[FFmpeg-devel] [PATCH 00/26] Major library version bump

[James Almer]

On 1/19/2023 4:26 AM, Anton Khirnov wrote:
> Quoting James Almer (2023-01-18 22:23:43)
>> On 1/18/2023 4:28 PM, Anton Khirnov wrote:
>>> Quoting James Almer (2023-01-16 14:38:14)
>>>> It's been a while since the last bump, so it's time to do some cleaning and
>>>> remove deprecated APIs. This will also give us an "Open ABI season" in which we
>>>> can do breaking changes (like changing public struct offsets, public enum
>>>> values, adding fields to structs that have their size tied to the ABI, etc) for
>>>> a few weeks.
>>>
>>> Last time this open season lasted something like half a year and only
>>> ended when I arbitrarily said it did.
>>>
>>> So I'd suggest to decide right now how long will the instability period
>>> last (6 weeks should be enough for everybody) and write the end date at
>>> the top of doc/APIchanges.
>>>
>>> Another thing I'm not entirely happy about is versioning during the bump
>>> and instability. While the remove-then-bump approach does make bisection
>>> easier, it also creates commits that lie about their ABI version.
>>
>> Does it really matter? All the patches will be pushed at the same time,
>> meaning one git fetch will give you a stable state pre bump and the next
>> will be right after it.
>> I think it's a bit farfetched to expect someone to pick a random commit
>> in the middle of the bump and try to use the resulting compiled
>> libraries with some program that was linked to some earlier version
>> libraries.
> 
> I agree that it's probably not a big practical problem, but it is ugly
> and goes against our claims of git master being stable.
> 
>>> I wonder if we couldn't come up with a better soltion. One thing that
>>> comes to mind is setting the major version to 0 until the instability
>>> period ends.
>>
>> This could have several undesired effects, mainly for users looking at
>> that define and not really expecting such value (There are several
>> projects supporting more than one ffmpeg release and "MAJOR <= xx"
>> checks are commonplace).
> 
> IMO users who don't expect such a value shouldn't be linking against
> unstable API/ABI anyway. We could also set the major version to
> something really big, like 999. We'll have to change deprecation macros,
> but that should be straightforward.
> 
>> Also, if we are going to code the instability period in some form into
>> the codebase, might as well make it so it starts with the first removal
>> commit, or immediately before it, so what you described above is no
>> longer a concern.
> 
> I'd rather say the two concerns merge into one, but it's not going away.
> There's currently very little user indication that API/ABI are unstable
> for several months.

How about making minor == 0 mean unstable? Some projects like GCC do it 
like this, for example. Said version would not guarantee anything at all 
and should not be linked against. Then once the period ends after the 
major bump, it's bumped to 1, and the usual "new API, minor bump" rule 
kicks in.
This also means that any API addition that takes place during the 
unstable period doesn't get it's own version, and they will all strictly 
speaking be introduced by minor 1.

As for your concern about first removing then bumping meaning we're 
lying about the ABI, the first commit in the set could maybe rollback 
minor to 0, then apply the removals, and finally the major bump in the 
last commit. It will that way be considered unstable for the whole thing.