Valya, And this is the problem! ) We have another implementation which minimizes movements - rendezvous. Fair affinity should not bother about partition migration. This is not a feature, but a bug in implementation of FairAffinityFunction. Let's fix that and forget about topology versions.
ср, 16 авг. 2017 г. в 3:23, Valentin Kulichenko < valentin.kuliche...@gmail.com>: > Vladimir, > > I would let other guys confirm, but I believe the reason is that if it > recalculates distribution every time from scratch, it would trigger too > much redundant data movement during rebalancing. Fair function not only > tries to provide best possible distribution, but also minimizes this data > movement, and for this it uses previous distribution. > > -Val > > On Tue, Aug 15, 2017 at 1:12 PM, Vladimir Ozerov <voze...@gridgain.com> > wrote: > > > I do not like the idea as it would make it very hard to reason about > > whether your SQL will fail or not. Let's looks at the problem from the > > different angle. I have this question for years - why in the world *fair* > > affinity function, whose only ultimate goal is to provide equal partition > > distribution, depends on it's own previous state? Can we re-design in a > way > > that it depends only on partition count and current topology state? > > > > On Thu, Aug 10, 2017 at 12:16 AM, Valentin Kulichenko < > > valentin.kuliche...@gmail.com> wrote: > > > > > As far as I know, all logical caches with the same affinity function > and > > > node filter will end up in the same group. If that's the case, I like > the > > > idea. This is exactly what I was looking for. > > > > > > -Val > > > > > > On Wed, Aug 9, 2017 at 8:18 AM, Evgenii Zhuravlev < > > > e.zhuravlev...@gmail.com> > > > wrote: > > > > > > > Dmitriy, > > > > > > > > Yes, you're right. Moreover, it looks like a good practice to combine > > > > caches that will be used for collocated JOINs in one group since it > > > reduces > > > > overall overhead. > > > > > > > > I think it's not a problem to add this restriction to the SQL JOIN > > level > > > if > > > > we will decide to use this solution. > > > > > > > > Evgenii > > > > > > > > > > > > > > > > > > > > 2017-08-09 17:07 GMT+03:00 Dmitriy Setrakyan <dsetrak...@apache.org > >: > > > > > > > > > On Wed, Aug 9, 2017 at 6:28 AM, ezhuravl <e.zhuravlev...@gmail.com > > > > > > wrote: > > > > > > > > > > > Folks, > > > > > > > > > > > > I've started working on a https://issues.apache.org/ > > > > > > jira/browse/IGNITE-5836 > > > > > > ticket and found that the recently added feature with cacheGroups > > > doing > > > > > > pretty much the same that was described in this issue. CacheGroup > > > > > > guarantees > > > > > > that all caches within a group have same assignments since they > > > share a > > > > > > single underlying 'physical' cache. > > > > > > > > > > > > > > > > > I think we can return FairAffinityFunction and add information to > > its > > > > > > Javadoc that all caches with same AffinityFunction and NodeFilter > > > > should > > > > > be > > > > > > combined in cache group to avoid a problem with inconsistent > > previous > > > > > > assignments. > > > > > > > > > > > > What do you guys think? > > > > > > > > > > > > > > > > Are you suggesting that we can only reuse the same > > FairAffinityFunction > > > > > across the logical caches within the same group? This would mean > that > > > > > caches from the different groups cannot participate in JOINs or > > > > collocated > > > > > compute. > > > > > > > > > > I think I like the idea, however, we need to make sure that we > > enforce > > > > this > > > > > restriction, at least at the SQL JOIN level. > > > > > > > > > > Alexey G, Val, would be nice to hear your thoughts on this. > > > > > > > > > > > > > > > > > > > > > > Evgenii > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > > View this message in context: http://apache-ignite- > > > > > > developers.2346864.n4.nabble.com/Resurrect-FairAffinityFunction- > > > > > > tp19987p20669.html > > > > > > Sent from the Apache Ignite Developers mailing list archive at > > > > > Nabble.com. > > > > > > > > > > > > > > > > > > > > >
