Any config object that needs to be Identifiable and Serializable should be identified as a CacheElement, but we don't have to go through all of the objects and identify them at once. When we are refactoring each individual commands and recognize the need, we will identify the object to be a CacheElement.
On Wed, Apr 25, 2018 at 3:41 PM, Patrick Rhomberg <prhomb...@pivotal.io> wrote: > Sorry, I forgot to step away from the In The Weeds where I am. > > o.a.g.cache.configuration.CacheElement is the experimental interface meant > to identify an extension's configuration classes for consumption with the > new cluster configuration interface. Extension developers can then use the > configuration service to disseminate these configuration objects via member > groups, consistent with other configuration elements (e.g., indices and > gateways). See [1] for the API's proposal. > > We also have been using this interface internally on some of our own > configuration objects, for instance the JDBC's ConnectorService. This > allowed us to use some utilities we had made for the easy retrieval of the > configuration objects. > > The lack of documentation on this interface is my core concern. The > initial intent was to label those extension's configuration classes, but > beyond that, its use is ambiguous. Even within that purview, it's not > clear whether an inner class should be required to implement CacheElement. > > Jinmei and I had a thread recently on a pull request of hers [2] suggesting > perhaps "anything you want to be searchable in the config" should have the > interface applied to it. That remains more subjective than I'd personally > like, but it would at least be documentable. > > So to return to my initial concerns: our current use of the interface is > inconsistent, and with it being an external API, we should properly define > the interface's scope for both external and internal use. I propose four > possible scopes in the initial message. > > @Udo, I haven't added a section on the linked proposal, but the page could > do with a larger update in any case. I'll start that straight away. > > [1] Public API for Cluster Configuration Proposal: > https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=75975896 > [2] PR #1853: https://github.com/apache/geode/pull/1853 > > On Wed, Apr 25, 2018 at 12:56 PM, Galen O'Sullivan <gosulli...@pivotal.io> > wrote: > > > It looks like there are two classes called CacheElement in the codebase; > > am I correct in thinking that you're referring to > > org.apache.geode.cache.configuration.CacheElement? > > > > This class doesn't have a Javadoc, so it's a little hard as an outsider > to > > understand exactly what it is or how it's used. Can you explain the > basics > > of how it's used? > > > > Thanks, > > > > Galen > > > > > > > > On 4/19/18 5:03 PM, Patrick Rhomberg wrote: > > > >> Hello all! > >> > >> We introduced the CacheElement interface as part of our experimental > >> API > >> to update the cluster configuration. I'd like to solidify and document > >> our > >> intent for the interface and the extent to which it is expected to > apply. > >> In its current form, the CacheElement interface extends Serlializable > >> and > >> Identifiable<String>. Serialization is required for communication > between > >> members, and Identifiable is useful during lookup, modification, and > >> removal of an existing configuration object. > >> At our current iteration, it is not entirely clear which of our own > >> configuration objects should or should not implement CacheElement. I > >> think > >> one of following interpretations may be best, but don't know which would > >> be > >> most natural. And, of course, if you have another that I've overlooked, > >> I'd love to hear it. > >> > >> Classes that implement CacheElement are... > >> > >> A) ... only custom configuration elements provided by extension > >> developers. This declaration will appear at a top-level class > declaration > >> that is directly consumed by the CacheConfig or RegionConfig. > >> > >> B) ... custom configuration elements, as well as any configuration > object > >> that we wish to be searchable within the configuration. > >> > >> C) ... custom configuration elements, as well as any configuration > object > >> appearing at the same level in the configuration hierarchy. That is, > >> every > >> getter of both CacheConfig and RegionConfig should return either a > >> CacheElement, or a List<CacheElement>. > >> > >> D) ... any object, excepting those in java.lang.*, that appears anywhere > >> in > >> the cluster configuration's hierarchy, including all inner classes. For > >> example, ConnectorService.RegionMapping implements CacheElement. > >> > >> > >> Each comes with its own potential pitfalls. > >> Option (A) seems restrictive and sparse. > >> Option (B) is subjective. > >> Many classes do not have a `name' or `id' field, making option (D) > >> difficult. > >> I like option (C), although we have already moved beyond that > >> specification in our current iteration, with > >> ConnectorService.RegionMapping > >> implementing CacheElement. > >> > >> Thoughts? > >> > >> Imagination is Change. > >> ~Patrick Rhomberg > >> > >> > > > -- Cheers Jinmei
