> After 1.16, only patches are accepted for 1.2.0-1.15.
I feel like this is a misunderstanding that both you and Danny ran into.
What I meant in the original proposal is that the last 2 _major_
/connector /versions are supported, with the latest receiving additional
features.
(Provided that the previous major version still works against a
currently supported Flink version!)
There will never be patch releases for a minor version if a newer minor
version exists.
IOW, the minor/patch releases within a major version do not form a tree
(like in Flink), but a line.
1.0.0 -> 1.0.1 -> 1.1.0 -> 1.2.0 -> ...
NOT
1.0.0 -> 1.0.1 -> 1.0.2
|-----> 1.1.0 -> 1.1.1
If we actually follow semantic versioning then it's just not necessary
to publish a patch for a previous version.
So if 2.x exists, then (the latest) 2.x gets features and patches, and
the latest 1.x gets patches.
I hope that clears things up.
On 20/09/2022 14:00, Jing Ge wrote:
Hi,
Thanks for starting this discussion. It is an interesting one and yeah, it
is a tough topic. It seems like a centralized release version schema
control for decentralized connector development ;-)
In general, I like this idea, not because it is a good one but because
there might be no better one(That's life!). The solution gives users an
easy life with the price of extra effort on the developer's part. But it is
a chicken and egg situation, i.e. developer friendly vs. user friendly. If
it is hard for developers to move forward, it will also be difficult for
users to get a new release, even if the version schema is user friendly.
I'd like to raise some questions/concerns to make sure we are on the same
page.
@Chesnay
c1) Imagine we have 2.0.0 for 1.15:
- 2.0.0-1.14 (patches)
- 2.0.0-1.15 (feature and patches)
===> new major release targeting 1.16 and we need to change code for new API
- 2.0.0-1.14 (no support)
- 2.0.0-1.15 (patches)
- 2.0.1-1.15 (new patches)
- 2.1.0-1.16 (feature and patches)
There is no more 2.1.0-1.15 because only the latest version is receiving
new features.
b1) Even if in some special cases that we need to break the rule, we should
avoid confusing users:
===> new major release targeting 1.16 and we need to change code for new API
- 2.0.0-1.14 (no support)
- 2.0.0-1.15 (patches)
- 2.1.0-1.16 (feature and patches)
===> now we want to break the rule to add features to the penultimate
version
- 2.0.0-1.14 (no support)
- 2.0.0-1.15 (patches)
- 2.2.0-1.15 (patches, new features) // 2.1.0-1.15 vs. 2.2.0-1.15,
have to choose 2.2.0-1.15 to avoid conflict
- 2.1.0-1.16 (feature and patches)
we have two options: 2.1.0-1.15 vs. 2.2.0-1.15, both will confuse users:
- Using 2.1.0-1.15 will conflict with the existing 2.1.0-1.16. The
connector version of "2.1.0-1.16" is actually 2.1.0 which means it has the
same code as 2.1.0-1.15 but in this case, it contains upgraded code.
- Using 2.2.0-1.15 will skip 2.1.0-1.15. Actually, it needs to skip all
occupied minor-1.16 versions, heads-up release manager!
c2) Allow me using your example:
===> new major release targeting 1.16
- 1.2.0-1.14 (no support; unsupported Flink version)
- 1.2.0-1.15 (patches; supported until either 3.0 of 1.17, whichever
happens first)
- 2.0.0-1.15 (feature and patches)
- 2.0.0-1.16 (feature and patches)
I didn't understand the part of "2.0.0-1.15 (features and patches)". After
1.16, only patches are accepted for 1.2.0-1.15.
It should be clearly defined how to bump up the connector's version number
for the new Flink version. If the connector major number would always bump
up, it would make less sense to use the Flink version as postfix. With the
same example, it should be:
===> new major release targeting 1.16
- 1.2.0-1.14 (no support; unsupported Flink version)
- 1.2.0-1.15 (patches; supported until either 3.0 of 1.17, whichever
happens first)
- 1.2.1-1.15 (new patches)
- 1.3.0-1.16 (feature and patches)
- 1.4.0-1.16 (feature and patches, new features)
- 2.0.0-1.16 (feature and patches, major upgrade of connector itself)
or
- 1.2.0-1.14 (patches)
- 1.2.0-1.15 (feature and patches)
- 2.0.0 -1.15 (feature and patches, major upgrade of connector itself)
===> new major release targeting 1.16
- 1.2.0-1.14 (no support; unsupported Flink version)
- 2.0.0-1.15 (patches)
- 2.0.1-1.15 (new patches)
- 2.1.0-1.16 (feature and patches)
- 2.2.0-1.16 (feature and patches, new features)
i.e. commonly, there should be no connector major version change when using
the Flink version postfix as the version schema. Special cases(rarely
happens) are obviously allowed.
Best regards,
Jing
On Tue, Sep 20, 2022 at 10:57 AM Martijn Visser<martijnvis...@apache.org>
wrote:
Hi all,
This is a tough topic, I also had to write things down a couple of times.
To summarize and add my thoughts:
a) I think everyone is agreeing that "Only the last 2 versions of a
connector are supported per supported Flink version, with only the latest
version receiving new features". In the current situation, that means that
Flink 1.14 and Flink 1.15 would be supported for connectors. This results
in a maximum of 4 supported connector versions.
b1) In an ideal world, I would have liked Flink's APIs that are used by
connectors to be versioned (that's why there's now a Sink V1 and a Sink
V2). However, we're not there yet.
b2) With regards to the remark of using @Interal APIs, one thing that we
agreed to in previous discussions is that connectors shouldn't need to rely
on @Interal APIs so that the connector surface also stabilizes.
b3) In the end, I think what matters the most is the user's perception on
versioning. So the first thing to establish would be the versioning for
connectors itself. So you would indeed have a <major.minor.patch> scheme.
Next is the compatibility of that scheme with a version of Flink. I do like
Chesnay's approach for using the Scala suffixes idea. So you would have
<major.minor.patch connector>_<major.minor Flink version>. In the currently
externalized Elasticsearch connector, we would end up with 3.0.0_1.14 and
3.0.0_1.15 as first released versions. If a new Flink version would be
released that doesn't require code changes to the connector, the released
version would be 3.0.0_1.16. That means that there have been no connector
code changes (no patches, no new features) when comparing this across
different Flink versions.
b4) Now using the example that Chesnay provided (yet slightly modified to
match it with the Elasticsearch example I've used above), there exists an
Elasticsearch connector 3.0.0_1.15. Now in Flink 1.16, there's a new API
that we want to use, which is a test util. It would result in version
3.1.0_1.16 for the new Flink version. Like Chesnay said, for the sake of
argument, at the same time we also had some pending changes for the 1.15
connector (let's say exclusive to 1.15; some workaround for a bug or smth),
so we would also end up with 3.1.0-1.15. I agree with Danny that we should
avoid this situation: the perception of the user would be that there's no
divergence between the 3.1.0 version, except the compatible Flink version.
I really am wondering how often we will run in that situation. From what
I've seen so far with connectors is that bug fixes always end up in both
the release branch and the master branch. The only exceptions are test
stabilities or documentation fixes, but if we only resolve these, they
wouldn't need to be released. If such a special occasion would occur, I
would be inclined to go for a hotfix approach, where you would end up with
3.0.0.1_1.15.
c) Branch wise, I think we should end up with <major.minor.patch
connector>_<major.minor Flink version>. So again the Elasticsearch example,
at this moment there would be 3.0.0_1.14 and 3.0.0_1.15 branches.
Best regards,
Martijn
