On Wed, Oct 8, 2025 at 2:53 PM Josh McKenzie <[email protected]> wrote:
> Thanks for the discussion - it looks like we agree on the problem and the > trade‑offs involved. > > Maintaining an extra branch would add toil: we’d have to merge bug fixes, > run CI (including upgrade tests), and extend the already lengthy upstream > path. A simpler alternative is to relax our backport restrictions on GA > branches. > Yes, this would mean CEP-37 would go to 5.0.x, for example. I do not have a problem with that in general, I am just not sure if we can "abuse" a patch release for this kind of addition. It would really have to be "non-disruptive as much as possible". This would be in line with what Jeff / Jeremiah / myself were suggesting as well. We would not need to introduce a new branch, upgrade paths would be tested as they are now, bug fixes would be added too ... Looking forward to the opinions of other people. > > What about the following idea: allow community-approved feature backports > to *latest GA* (with a brief window allowing backports to 2 GA branches) > and tier releases as follows: > > *When a new release is cut:* > > - New release / Latest GA (6.0): stabilizing (backports accepted) > - Middle GA (5.0): backports accepted > - Oldest GA (4.1): stable > > *When the new release stabilizes:* > > - Latest GA (6.0): backports accepted > - Middle GA (5.0): stable > - Oldest GA (4.1): stable > > This approach gives users a clear choice - stable, backport‑enabled, or a > temporary stabilizing branch - without adding CI overhead. > > On Wed, Oct 8, 2025, at 5:42 AM, Štefan Miklošovič wrote: > > Hi Dinesh, > > thanks for reassuring that this branch would be really just for crucial > functionality where benefits justify the backport. I would be more willing > to entertain that idea but I still think that once people see 5.1 is up > they will want to support their case and there will be a lot of pressure to > backport this and that. Not all CEPs / features will make it. We need to be > very selective. There will be a lot of "massaging" around what can go in > and what not and the expectations would need to be set from the very > beginning and followed. > > However, I am not still quite sure how that would work in general, reading > Josh email here: > > "The branch would selectively accept non‑disruptive improvements that meet > criteria we define together." > > Once people are on 5.1, they will want to have all the bug fixes in > there as well. So instead of merging from 4.0 to 4.1, 5.0 and trunk, we > will be doing 4.0. 4.1, 5.0, 5.1 and trunk? > > If the answer is yes, then we will have just another branch we need to > fully maintain. > > If the answer is no, as in we will skip 5.1 on merges from 4.0 to trunk, > then I think this will be met with disappointment and questions as to why > we are not patching 5.1 as well. > > Basically we go all in and maintain 5.1 with all the patches from lower > branches or we just maintain and backport important features but then ... > who is going to use it like that - without receiving bug fixes. > > > > On Wed, Oct 8, 2025 at 9:46 AM Dinesh Joshi <[email protected]> wrote: > > Stefan, Sam – your concerns are absolutely valid and have come up in > various discussions. > > Here's the reality though – many large operators of Cassandra are > maintaining backports of various features. The proposal here is to try and > allow these contributors to maintain them in the community instead of > internally. This is a limited time pilot to see if this model could work. > > When we "open the flood gates" then the existence of a backporting branch > will be the justification of anything they want to see there because they > do not want to upgrade. > > > Stefan — nobody is talking about “opening the floodgates” here. The > expectation is that small, self contained features could be back ported on > a case by case basis. Let’s engage on the criteria that makes sense. > > On the subject of avoiding backports and using it as a tool to “force” > people to upgrade, I’d like to point out that if upgrades were easier we > would not be having this discussion. The simple fact is that upgrades are > not easy and they are riskier than maintaining backports hence we see this > pattern. > > If the community gets together and makes upgrades easier we will likely > not have a need for backports. > > My suggestion is to engage with “how” this pilot would look like to shape > it. It is a limited time experiment that might benefit the community. A > number of contributors have shown interest so ideally we should be open to > trying it out. > > > > On Wed, Oct 8, 2025 at 12:12 AM Sam Tunnicliffe <[email protected]> wrote: > > I second Štefan's concerns here. The proposal reduces the incentive to > upgrade or even test trunk, meaning that the things users want to avoid > (features etc, but also just refactorings/re-implementations) because they > are as-yet "untrusted" or "unqualified" remain that way for longer. This > feels pretty antithetical to the direction we've been aiming to travel in, > toward more regular release cycles. > > > > On 8 Oct 2025, at 06:41, Štefan Miklošovič <[email protected]> > wrote: > > > > This is indeed an interesting idea but please let me share my point of > view and somehow different opinion on that. > > > > I share the questions with Jeff and Jeremiah a lot. I see it similarly > and they got the point. > > > > Before 5.0 was out, we had quite a situation where we officially had to > take care of 3.0, 3.11, 4.0 and 4.1 at the same time. If a bug was found, > we had to patch 5 branches at once (trunk as well). That meant 5 CI jobs. > The patching was an endeavour spanning multiple days, realistically. Once > 5.0 got out, we officially discontinued 3.0 and 3.11. But what I have been > experiencing was that this information about not supporting 3.0 / 3.11 was > spreading very slowly among people / customers and I / we had to repeatedly > explain to everybody that yes, 3.0 and 3.11 and done. What are they? Done? > Yes, done. 3.0 and 3.11 are finished. Finished you say? That means no > patches? Yes, no patches. Aha right ... For real? ... you got it. People > had to internalize that it is just not going to happen. > > > > When we "open the flood gates" then the existence of a backporting > branch will be the justification of anything they want to see there because > they do not want to upgrade. Instead of us working towards a more smooth > upgrade we are burying ourselves with older stuff. That slows adoption of > new majors a lot. People will not be forced to, there will be way less > incentive to do that when all the important goodies are backported anyway. > > > > I see that "the backports would be non-disruptive but potentially higher > risk". I do not completely understand what this means in practice. Let's > say CEP-37. Is that disruptive or not? What's the definition of that? To > me, correct me if I am wrong, is that something is disruptive if I just can > not turn it off even if I do not want to use it. Does one _have to_ use > CEP-37 when it is backported? No. They can just turn it off. So what is > exactly the risk of introducing it to e.g. 5.0.x ? > > > > Also, how are upgrades done? People are going to upgrade from 5.0.x to > 5.1 and then it will be possible to upgrade to 6.0 from 5.1? This would > need us to make the pipelines, incorporate this new path into upgrade tests > and so on ... a lot of work. > > > > I think that the current policy - "only bug fixes to older branches" > might be relaxed a bit instead and leverage already existing upgrade paths > and infrastructure to test it all instead of creating brand new branches we > need to take care of. > > > > Regards > > > > On Mon, Oct 6, 2025 at 6:04 PM Josh McKenzie <[email protected]> > wrote: > > Many large‑scale Cassandra users have had to maintain private feature > back-port forks (e.g., CEP‑37, compaction optimization, etc) for years on > older branches. That duplication adds risk and pulls time away from > upstream contributions which came up as a pain point in discussion at CoC > this year. > > > > The proposal we came up with: an official, community‑maintained backport > branch (e.g. cassandra‑5.1) built on the current GA release that we pilot > for a year and then decide if we want to make it official. The branch would > selectively accept non‑disruptive improvements that meet criteria we define > together. There’s a lot of OSS prior art here (Lucene, httpd, Hadoop, > Kafka, Linux kernel, etc). > > > > Benefits include reduced duplicated effort, a safer middle ground > between trunk and frozen GA releases, faster delivery of vetted features, > and community energy going to this branch instead of duplicated on private > forks. > > > > If you’re interested in helping curate or maintain this branch - or have > thoughts on the idea - please reply and voice your thoughts. > > > > ~Josh > > >
