On Fri, 6 Nov 2015 16:53:00 -0800, Gary Gregory wrote:
On Fri, Nov 6, 2015 at 4:42 PM, Gilles <gil...@harfang.homelinux.org>
wrote:
On Fri, 6 Nov 2015 17:02:01 -0700, Phil Steitz wrote:
On 11/6/15 4:46 PM, Gary Gregory wrote:
On Fri, Nov 6, 2015 at 3:01 PM, Phil Steitz
<phil.ste...@gmail.com>
wrote:
On 11/6/15 2:51 PM, Gary Gregory wrote:
On Fri, 6 Nov 2015 09:17:18 -0700, Phil Steitz wrote:
Here is an idea that might break our deadlock re backward
compatibility, versioning and RERO:
Agree that odd numbered versions have stable APIs -
basically
adhere
to Commons rules - no breaks within 3.0, 3.1, ..., 3.x...
or 5.0,
5.1... but even-numbered lines can include breaks -
...
This sounds awfully complicated for my puny human brain.
How, exactly? Seems pretty simple to me. The even-numbered
release
lines may have compat breaks; but the odd-numbered do not.
It's bad enough that I have to remember how each FOSS project
treats
versions numbers, but having an exception within a Commons
component is
even worse. This is a non-starter for me.
Do you have any better suggestions? The problem we are trying to
solve is we can't RERO while sticking to the normal compat rules
without turning major versions all the time, which forces users
to
repackage all the time and us to support more versions
concurrently
than we have bandwidth to do.
I do not see how a different version scheme will determine how
many
branches the community supports.
If we just keep one 4.x branch that keeps cutting (possibly
incompatible) releases, that is just one line, one branch. If we
have to cut 4.1, 4.2, 4.3 as 4, 5, 6 instead and we don't allow any
compat breaks, we end up having to maintain and release 4.0.1,
5.0.1, 6.0.1 instead of just 4.3.1, for example, or we just strand
the 4, 5 users in terms of bug fixes as we move on to 6.
Breaking BC without a package and coord change is a no-go.
We have done this before and we will probably do it again - and
more
if we have to don't separate out an unstable line.
You have to
think about this jar as a dependency that can be deeply nested in
a
software stack. Commons components are such creatures. I
unfortunately
run
into this more than I'd like: Big FOSS project A depends on B
which
depends
on C. Then I want to integrate with Project X which depends on Y
which
depends on different versions of B and C. Welcome to jar hell if B
and C
are not compatible. If B and C follow the rule of break-BC -> new
package/coords, then all is well.
The mitigation here is that we would not expect the even-numbered
releases to be deployed widely.
Or we can prevent any potential JAR hell by changing the top-level
package name for every release in the unstable series:
4.0 -> org.apache.commons.math4u0
4.1 -> org.apache.commons.math4u1
This would also offer the possibility to write a code using
classes from several unstable releases (e.g. to compare their
performance).
Pardon my daftness, but if you change package/coord names, what are
you
gaining compared to the usual major version changes?
Objectively, nothing...
But IIUC Phil's paragraph, above, using a dual numbering would
highlight the "unstable" nature of some releases: the whole 4.x
series would be known (in advance) to be unstable, while if we
release a usual 4.0, some user might assume it to be "stable",
only to see it abandoned a few months later when 5.0 is released.
Moreover, at each major number bump, we'd have to advertize (in
words) whether that release is stable or not, while with the dual
numbering the warning would be carried by number in the package
name.
Maybe it's only a psychological trick. But if it can help to
release more often, it would have led to a real improvement! :)
Gilles
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org
For additional commands, e-mail: dev-h...@commons.apache.org
