I think it's a good point. Culturally we have been willing to break extension APIs for relatively small benefits. But we have generally been unwilling to make breaking changes on the operations side quite so liberally. Also, most cluster operators don't have their own custom extensions, in my experience. So it does make sense to differentiate them. I'm not sure how it makes sense to differentiate them, though. It could be done through the version number (only increment the major version for operations breaking changes) or it could be done through an "upgrading" guide in the documentation (increment the major version for operations or extension breaking changes, but, have a guide that tells people which versions have operations breaking changes to aid in upgrades).
Coming back to the question in the subject of your mail: IMO, for "graduation" out of 0.x, we should talk as a community about what that means to us. It is a milestone that on the one hand, doesn't mean much, but on the other hand, can be deeply symbolic. Some things that it has meant to other projects: 1) Production readiness. Obviously Druid is well past this. If this is what dropping the 0. means, then we should do it immediately. 2) Belief that the APIs have become relatively stable. Like you said, the extension APIs don't seem particularly close to stable, but maybe that's okay. However, the pace of breaking changes on the operations and query side for non-experimental features has been relatively calm for the past couple of years, so if we focus on that then we can make a case here. 3) Completeness of vision. This one is the most interesting to me. I suspect that different people in the community have different visions for Druid. It is also the kind of project that may never truly be complete in vision (in principle, the platform could become a competitive data warehouse, search engine, etc, …). For what it's worth, my vision of Druid for the next year at least involves robust stream ingestion being a first class ingestion method (Kafka / Kinesis indexing service style) and SQL being a first class query language. These are both, today, still experimental features. So are lookups. All of these 3 features, from what I can see, are quite popular amongst Druid users despite being experimental. For a 'completeness of vision' based 1.0 I would want to lift all of those out of experimental status and, for SQL in particular, to have its functionality rounded out a bit more (to support the native query features it doesn't currently support, like multi-value dimensions, datasketches, etc). 4) Marketing / timing. Like, doing a 1.0 around the time we graduate from the Incubator. Not sure how much this really matters, but projects do it sometimes. Another question is, how often do we intend to rev the version? At the rate we're going, we rev 2-3 major versions a year. Would we intend to keep that up, or slow it down by making more of an effort to avoid breaking changes? On Thu, Dec 20, 2018 at 2:17 PM Roman Leventov <leventov...@gmail.com> wrote: > It may also make sense to distinguish "operations" breaking changes from > API breaking changes. Operations breaking changes establish the minimum > cadence of Druid cluster upgrades, that allow rolling Druid versions back > and forward. I. e. it's related to segment format, the format of the data > kept in ZooKeeper and the SQL database, or events such as stopping support > of ZooKeeper for certain things (e. g. forcing using of HTTP > announcements). So Druid cluster operators cannot update Druid from version > X to version Z skipping the version Y, if both Y and Z have some operations > breaking changes. (Any such changes should support rollback options at > least until the next version with operations breaking changes.) > > API breaking changes are just changes in Druid extensions APIs. Druid > cluster operators could skip any number of releases with such breaking > changes, as long as their extension's code is updated for the latest > version of API. > > On Thu, 20 Dec 2018 at 20:03, Roman Leventov <leven...@apache.org> wrote: > > > It doesn't seem to me that Druid API is going to stabilize in the near > > future (if ever), because there are so many extension points and > something > > is broken in every release. On the other hand, Druid is not Hadoop or > > Spark, which have applications API. Druid API for extensions, not > > applications. It is used by people who are closer to Druid development > and > > fixing their extensions is routine. > > > > With that, I think it make sense to drop "0." from the Druid version and > > call it Druid 14, Druid 15, etc. > > >
