On Fri, 6 Nov 2015 07:02:58 -0700, Phil Steitz wrote:
On 11/6/15 2:04 AM, luc wrote:
Le 2015-11-06 02:34, Gilles a écrit :
On Thu, 5 Nov 2015 15:41:57 -0700, Phil Steitz wrote:
On 11/5/15 1:58 PM, Luc Maisonobe wrote:
Le 05/11/2015 12:25, Gilles a écrit :
Hello.
On Wed, 04 Nov 2015 10:13:00 +0100, luc wrote:
Hi all,
I would like to release 3.6 in the upcoming weeks.
There have been a bunch of bug fixes and a few evolutions
that are
important to me.
s/3.6/4.0/
And the statement is still true.
[...]
Of course, once 3.6 is out the MATH_3_X branch will remain
alive and
we could also release other versions later on in the 3.x
series.
Should we worry that the next major release will be endlessly
delayed?
I think we are really quite far from releasing 4.0 as in many
packages
we did not even start refactoring API. Optimization is clearly
the most
well known example, but there are also other things waiting in
the pipe
for geometry and ode.
Is there any specific target for 4.0?
Yes, at least having changed public API.
Could we judge how far we are from releasing a major release with
the
same arguments which you used for 3.6 (many additions, bug fixes,
things someone would like to use, etc.)?
4.0 does not need to be perfect. [3.0 was supposed to be perfect
;-).]
4.1 will be! Or 4.2, or 5.0...
4.x for x > 1 will have exactly the same problems as 4.0 API-wise
since
once 4.0 is out the API will be fixed.
Let's not forget that we agreed to work towards 4.0, and that the
3.x
branch was an afterthought.
I agree and was slightly reluctant to continue on 3.X at that
time. Deciding
to still keep this branch was indeed a good idea. I properly
address the problem
that we did not find time to work on 4.0 as we wanted.
Since we now effectively maintain two versions of CM, it makes
sense
to allow people to benefit from the extra work.
Yes, but our own overzealous rule about compatibility (I can take
the
blame here, I am guilty for this) induces that we change API only
when
major versions are published. We do have the opportunity to do
this for
4.0, lets use it at least and not again postpone needed changes.
Our 3.X
API sucks in many places and we know it.
My viewpoint is that we can have releases from both branches,
making
it clear what is old/deprecated/broken in 3.x and what is still
missing in 4.y.
If it were only missing features, that can be added, I would
agree. However
some of the changes are really modifications of existing interfaces.
I agree on this. One thing I forgot to mention above is I think
we
may have a few places in 3.x optimization where every way to do
something is deprecated. I suggest we take a careful look and
undeprecate where necessary to make 3.6 usable without warnings.
I
may be wrong since I don't use that code myself; but I think its
worth taking a careful look and considering removal of some
deprecations.
I'm against this. And is why I started the
sure-to-be-controversial
discussion on 4.0.
I also don't really like the idea of undeprecating these things.
We agreed that things were (relatively) broken, and we agreed on a
way forward (fluent API, refactoring of the "optim" classes); yet
things don't move, because there is no urge to fix them since new
features can be back-ported indefinitely to the 3.x series.
No, things don't move because I didn't find time. I am really,
really busy doing lots of different stuff. I am also really, really
aware this API should be improved and fluent API is still a way I
would
like to explore for this. And no, I am not sure this will work and
4.x will see the end of these problems.
Undeprecating what we agreed should be deprecated would only
reinforce that feeling, and certainly won't attract attention that
we need help to make progress.
[And, in addition, 3.x is tied to old Java5 (known tune)...]
In summary, I think that new features should only go to the master
branch, while only bug-fixes would be back-ported to MATH_3_X.
Thus everybody can have the best of both, while reducing the
amount of work for the developers. Continuing in this way, and
we'll soon have to also "forward-port" bugs reported against the
3.x series. :-/
Hey, I already do that! The following one-liner is my new
favorite:
git diff -p MATH_3_X~1 MATH_3_X | sed 's,math3,math4,g' | patch
-p1
Yes, it is cumbersome.
I am as busy and over-subscribed as anyone and agree it is a little
cumbersome; but I am happy to do it so users can have something
stable to work with.
Do you offer to do it for all the commits *I* might push to the
"master" branch?
That would be nice; thank you!
But is only part of the request: that new features exist in 3.x
should not frustrate those of use who would like to rely on an
official release of compatibility-breaking features.
Realize that they too are very busy and much
as we may like to ponder over the best way to keep refactoring our
API, they just want to solve practical problems (why [math] was
created in the first place) and once they have invested the time
necessary to figure out how to get something done, they are not
going to be thrilled about the prospect of having to invest more
time learning how we have decided to change something.
It would be nice to consider what I wrote rather than summarily
imply that, should my viewpoint come to pass, we'd be forcing
anyone to learn new ways to do the same things.
A. People are happy with 3.5; they don't need to touch anything
(i.e. they keep using 3.5).
B. People fear that they might be hit by a bug: they switch to 3.6.
C. People want to use new features: "math4" must be an *additional*
dependency. For refactored functionalities, people are _free_
to update their codes, or not. Whatever they choose, their
code will depend on the "current" state of CM.
IIUC, your scenario is
D. People ask that we do more work so that they can benefit from
new features without adding one line in their dependency file.
I consider this an unreasonable expectation.
Then the following may indeed seem problematic:
E. People are hit by a bug which we deem to complicated to fix
without breaking compatibility.
But in fact, it's their call: either they help us by providing
a patch that we can include in 3.x, or they have to accept to
share the effort (we fix in 4.0, they upgrade to it).
I know we have some API bugs that are so bad that they keep us from
being able to deliver functionality. I know those need to be
fixed. But I also know a library without a stable API is not
suitable for widespread use and like it or not, we have become
widely used.
So, in short, you've just stated that
1. API must be changed,
2. API must not be changed.
That general problem of CM exists because it is inherent to an
open-ended development (in terms of features) with no specific
targets other than "[someone needs this code now] to solve a
practical problem".
It's neither good nor bad. In this situation, my position is
that we should be ready, even willing, to make disruptive changes
in new releases. Knowing that it won't hurt anyone but the most
lazy fellows.
On the other hand, it is perfectly reasonable to decide that
CM must become a legacy project, bent on accepting utterly
contorted ways to fix bugs without ever touching on its API
again.
Personally, I clearly don't want this (both as a user and
as a developer).
The research that I did in preparation for my
apacachecon talk this year showed that we 3.x is very widely
deployed.
Would be interesting to know the above "A", "B", "C", "D", "E"
user types repartition.
I think we should continue to support this series and I
would like to do that, both in terms of bug fixes
Good IMO (as said in previous post).
and new features.
Bad IMO (ditto). And unnecessary burden.
Bad for CM. Bad for developers, bad for users, bad for attracting
contributors.
I also think we should change our deprecation strategy to only
deprecate where there is a recommended alternative in the released
code.
It's not because you *hide* what needs improving that the problem
disappears.
IMHO, it is good that users see that we think about our code,
that we have identified problems. And that we lack the resources
to fix all of them.
[Perhaps deprecation warnings will get us more visibility than
the "caveat" warnings in the release notes!]
I ask that we think about users actually using the library.
I ask that we have consideration for the whole food-chain.
Who likes to have deprecation warnings when they write new code
using something?
Those who cannot bear it will find how to redirect the messages to
/dev/null. ;-)
Who likes to struggle to get something working
using a difficult-to-work-with API and then have to struggle again
to figure out how to use a maybe-slightly-better API to get new
features?
Not many people I guess.
As I suggest above, shall we define what kind of project CM is
("legacy" or not) and agree on what it means in terms of true
expectations, for users and developers?
I'm responsible of the half-baked versions of the successively
and iteratively revised APIs of "optim".
The history is somewhere in the ML archive.
Think about your own experience using other people's
APIs.
I think that we are overly cautious about compatibility with
outdated code, tools and practice.
I am also quite convinced that nobody can come up with a
case that we performed gratuitous changes (assuming that
CM is not "legacy").
Please lets think about our users, people.
If people want to complain, they are welcome in this forum
where we help them help us (!) find a way to solve their
"practical problems".
If you are happy about how the CM community has not grown
to match the growth of the codebase, then we agree to
disagree on what CM should be, and the community will
probably shrink as a consequence of not taking *all* of
us into account.
Best regards,
Gilles
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org
For additional commands, e-mail: dev-h...@commons.apache.org
