On Fri, Oct 5, 2018 at 3:59 AM Chamikara Jayalath <chamik...@google.com> wrote:
> > On Thu, Oct 4, 2018 at 9:39 AM Ahmet Altay <al...@google.com> wrote: > >> I agree that LTS releases require more thought. Thank you for raising >> these questions. What other open questions do we have related LTS releases? >> >> One way to do this would be to add them to a particular tracking list >> (e.g. 2.9.0 blocking list) that way we would be ready for an LTS release >> ahead of time. >> >> Related to dependencies, I agree with Thomas. If we block on waiting for >> dependencies, we may end up taking a long time before making any LTS >> release. And the reality of Beam releases right now is that there are no >> supported releases today that and in the long term that might hurt user >> trust. In my opinion, we need to fix that sooner rather than later. >> > > Agree on the idea of focussing on stability instead of feature set when it > comes to LTS releases. Based on the previous discussion on this [1] looks > like the intended audience is enterprise customers that would like to > maintain a stable environment and that usually have a long testing cycle > before deploying a new version. Seems like, rushing in features or > dependency upgrades for an LTS release will be counterintuitive for such > use cases. > > On the other hand, I think a part of Ismaël point was that all major > features of Beam (Spark Runner, Flink Runner, Kafka IO, etc ..) should be > supportable for the duration of the time Beam LTS release is supported (1 > year based on [1]). If a major dependency is expected to become unsupported > within an year, it makes sense to upgrade that before the LTS release. I > suggest we do this at least one release before the planned LTS release (so > 2.9.0 if we want to make 2.10.0 a LTS release) so that we have all major > dependency/feature updates battle tested before committing to them. > Dependencies are a significant concern. What happens if one of our dependencies has no version they're going to support for a year hence? It could happen that there's never a time when all our dependencies will be supported for at least a year as of any given date. And rushing to get the "latest" version of a dependency in is often contrary to the goals of an LTS. (Better to have those depend on the battle-hardened versions.) The long-term goal is probably to avoid pinning specific versions as much as possible. E.g. if we worked with Flink 1.5 or 1.6 (whichever the user provided), and considered backporting support for 1.7 when it came out, we'd be in much better shape here. How feasible this is depends on how backwards-compatible the dependency is. On the other hand, forcing a user to upgrade a (user-visible, which in particular includes runners) dependency seems contrary to the goals of an LTS. Even if that dependency becomes unsupported.
