IIRC @PublicEvolving had no /real/ guarantees, from the 1.0.0 announcement:
/"Flink 1.0.0 introduces interface stability annotations for API classes
and methods. Interfaces defined as //|@Public|//are guaranteed to remain
stable across all releases of the 1.x series. The
//|@PublicEvolving|//annotation marks API features that may be subject
to change in future versions."/
1) We don't have a process for promoting things to @Public because we
have actually never done that. Based on our current rules, it is an
addition to the public API (technically even more so than most FLIPs),
and hence should go through a FLIP and vote.
Ideally we evaluate existing @PublicEvolving API's for promotion on each
release.
2) Yes, I think this is an excellent idea; as Ufuk said users should not
be worried about upgrading to the next minor version. Not sure how easy
this is to implement; but it basically means either modifying or adding
new new japicmp execution when forking the release-X.Y branches.
On 13/05/2020 15:19, Till Rohrmann wrote:
A small addendum: The project currently enforces only binary compatibility
for classes which are annotated with @Public. The StreamingFileSink is
annotated with @PublicEvolving.
I am not sure whether the community properly defined what kind of
stability/compatibility guarantees we provide for @PublicEvolving classes.
We can derive two follow-up discussions from this:
1) Should the StreamingFileSink be made @Public? Should
other @PublicEvolving classes be promoted? What is the process here?
2) Should we enforce binary compatibility of @PublicEvolving classes
between bug fix releases and allow breaking changes between minor/major
releases?
Cheers,
Till
On Wed, May 13, 2020 at 2:30 PM Ufuk Celebi <u...@apache.org> wrote:
Thanks for the analysis Till.
1/ I think breaking binary compatibility is acceptable between minor
releases (1.10 -> 1.11) as you suggested since the API is marked
as @PublicEvolving.
2/ I'm quite torn about how to proceed with the 1.10.x patch release, but
slightly leaning towards the "pragmatic" solution of keeping the change and
adding a big warning to the 1.10.1 release announcement*. What do others
think? If we do want to revert the change and follow up with a 1.10.2
release we should do it _before_ we announce 1.10.1 publicly. Doing it
after 1.10.1 has been announced would only cause more friction for users.
– Ufuk
*I hope we don't lose user trust with this. I really would like users to be
able to upgrade to the latest patch release without hesitation.
On Wed, May 13, 2020 at 11:45 AM Till Rohrmann <trohrm...@apache.org>
wrote:
Thanks for reporting this issue Seth. This is indeed a big problem.
I looked into the problem and it seems we have the following situation:
# 1.9.x --> 1.10.0
There is an API breaking change between 1.9.x and 1.10 due FLINK-13864
because it introduced another generic parameter. I expected only few
people
will be affected by this because one would have to store the builder in a
variable.
1.9.x is binary compatible with 1.10.0.
# 1.10.0 -> 1.10.1
There is no API breaking change between 1.10.0 and 1.10.1.
I changed the return type of the StreamingFileSink.forRowFormat and
.forBulkFormat to a subtype of StreamingFileSink.BulkFormatBuilder in
FLINK-16684. This causes the java.lang.NoSuchMethodError and the binary
incompatibility between 1.10.0 and 1.10.1. This is should not happen for
bug fix releases.
W/o FLINK-16684, the StreamingFileSink builders cannot be used with
Flink's
Scala API.
# Options
I think we should keep the fix for 1.11.0 and add a release note that we
are violating binary compatibility between 1.10 and 1.11.
Now the question is what to do with the 1.10.x branch:
a) We can revert the change to re-establish binary compatibility between
1.9.x, 1.10.0 and 1.10.2 but not between 1.10.1 and 1.10.2. This would
also
imply that we cannot use the StreamingFileSink builders with the Scala
API.
b) Keep the change and add a release note that our users need to
recompile
their jobs. This would allow Flink 1.10.x with x >= 1 users to use
StreamingFileSink builders with the Scala API.
I am not entirely sure which need weighs more: binary compatibility
between
bug fix releases or a working API (small subset of it). Another aspect to
consider is how many people will migrate from 1.10.0 to 1.10.1 compared
to
1.y to 1.10.1 with y <= 9. If the former is very small then one might
make
a case for keeping the change.
What do the others think?
https://issues.apache.org/jira/browse/FLINK-13864
https://issues.apache.org/jira/browse/FLINK-16684
Cheers,
Till
On Tue, May 12, 2020 at 7:26 PM Thomas Weise <t...@apache.org> wrote:
We also noticed that and had to make an adjustment downstream.
It would be good to mention this in the release notes (if that's not
already the case).
Thomas
On Tue, May 12, 2020 at 10:06 AM Seth Wiesman <sjwies...@gmail.com>
wrote:
Hi Everyone,
I realize I'm about 7 hours late but I just realized there is a
breaking
API change in 1.10.1. FLINK-16684 changes the API of the streaming
file
sink in a binary incompatible way. Since the release has been
approved
I'm
not sure the correct course of action but I wanted to bring this to
the
communities attention.
Seth
https://issues.apache.org/jira/browse/FLINK-16684
