixManager level. A mechanism similar to
> HelixManager.addCurrentStateChangeListener() would work for example. Am I
> missing something obvious, or is this capability just not exposed?
>
> Reasons why we would like to do this include, for example:
>
> * Be able to externally know if a disconnect is due to a flapping issue
> * Be able to differentiate between a temporary disconnection vs. a
> permanent one, or one of longer duration
>
> Any suggestions?
>
> Thanks a lot,
> Kevin
>
> --
> Kevin Gao
> kgao@box.com
>

Mime
View raw message