Yes. Agree with Ash that when we have "sdk" that will be all-but-guaranteed
(or at least intended to be) backwards compatible and no DB access from
workers, or DAG processor, then yes we can change the rules for Airflow 3
going forward, but Airflow 2 support question remains.

Basically it boils down to "How long are we going to support Airflow 2 in
providers released while Airflow 3 is already out?". 12 months
after Airflow 3.0 is released is if we use the current rule.

No strong opinions of that but that looks good.

J



On Thu, Aug 15, 2024 at 1:57 PM Ash Berlin-Taylor <a...@apache.org> wrote:

> Going forward with Airflow 3 I also think we should re-consider what we
> version/depend upon from providers.
>
> If we move all providers (and likley also utility functions!) out of
> airflow core then the thing that a given provider needs to depend upon is
> the shared library/provider versions and the "Task SDK” that we are
> introducing with AIP-72, and the version of Airflow core doesn’t matter
> anymore.
>
> This doesn’t take away the need to think about how long we continue to
> support Airflow 2 in our providers, but that going forward I think we need
> to subtly shift how we think about this. The first “key use case” we
> mention in AIP 72 is this:
> >
> > 1. Ensure that this interface reduces the interaction between the code
> running within the Task and the rest of Airflow. This is to address
> unintended ripple effects from core Airflow changes which has caused
> numerous Airflow upgrade issues, because Task (i.e. DAG) code relied on
> Core Airflow abstractions. This has been a common problem pointed out by
> numerous Airflow users including early adopters. This proposal would enable
> Airflow users to upgrade Airflow system components (Scheduler, et al),
> without impacting DAG user code.
>
> I don’t yet have a clear proposal of what it should be, but I do want to
> surface this.
>
> -ash
>
> > On 15 Aug 2024, at 12:05, Amogh Desai <amoghdesai....@gmail.com> wrote:
> >
> > Nice points to ponder upon, Jarek. Both topics are essential for ensuring
> > that users have a clear path forwars.
> >
> >> We currently guarantee that the minimum Airflow version supported by a
> > provider is the release date of the next minor version, plus 12 months.
> > You’re asking if this should be adjusted for Airflow 3, correct?
> > When we say adjust, what are we looking at? Are we thinking of extending
> > the 12 month period, to lets say 18 months
> > or something else?
> >
> > Few thoughts that come to my mind when I read this email:
> > - A potential adjustment would be to extend the window to 18 months post
> > the first Airflow 3 release,
> > allowing users more time to transition without feeling
> > rushed. Communicating this timeline clearly
> > to the community will be key here. We should *emphasize* that while they
> > extra time for the transition,
> > they will eventually need to migrate to Airflow 3 to access the newer
> > features and bug fixes.
> > - We can introduce a policy (and document it well, maybe in
> >
> https://github.com/apache/airflow/blob/main/PROVIDERS.rst#upgrading-minimum-supported-version-of-airflow
> > )
> > where providers are guaranteed to work with all Airflow versions released
> > within 12 months (or 18)
> > of the provider release. After that, users *should* expect that older
> > providers may not be fully compatible
> > with newer Airflow versions, and any issues arising from this won't be
> > treated as bugs.
> >
> > We should also document these rules clearly to provide users with the
> > context and rationale behind these decisions.
> >
> > Thanks & Regards,
> > Amogh Desai
> >
> >
> > On Wed, Aug 14, 2024 at 3:37 PM Jarek Potiuk <ja...@potiuk.com> wrote:
> >
> >>> We currently guarantee that the minimum Airflow version supported by a
> >> provider is the release date of the next minor version, plus 12 months.
> >> You’re asking if this should be adjusted for Airflow 3, correct?
> >>
> >> Yes
> >>
> >>> Are you asking for guarantees the other way around, so Airflow version
> ->
> >> guaranteed provider package min version support?
> >>
> >> Yes.
> >>
> >> On Wed, Aug 14, 2024 at 12:04 PM Bas Harenslak
> <b...@astronomer.io.invalid>
> >> wrote:
> >>
> >>> Sorry I’m a bit lost in the long text. Is my understanding of the 2
> >>> questions correct?
> >>>
> >>> We currently guarantee that the minimum Airflow version supported by a
> >>> provider is the release date of the next minor version, plus 12 months.
> >>> You’re asking if this should be adjusted for Airflow 3, correct?
> >>> Are you asking for guarantees the other way around, so Airflow version
> ->
> >>> guaranteed provider package min version support?
> >>>
> >>> Bas
> >>>
> >>>> On 14 Aug 2024, at 11:22, Jarek Potiuk <ja...@potiuk.com> wrote:
> >>>>
> >>>> Hello everyone,
> >>>>
> >>>> I have two important topics to discuss regarding Provider <> Airflow
> >>>> version support rules.
> >>>>
> >>>> *1) What is the "min Airflow version" support while we are
> >> transitioning
> >>> to
> >>>> Airflow 3? *
> >>>>
> >>>> So far we had the [1] rule:
> >>>>
> >>>>> The default support timespan for the minimum version of Airflow
> (there
> >>>> could be justified exceptions) is that we increase the minimum Airflow
> >>>> version to the next MINOR release, when 12 months passed since the
> >> first
> >>>> release for the MINOR version of Airflow.
> >>>>
> >>>>> For example this means that by default we upgrade the minimum version
> >> of
> >>>> Airflow supported by providers to 2.8.0 in the first Provider's
> release
> >>>> after 18th of December 2024. 18th of December 2023 is the date when
> the
> >>>> first PATCHLEVEL of 2.8 (2.8.0) has been released.
> >>>>
> >>>> The next cutoff time (2.9.0 + 12 months) will be in April 2025, And
> the
> >>>> next one - 2.10.0 - assuming we release 2.10.0 in August - for min
> >>> airflow
> >>>> version = 2.10.0 will be in August 2025. If we have "bridge" 2.11.0
> >>> release
> >>>> (very likely) - say in December, then min_version = 2.11.0 will be in
> >>>> December 2025.
> >>>>
> >>>> Assuming that we release Airflow 3.0.0 in March 2025, if we follow
> this
> >>>> rule, we will be able to drop Airflow 2 support in providers in March
> >>> 2026
> >>>> - so we give about a year to Airflow 2 users to migrate to Airflow 3
> -
> >>> if
> >>>> they will want new features and bug-fixes from new providers. Also it
> >>> means
> >>>> that we have 1.5 year of supporting both Airflow 2 and Airflow 3 in
> >>>> providers (that for example includes the old Templated Field
> >>>> back-compatibility support assuming that we will want to go ahead with
> >>> the
> >>>> change)
> >>>>
> >>>> That provides (as was the original idea of this rolling min-version
> >>>> support) additional incentive for users to migrate (want new features
> >>> from
> >>>> providers - migrate to latest Airflow). With Airflow 3 that will be a
> >>> more
> >>>> difficult decision to migrate, so I wonder if this rule should be
> >>> somewhat
> >>>> adapted.
> >>>>
> >>>> *2) Should we introduce a rule for a "rolling" min version Provider's
> >>>> support in Airflow 3?*
> >>>>
> >>>> In Airflow 2 we never had a rule describing "how old" providers are
> >> still
> >>>> supported in newer versions of Airflow. For example we never said you
> >>> need
> >>>> to have at least "2.10.0" version of a google provider. This was never
> >> a
> >>>> huge problem, because vast majority of users were upgrading providers
> >>> with
> >>>> Airflow and only occasionally downgraded providers or kept them with
> >> old
> >>>> versions because of some breaking changes.That had never caused a big
> >>>> problem as far as I know and remember - and we can safely assume that
> >>>> "really old" providers might  not work with "new" airflow even if we
> >> had
> >>> no
> >>>> formal rule for it.
> >>>>
> >>>> However that required some back-compatibility code sometimes - like
> >> this
> >>>> code removal that started breaking our main after removing old
> >> "imports"
> >>>> [2]. This means that some very old SB providers will not work with the
> >>> new
> >>>> common.sql provider when released (and they also will not work with
> >>> Airflow
> >>>> 3). I think it will never cause a "real" serious issue - but would be
> >>> great
> >>>> to formalize it to be able to set expectations of our users and have
> >> some
> >>>> ways to respond to potential issues they might raise.
> >>>>
> >>>> Should we formalize it somehow? For example add a rule that providers
> >>> will
> >>>> be guaranteed to support Airflow versions released up to 12 months
> >> after
> >>>> the provider was released? We could potentially add tests to check it
> >> in
> >>>> the future (though it would be rather complex and time consuming). But
> >> I
> >>>> think we should start with the rule first.
> >>>>
> >>>> Example to illustrate it (hypothetical):
> >>>>
> >>>> * Google provider 11.0.4 is released in December 2024 - just before
> >>>> Airflow 2.11.0
> >>>> * Airflow 3.0 - 3.4 are released between December 2024 and December
> >> 2025
> >>> -
> >>>> all of them support google provider 2.11.0 (and if not - this is a bug
> >>> that
> >>>> needs to be fixed)
> >>>> * Airflow 3.5 is released in January 2026 - and if someone wants to
> >>> install
> >>>> Airflow 3.5 and use Google provider 11.0.4 or below - they might, or
> >>> might
> >>>> not work. But if they don't we do not treat it as a bug and we
> >> recommend
> >>>> the users to upgrade to a newer Google provider version.
> >>>>
> >>>> I think we should not make "intentional" changes to break those
> >>> providers,
> >>>> unless we have good reasons, and it's mostly about some weird edge
> >> cases,
> >>>> where people have new Airflow and old providers, but I think it's
> worth
> >>>> introducing and documenting some rules about it.
> >>>>
> >>>>
> >>>> Let me know what you think on both accounts.
> >>>>
> >>>> J.
> >>>>
> >>>> [1]
> >>>>
> >>>
> >>
> https://github.com/apache/airflow/blob/main/PROVIDERS.rst#upgrading-minimum-supported-version-of-airflow
> >>>> [2] https://github.com/apache/airflow/pull/41461
> >>>
> >>>
> >>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscr...@airflow.apache.org
> For additional commands, e-mail: dev-h...@airflow.apache.org
>
>

Reply via email to