Val, Dima, We already have per-cache key configuration: CacheConfiguration.keyConfiguration.
Personally I vote for complete restriction of any runtime changes. Everything should be specified in advance *before* cache is started. @AffinityKeyMapped should be deprecated, as well as "affinityKey" should be removed from binary metadata. Predefined affinity configuration would solve a lot of issues: 1) We will be able to use it for SQL engine optimizations and UX improvements. E.g. SqlFieldsQuery.collocated flag will not be needed anymore, we will be able to determine whether distributed join is needed or not, etc.. 2) We will get consistent and predictable behavior. For example, currently if I put an object with this annotation to cache, it will be used. But if I put the same object through BinaryObjectBuilder, it will not be used ever. On Wed, Sep 13, 2017 at 3:00 AM, Valentin Kulichenko < valentin.kuliche...@gmail.com> wrote: > Dmitry, > > IMO, it's actually pretty typical for data grid use cases where affinity > key is usually provided as part of key itself, i.e. after cache creation. > In vast majority of cases I've seen, this is done via very popular > @AffinityKeyMapped annotation. > > My only point is that the annotation can't always be used due to > restrictions on application level. For example, an application I was > recently working with has a platform that uses Ignite internally and > provides its own API. Therefore it can't expose @AffinityKeyMapped to its > users, but instead provides another annotation (and probably some other > custom mechanisms). It currently uses deprecated AffinityKeyMapper to > support this and there is no alternative. Just getting rid of it would mean > changing and complicating the platform API. > > -Val > > On Mon, Sep 11, 2017 at 6:25 PM Dmitriy Setrakyan <dsetrak...@apache.org> > wrote: > > > On Mon, Sep 11, 2017 at 4:47 PM, Valentin Kulichenko < > > valentin.kuliche...@gmail.com> wrote: > > > > > Even if CacheKeyConfiguration is part of CacheConfiguration, the > affinity > > > key field name can be provided only on cache startup. In many cases > this > > > name can be resolved only based on the actual key instance, e.g. during > > > first put. Per my understanding, this already works with annotation, I > > just > > > propose more flexible solution for rare cases when annotation can't be > > > used. Basically, the logic we currently have would become the default > > > implementation of the resolver. > > > > > > > Val, the use case seems very strange to me. How can you not know the > > affinity key field in advance? Can you provide an example from the field? > > >
