Re: OutOfMemoryError while streaming

[breischl <br...@gmail.com>]

The keys are hashed so in theory they should be distributed relatively
evenly. I traced the logic to do the hashing once and it seemed ok, but it's
pretty complicated so I won't claim to know it that well. We use UUIDs and
it does seem to distribute pretty much evenly. 

Are you sure all the nodes have properly joined the cluster? ie, that you
have a single cluster with 8 nodes, instead of a bunch of 1-2 node clusters.
Depending on how your discovery is set up, that can be tricky and
potentially have timing bugs. Although that may be more due to us having a
custom DiscoverySPI that is a bit finicky. 

For #2, I would like to know the answer too. :) We are using
OptimizedMarshaller, but that's due to some problems with serializing some
of our complex objects which would not apply to your case. You may want to
just test it. 



--
Sent from: http://apache-ignite-users.70518.x6.nabble.com/

Re: OutOfMemoryError while streaming

[Denis Mekhanikov <dm...@gmail.com>]

How many entries did you put into a cache?
How do you check, that nodes have entries?
Did you configure a node filter for the cache?

Rendezvous affinity function may give unfair distribution on small
datasets,
but it gets better, when number of entries grows.
So, if you've only put 5-10 entries, then you may get unfair data
distribution.

Denis

вс, 1 июл. 2018 г. в 18:33, breischl <br...@gmail.com>:

> The keys are hashed so in theory they should be distributed relatively
> evenly. I traced the logic to do the hashing once and it seemed ok, but
> it's
> pretty complicated so I won't claim to know it that well. We use UUIDs and
> it does seem to distribute pretty much evenly.
>
> Are you sure all the nodes have properly joined the cluster? ie, that you
> have a single cluster with 8 nodes, instead of a bunch of 1-2 node
> clusters.
> Depending on how your discovery is set up, that can be tricky and
> potentially have timing bugs. Although that may be more due to us having a
> custom DiscoverySPI that is a bit finicky.
>
> For #2, I would like to know the answer too. :) We are using
> OptimizedMarshaller, but that's due to some problems with serializing some
> of our complex objects which would not apply to your case. You may want to
> just test it.
>
>
>
> --
> Sent from: http://apache-ignite-users.70518.x6.nabble.com/
>