I disagree on freezing the API before 2.0. For one, it seems to indicate a shift in policy over what a major release means if we’ve previously said that API changes are allowed in 1.x releases. More important, if we’re looking at potentially disruptive changes in the API for 2.0 (for example, I believe John Vines said he didn’t expect the transition to 2.0 to be seamless), it means integrators may take a long time to adopt it. Many might even avoid 2.0 to wait for 2.1 to wait for things to stabilize after the “big changes” in 2.0. My project, for example, might not be able to upgrade to the 2.x line for a very long time for just that reason plus some other factors beyond our control. I believe “forcing” people to decide on 2.0 by abandoning the addition of features in the 1.x line is just the wrong way to go.
Sean makes the argument that porting backwards is just as difficult as porting forwards in our code. That is certainly true, but I don’t see how that’s a reason to avoid all API changes. Take the hypothetical case he provides of an integrator that gets a regulatory requirement to use an earlier version. Why is it more likely that this requirement would dictate they go from 1.7 to 1.6 instead of from 1.7 to 1.3? What if the integrator already used new features that are currently in 1.7? It seems to me there’s no way to prevent that potential pain in this hypothetical situation. We have a very real situation where preventing API additions to 1.7 can cause problems for our integrators. In fact it already has in the case of ACCUMULO-3176 since we wanted the feature it provides for my project. I agree that there will be temptation to squeeze out just one more release before 2.0, but I’d rather it be a little harder on us instead of making it easier on ourselves at the expense of alienating users. Sure these users could backport features they need, but if they’re only going to be coming from the 2.0 line with a very different API, backporting may be a lot of work. > On Dec 3, 2014, at 9:48 AM, Sean Busbey <bus...@cloudera.com> wrote: > > -1 also, ATM. I'd like to see us freeze APIs between now and the 2.0 > release. > > Downstream users have to plan when they invest effort in migrating Accumulo > versions. We've already signaled that 2.0 will be the start of a new API > with long-lived compatibility promises. (We should keep signaling this.) > That makes it a promising place to make a jump (in some cases, from 1.4 I'm > sure). > > I would like to avoid, however possible, leaning those users towards > ignoring releases between now and 2.0. For those who are back on 1.4 or 1.5 > we can't really do too much. For those on 1.6 we can make it so there is > relatively little risk in moving forward. > > API additions matter here because when a system integrator makes an > application on top of Accumulo they often start at the latest version they > can find. Later, they may have a client with a regulatory requirement to > use an earlier version. Porting backwards is just as hard as porting > forwards in our code base. > > I'd also like to see the "no removing of deprecated" language strengthened > to remove the exception for things deprecated prior to 1.7. > > Yes, this will severely constrain what we can do prior to 2.0. But I think > doing otherwise will just encourage us to keep squeezing in "just one more" > major pre-2.0 release to get some additional client facing feature out the > door. > > If we have some downstream users with different compatibility needs and > with particular operational needs for features that are delayed to 2.0 > because of this decision, it should be straight forward for them to > backport the things they need and run their own packaging. Plenty of folks > who don't need the legal indemnification that the ASF provides do this for > a wide variety of projects. > > > On Tue, Dec 2, 2014 at 2:07 PM, John Vines <vi...@apache.org> wrote: > >> -1 I do not like the idea of committing to 1.7.0-1.9.9... API additions for >> the 2.0 API. We have already come to the consensus that 2.0 will break the >> 1.x API which provides a lot of breathing room and freedom from old >> decisions. This causes this issue to come roaring back and an even larger >> amount of scrutiny to be required for all 1.7.0-1.9.9... API changes. I >> would go so far as to say an undefinable amount of scrutiny since we still >> don't have solid foundation of a 2.0 API. We cannot judge API items for how >> well they belong in an API that does not exist yet. >> >> Tangential- I would like to see a clause about all current API items will >> not be removed (still could be deprecated) until 2.0.0, as I feel this may >> ease some concerns about API alteration in 1.7+. >> >> On Tue, Dec 2, 2014 at 3:01 PM, Christopher <ctubb...@apache.org> wrote: >> >>> Following the conversation on the [VOTE] thread for ACCUMULO-3176, it >> seems >>> we require an explicit API guidelines at least for 1.7.0 and later until >>> 2.0.0. >>> >>> I hereby propose we adopt the following guidelines for future releases >> (if >>> we produce any such releases) until 2.0.0: >>> >>> API additions are permitted in "major" 1.x releases (1.7, 1.8, 1.9, 1.10, >>> etc.). >>> API should be forwards and backwards compatible within a 1.x release (no >>> new additions to the API in a "bugfix" release; e.g. 1.7.1). >>> New API in 1.7.0 and later 1.x releases will not be removed in 2.0 >> (though >>> they may be deprecated in 2.0 and subject to removal in 3.0). >>> Existing API in 1.7.0 will be preserved through 2.0, and should only be >>> subject to removal if it was already deprecated prior to 1.7.0 (though >> they >>> may be deprecated in 2.0 and subject to removal in 3.0). >>> >>> The purpose of these guidelines are to ensure the ability to add >> additional >>> functionality and evolve API naturally, while minimizing API disruptions >> to >>> the user base, in the interim before 2.0.0 when we can formally adopt an >>> API/versioning policy. >>> >>> Exceptions to these guidelines should be subject to a majority vote, on a >>> case-by-case basis. >>> >>> Because these relate to release planning, this vote will be subject to >>> majority vote, in accordance with our bylaws pertaining to release >> planning >>> and voting, and will be open for 3 days, concluding at 2000 on 5 Dec 2014 >>> UTC. >>> >>> -- >>> Christopher L Tubbs II >>> http://gravatar.com/ctubbsii >>> >> > > > > -- > Sean
