ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Andrey Kornev <andrewkor...@hotmail.com>
Subject RE: Cluster group affinity
Date Wed, 07 Oct 2015 12:14:13 GMT
Thanks, Yakov! I'm going to use your code (I hope it's Apache2 licensed:)))

As for the use case, it's quite simple. I have a group of compute nodes that are being sent
a stream of work items and I'd like to parallelize processing by partitioning the stream across
all the nodes in the group by some attribute. This is pure compute, no caches.

Thanks
Andrey

> Date: Wed, 7 Oct 2015 13:46:23 +0300
> Subject: Re: Cluster group affinity
> From: yzhdanov@apache.org
> To: dev@ignite.apache.org
> 
> Andrey, probably it is, but I am not sure if I have ever had a thought for
> mentioned scenario.
> 
> I think you should get existing implementation and use it like this. Make
> sure to cache assignments as this may be quiet expensive operation.
> 
>         Ignite ignite = Ignition.start(cfg);
> 
>         ClusterGroup group = ignite.cluster().forPredicate(new Predicate());
> 
>         final List<ClusterNode> snapshot = new ArrayList<>(group.nodes());
> 
>         RendezvousAffinityFunction aff = new RendezvousAffinityFunction();
> 
>         List<List<ClusterNode>> parts = aff.assignPartitions(new
> AffinityFunctionContext() {
>             @Nullable @Override public List<ClusterNode>
> previousAssignment(int part) {
>                 return null;
>             }
> 
>             @Override public int backups() {
>                 return 0;
>             }
> 
>             @Override public List<ClusterNode> currentTopologySnapshot() {
>                 return snapshot;
>             }
> 
>             @Override public AffinityTopologyVersion
> currentTopologyVersion() {
>                 return null;
>             }
> 
>             @Nullable @Override public DiscoveryEvent discoveryEvent() {
>                 return null;
>             }
>         });
> 
>         // Picking node.
>         ClusterNode node = parts.get(aff.partition(key)).get(0);
> 
> 
> --Yakov
> 
> 2015-10-07 11:39 GMT+03:00 Andrey Kornev <andrewkornev@hotmail.com>:
> 
> > Dmitriy,
> >
> > This approach would definitely work, if it wasn't for the fact that the
> > cluster groups in my case are created dynamically and may include any
> > combination of nodes in the cluster (where the number of combinations grows
> > exponentially with the number of nodes in the cluster). I don't think it's
> > practical to create that many caches.
> >
> > I still can't get why the affinity function can't be applied to an
> > arbitrary cluster group, and why it must necessarily be a cache. Isn't the
> > cache affinity just a special case of the cluster group affinity defined as
> > ClusterGroup.forCache()?
> >
> > Thanks
> > Andrey
> >
> > > From: dsetrakyan@apache.org
> > > Date: Tue, 6 Oct 2015 12:07:39 -0700
> > > Subject: Re: Cluster group affinity
> > > To: dev@ignite.apache.org
> > >
> > > On Tue, Oct 6, 2015 at 8:46 AM, Andrey Kornev <andrewkornev@hotmail.com>
> > > wrote:
> > >
> > > > Thanks, Andrey! This definitely helps.
> > > >
> > > > It's just that implementing such a simple feature in the "user space"
> > > > feels awkward and requires intimate knowledge of  fairly low-level
> > details
> > > > of how things work in the current version.
> > > >
> > > > Just curios, how about providing an override for Ignite.affinity()
> > method
> > > > that ClusterGroup? Is there something fundamentally wrong about
> > calculating
> > > > the affinity for an arbitrary collection of nodes (such as a
> > ClusterGroup
> > > > is)?
> > > >
> > >
> > > Affinity is usually associated with data. In your case you have no data,
> > > but you still need keys to be always mapped to the same node.  How about
> > > creating an empty cache and using standard cache API for determining the
> > > affinity for a key?
> > >
> > >
> > > > Regards
> > > > Andrey
> > > >
> > > > > Date: Tue, 6 Oct 2015 18:12:48 +0300
> > > > > Subject: Re: Cluster group affinity
> > > > > From: agura@gridgain.com
> > > > > To: dev@ignite.apache.org
> > > > >
> > > > > Andrey,
> > > > >
> > > > >
> > > > > > 1) I'm expected to return an instance of the internal class
> > > > > > AffinityTopologyVersion.
> > > > >
> > > > >
> > > > > If you are talking about
> > AffinityContextFunction.currentTopologyVersion
> > > > > method then for now this method is nowhere uses. But it make sense
to
> > > > > return non null value in order to avoid problems in the future.
> > > > >
> > > > > 2) the consequences of returning null from
> > > > > > AffinityFunctionContext.previousAssignment and
> > > > > > AffinityFunctionContext.discoveryEvent methods (because I can't
> > > > provide any
> > > > > > meaningful implementation for them) are not clear.
> > > > > >
> > > > >
> > > > > Both methods declared as @Nullable, so affinity function developer
> > should
> > > > > correctly handle this cases. In Ignite only FairAffinityFunction
uses
> > > > these
> > > > > methods. FairAffinityFunction tries to obtain left node Id from
> > event of
> > > > > EventType.EVT_NODE_LEFT or EventType.EVT_NODE_FAILED type. It needs
> > to
> > > > > exclude this node assignment from previous assignments. So if your
> > > > cluster
> > > > > group lost node you can return EVT_NODE_LEFT discovery event with
Id
> > of
> > > > > lost node from discoveryEvent method and assignments for previous
> > cluster
> > > > > group state from previousAssignment method.
> > > > >
> > > > > RendezvousAffinityFunction uses only currentTopologySnapshot() and
> > > > > backups() methods of AffinityFunctionContext interface.
> > > > >
> > > > >
> > > > > On Tue, Oct 6, 2015 at 5:07 PM, Andrey Kornev <
> > andrewkornev@hotmail.com>
> > > > > wrote:
> > > > >
> > > > > > Andrey, thanks!
> > > > > >
> > > > > > But a "properly formed AffinityFunctionContext" is the problem:
> > > > > > 1) I'm expected to return an instance of the internal class
> > > > > > AffinityTopologyVersion.
> > > > > > 2) the consequences of returning null from
> > > > > > AffinityFunctionContext.previousAssignment and
> > > > > > AffinityFunctionContext.discoveryEvent methods (because I can't
> > > > provide any
> > > > > > meaningful implementation for them) are not clear.
> > > > > >
> > > > > > Please advise.
> > > > > >
> > > > > > Thanks
> > > > > > Andrey
> > > > > >
> > > > > > > Date: Tue, 6 Oct 2015 16:43:10 +0300
> > > > > > > Subject: Re: Cluster group affinity
> > > > > > > From: agura@gridgain.com
> > > > > > > To: dev@ignite.apache.org
> > > > > > >
> > > > > > > Andrey,
> > > > > > >
> > > > > > > See AffinityFunction.assignPartitions method. It returns
> > assignment
> > > > list
> > > > > > as
> > > > > > > List<List<ClusterNode>> where index of element
in returned list
> > > > > > corresponds
> > > > > > > to partition number. Assignment for each partition represented
as
> > > > list of
> > > > > > > nodes where primary node is always the first. So you can
use
> > existing
> > > > > > > affinity functions for you case just passing properly formed
> > > > > > > AffinityFunctionContext to assignPartitions method.
> > > > > > >
> > > > > > > On Tue, Oct 6, 2015 at 4:25 PM, Andrey Kornev <
> > > > andrewkornev@hotmail.com>
> > > > > > > wrote:
> > > > > > >
> > > > > > > > Dmitriy,
> > > > > > > >
> > > > > > > > The affinity function only maps a key to a partition
id and it
> > > > doesn't
> > > > > > > > seem to provide a way to map the partition id to a
cluster
> > node. So
> > > > > > I'm a
> > > > > > > > little bit confused right now.
> > > > > > > >
> > > > > > > > Could you please clarify?
> > > > > > > >
> > > > > > > > Thanks a lot
> > > > > > > > Andrey
> > > > > > > >
> > > > > > > > > From: dsetrakyan@apache.org
> > > > > > > > > Date: Mon, 5 Oct 2015 09:53:25 -0700
> > > > > > > > > Subject: Re: Cluster group affinity
> > > > > > > > > To: dev@ignite.apache.org
> > > > > > > > >
> > > > > > > > > On Mon, Oct 5, 2015 at 2:28 AM, Andrey Kornev
<
> > > > > > andrewkornev@hotmail.com>
> > > > > > > > > wrote:
> > > > > > > > >
> > > > > > > > > > Hello,
> > > > > > > > > >
> > > > > > > > > > I have a user-defined cluster group and
I'd like to be
> > able to
> > > > > > > > > > consistently pick the same node in the group
for a given
> > key.
> > > > > > > > Essentially,
> > > > > > > > > > what I want is a cluster group affinity
that is not
> > associated
> > > > > > with any
> > > > > > > > > > cache. How can I do it?
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > Andrey, perhaps you could just take our affinity
function and
> > > > use it
> > > > > > > > > directly, no?
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > Thanks
> > > > > > > > > > Andrey
> > > > > > > > > >
> > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > --
> > > > > > > Andrey Gura
> > > > > > > GridGain Systems, Inc.
> > > > > > > www.gridgain.com
> > > > > >
> > > > > >
> > > > >
> > > > >
> > > > >
> > > > > --
> > > > > Andrey Gura
> > > > > GridGain Systems, Inc.
> > > > > www.gridgain.com
> > > >
> > > >
> >
> >
 		 	   		  
Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message