Re: [PATCH 00/13] remote-hg: general updates

[Max Horn]

On 03.04.2013, at 03:31, Felipe Contreras wrote:

> On Tue, Apr 2, 2013 at 4:23 PM, Max Horn <m...@quendi.de> wrote:
>>
>> On 02.04.2013, at 22:09, John Keeping wrote:
>>
>>> On Tue, Apr 02, 2013 at 01:02:49PM -0600, Felipe Contreras wrote:
>>>> Here is the next round of patches for remote-hg, some which have been
>>>> contributed through github.
>>>>
>>>> Fortunately it seems to be working for the most part, but there are some
>>>> considerable issues while pushing branches and tags.
>>>
>>> How does this compare to the current state of gitifyhg[1]? That's built
>>> on top of this git-remote-hg script but seems to have been more actively
>>> developed recently.
>
> I only learned about it recently, I've looked at the history and to me
> it seems rather chaotic, and a lot of the code was simply copied from
> git-remote-hg without comment.

gitifyhg was scrapped and completely restarted from scratch at some point. Based largely on your git-remote-hg code. A bit more on its history can be read here:
http://archlinux.me/dusty/2013/01/06/gitifyhg-rewritten/


>
>> * added many new test cases, sadly still including some xfails. Several of these (both passing and xfailing) also apply to remote-hg (i.e. the issue is also present in contrib's remote-hg)
>
> I ran these test-cases with remote-hg, and the same test-cases pass. I
> only had to do minor modifications, most of the failures came from
> subtle differences such as different strategies to sanitize authors,
> and which branch to pick for HEAD.

Yeah, that's because what I wrote was based on the situation before your recent patch series. Glad to see that git/contrib's remote-hg is improving!


>
>> * improved handling of hg user names (remote-hg is not able to deal with some pathological cases, failing to import commits). Sadly, mercurial allows arbitrary strings as usernames, git doesn't...
>
> I wouldn't call it improved. In some cases the remote-hg result is
> better, in others gitifyhg is,

I'd love to learn about cases where remote-hg's result is better in your opinion, so that I can see if the mapping in gitifyhg could be improved for those cases.

That said, this part is really quite subjective I guess. In the end, since Mercurial names can be *anything*, one can never get a "perfect" mapping. Luckily, for most real repositories out there, user names will be quite sane and remote-hg and gitifyhg will produce identical results. (Although some hg repos out there have some really weird stuff going on. Yuck.)

> but there's only a single case where
> the author name becomes a significant problem. It's a trivial fix.

Excellent, looking forward to it then.

>
>> * failed pushes to hg are cleanly rolled back (using mq.strip() from the mq extension), instead of resulting in inconsistent internal state. This is quite important in real life, and has bitten me several times with remote-hg (and was the initial reason why I switched to gitifyhg). A typical way to reproduce this is to push to a remote repository that has commits not yet in my local clone.
>
> This is not an issue in remote-hg any more since now we force the
> push. It's not nice, but there's no other way to push multiple
> bookmarks (aka git branches) to the same branch (aka commit label).

Uhh, what? The push is forced? That sounds to me like a much, much bigger issue with remote-hg than anything else ever was, from my point of view. That seems tobe akin to making "--force" default to "git push", and I don't think anybody here would consider that a good idea.

> I doubt these inconsistent states can happen any more, but if they do,

Seriously? This is triggered quite frequently in real life. And it will very likely cause somebody to mess up a hg repository they work on. As long is this in, using remote-hg is a total no-go for me. Just consider the following scenario:
* user A clones a hg repository into a git repository
* user A commits some commits in the git clone
* meanwhile, user B pushes changes to the hg repository
* user A tries to push his changes to the hg remote

The last step causes this result in gitifyhg (similar to what one gets when the remote is a git repos):

! [rejected] master -> master (non-fast-forward)
error: failed to push some refs to 'gitifyhg::URL'
hint: Updates were rejected because the tip of your current branch is behind
hint: its remote counterpart. Merge the remote changes (e.g. 'git pull')
hint: before pushing again.
hint: See the 'Note about fast-forwards' in 'git push --help' for details.


With remote-hg, you just force push the change, creating a new head in the remote repo. So, yeah, failed pushes which mess up the internal state don't happen anymore. But I rather have those than potentially mess up the upstream repository like that.


> the plan in remote-hg is to simply ignore those revisions, and only
> push the ones that have git refs. I have the code for that, but I'll
> not be pushing it to git.git for the time being.

I am not quite sure what you mean here, but I'll just wait for your code and hope it'll explain itself.

>
>> * git notes are used to associate to each git commit the sha1 of the corresponding hg commit, to help users figure out that mapping
>
> This is a minor feature. I've had the code for this for quite some
> time, but for the moment I think there are higher priorities.

Well, I guess it depends on how you use this. In my daily work with hg repositories, that's a quite important feature.

Actually, I wonder: Are you actually using remote-hg yourself for day-to-day development? Or for any other regular activities? You seem to have quite different usage patterns from me and other people involved with gitifyhg.

>
>> * internally, the marks are using the hg sha1s instead of the hg rev ids. The latter are not necessarily invariant, and using the sha1s makes it much easier to recover from semi-broken states.
>
> I doubt this makes any difference (except for more wasted space).

I disagree. Especially since remote-hg does treat local hg repositories differently than remote ones, by not cloning them but rather working with them directly. As soon as somebody uses "hg strip" or "hg graft", etc. on that, your mark files will contain incorrect data.

With a remote hg repository, that is of course not as much of an issue, since you have a local hg clone stored inside of .git/. In theory, the user would never touch that, so nothing in there changes. In reality, though, sometimes the user still has to dig in there and modify things (e.g. to undo a "bad" push; although of course for now you "fixed" that particular problem by force pushing). Of course in an ideal world, a user never should have to do that, but as long as there are bugs, users may have to. And based on painful experience, it is *much* easier to do that when the marks use the invariant hashes instead of the transient rev ids.

As for the wasted space: For a repository with 999,999 commits, the size of the "marks-hg file should increase from about 1.5 MB to about 5 MB. To use your words: Hardly worth mentioning.


>
>> * Better handling of various hg errors, see e.g. [2]. More work is still needed there with both tools, though [3].
>
> This is literally a three lines fix, and it simply makes one error
> nicer. Hardly worth mentioning.

Perhaps not for you, but for users who just want to use remote-hg resp. gitifyhg, getting a helpful error message instead of seeing an (to them) unreadable stack trace is quite important. I guess this just shows that we have quite different goals and ideas about how to interact with users of our respective tools :-).

Anyway, I hope you'll consider applying something like that three lines fix to remote-hg.

>
>> * Support for creating hg tags from git (i.e. pushing light git tags to heavy hg tags)
>
> remote-hg has the same.

Excellent.

>
>> * The gitifyhg test suite is run after each push on Travis CI against several git / mercurial combinations [4].
>> In particular, unlike all other remote-hg implementations I know, we explicitly promise (and test) compatibility with a specific range of Mercurial versions (not just the one the dev happens to have installed right now). This has been a frequent issue for me with the msysgit remote-hg
>
> I've personally checked against multiple versions of Mercurial. It's
> possible that some error might slip by, but it would get quickly
> noticed.

Really? This sounds close to some people who say things like "I don't need a test suite, I personally run some tests every now and then on my machine."
Of course clearly that is not at all how you operate. Rather you are very sensible and strive to provide a good testsuite which strives to test as much stuff as possible. As it should be. Which is why I am surprised that you compare (a) "personally" checking against some unspecified "multiple versions of Mercurial" at an unspecified frequency with (b) automatically running a testsuite after each push against a specific set of Mercurial versions. Versions which were specifically selected to match what Ubuntu 11.10, 12.04 LTS and 12.10 resp. Debian Wheezy provide, as well as recent versions. See also <https://github.com/buchuki/gitifyhg/blob/master/.travis.yml>. I would like to also let this check with multiple git versions, but that requires collecting / making suitable .debs which I have not yet gotten around to.

Furthermore, you don't seem to document what versions of Mercurial are supposed to work / not work. (Indeed, as far as I can tell, remote-hg has no documentation whatsoever, another difference. Granted, the gitifyhg README is not particular great at this point, but at least it exists and tells people how to get started and where to get help).

In contrast, the gitifyhg README clearly states that it "requires at least Mercurial 1.9". And its setup.py refuses to install it if the Mercurial version is too old. Again, for us devs that's not very important, but for users, I think it is. In recent months, I had to provide assistance with using hg to tons of people (*sigh*), and old Mercurial versions came up in a considerable portion of those (perhaps 30-40% or so).

>
>> * Renaming a gitifyhg remote just works [5]. Doing that with remote-hg triggers a re-clone of the remote repository (if it works at all, I don't remember).
>
> Yeah, now you can change the alias of the remote, but you can't change
> the remote url.

That's simply wrong. You still can change the remote URL, it will just lead to the creation of a fresh separate local clone.

In contrast, with remote-hg, renaming the remote will create a fresh local clone, while changing the remote URL will *not* do that -- instead, the changes from the new remote will be pulled into the existing local clone. Which in some cases may be exactly what you want (e.g. if a repository just moved to a different remote URL). But in other cases, it won't be, namely when the repository at the new URL contains different content. This is not that unlikely when you consider that it comment Mercurial workflow to use separate repositories for different branches.


> This is not really an advantage, simply an almost
> imperceptible different choice.

These are indeed different design choices, but at least to me, they are very much perceptible :-). And we made this design choice quite consciously, after looking at how it is done in remote-hg, and not liking that.

Anyway: At least in my day-to-day operations, I occasionally rename a remote (very rare, but it happens). So far I never had to change a remote URL. Of course that is just me, perhaps others occasionally (well, more of than me :-) have to change remote URLs. But as I said, that's still possible in gitifyhg.

That said, the situation is certainly not ideal. The fact that remote helpers have no good way to notice if the remote name (or URL, or anything) have changed is the root problem at hand here, I'd say. But for now, we are happy enough with the solution we implemented in gitifyhg, at least for our purposes.


> I still don't see any good reason why a user might prefer gitifyhg,

Well, for me it is exactly the other way around :-). That said, gitifyhg certainly also still has issues and problems, but I am confident it will even better in the coming weeks. And it certainly wouldn't be were it is now without your great work on remote-hg, and also on related improvements you got into git.git. Kudos!


> even more importantly, why gitifyhg developers don't contribute to
> remote-hg.

I can only speak for myself, as a (minor) contributor to both remote-hg and gitifyhg. But here are my reasons why I prefer contributing to remote-hg over gitifyhg:

1) Apparent difference between your goals and mine / those of gitifyhg. I think this is quite visible if one looks at our exchange above. Features that are quite important to me, even crucial, are unimportant for you, or even outright rejected. In contrast, so far with the couple people working on gitifyhg we always managed to arrive at solutions that satisfy all of us. I.e. it seems our ideas and goals are much better aligned within that group.

2) Lack of reactions on pull requests and bug reports on your github pages. Perhaps you never intended this to be used for pull requests / bug reports, but I never (until recently) saw you state that, nor can one read such a statement on your github pages.

3) Not shipping gitifyhg as part of git/contrib but rather as a separate package is important to me, too. It means that we can make new releases whenever we want, not tied to git. Users can install it easily via "pip", and I also think it gets a lot more visibility this way.


>
> Also, unlike remote-hg, which basically passes all the tests of
> gitifyhg, gitifyhg barely passes any tests of remote-hg (three).

Heh, bad, but OK (as I said, my message was based on an older version of remote-hg, and actually also on an older gitifyhg). Thank you for the report, I'll look into it as soon as I can (or somebody else might).

BTW, I just pulled you hg-next branch, and run "make test" in that. The tests in test-hg-hg-git.sh actually all failed (with remote-hg). Do I need to do something special for those to work?


Cheers,
Max

[Max Horn]

On 04.04.2013, at 08:46, Felipe Contreras wrote:

> On Thu, Apr 4, 2013 at 12:42 AM, Felipe Contreras
> <felipe.c...@gmail.com> wrote:

>> On Wed, Apr 3, 2013 at 6:25 PM, Max Horn <m...@quendi.de> wrote:
>>> On 03.04.2013, at 03:31, Felipe Contreras wrote:
>>
>>>> I only learned about it recently, I've looked at the history and to me
>>>> it seems rather chaotic, and a lot of the code was simply copied from
>>>> git-remote-hg without comment.
>>>
>>> gitifyhg was scrapped and completely restarted from scratch at some point. Based largely on your git-remote-hg code. A bit more on its history can be read here:
>>> http://archlinux.me/dusty/2013/01/06/gitifyhg-rewritten/
>

> Please don't CC the gitifyhg mailing list, unlike vger mailing lists
> (or any other sane list), it doesn't accept mail from non-subscribers,
> which makes communication with outsiders much more difficult, as
> demonstrated by this.

I changed the settings of the gitifyhg list settings to accept emails from anybody.

Moreover, I would appreciate if you could refrain from injecting all those snide side remarks, such as the one you just needlessly made about how moderated mailing lists are insane.

[Felipe Contreras]

On Wed, Apr 3, 2013 at 6:25 PM, Max Horn <m...@quendi.de> wrote:
> On 03.04.2013, at 03:31, Felipe Contreras wrote:

>> I only learned about it recently, I've looked at the history and to me
>> it seems rather chaotic, and a lot of the code was simply copied from
>> git-remote-hg without comment.
>
> gitifyhg was scrapped and completely restarted from scratch at some point. Based largely on your git-remote-hg code. A bit more on its history can be read here:
> http://archlinux.me/dusty/2013/01/06/gitifyhg-rewritten/

Yeah, I'm aware of that, but look at the README:

---
To the best of our knowledge, this is the most robust and usable git
to hg bridge currently available. It has a large test suite and better
documentation than the alternatives we know about. It has been tested
on several large mercurial repositories (including that of mercurial
itself and the pypy repository) that break with various other
git-to-hg bridge projects and is used daily in normal workflow
scenarios.
---

It reads as though gitifyhg was this marvelous piece of code developed
independently, and everything else (including remote-hg) pales in
comparison. There's no mention at all of remote-hg until the very
bitter end where it's mentioned that it's "heavily inspired by and
borrows code" from it, but it's much more than that, it's *based* on
this code. When I read the code, not only do I see it's the same
design, has the same function names, the same variables, the same
order, huge tracks of code are exactly the same, and in fact, I have
too look very carefully to see the differences.

Take a look at this commit:
http://github.com/buchuki/gitifyhg/commit/2c4fd7329fa0db8dd56df3ac8c62fa292580b74e

Geez, that's an awfully nice author sanitize function, but Dusty
Phillips was not the author, I am, he just added 2 lines of code.

But what really bothers me is that the first public version of
remote-hg happened after many many tries and commits, different
designs were discarded based on bad performance, feasibility, or
complexity, and gitifyhg just copies this code and runs with it.

Take a look at this function, which I took a long time to develop,
clean, and improve so it performs as efficiently as possible through
arduous profiling in huge repositories:

def get_filechanges(self, context, parent):
modified = set()
added = set()
removed = set()

current = context.manifest()
previous = self.repo[parent].manifest().copy()

for fn in current:
if fn in previous:
if (current.flags(fn) != previous.flags(fn) or
current[fn] != previous[fn]):
modified.add(fn)
del previous[fn]
else:
added.add(fn)
removed |= set(previous.keys())

return added | modified, removed

Nobody from gitifyhg even bothered to try to understand this function,
what it does, or why, just copied it:

http://github.com/buchuki/gitifyhg/commit/a46d518e2b8df5e8339c8caa9fa113642bc7ac3a

In fact, it's co clear the code was simply copied without much
thought, that the parent caller was introduced *before* the actual
function:

http://github.com/buchuki/gitifyhg/commit/54be060e1928b7e750bdf3981c32d5bc88851ed1

To me the strategy was clear: copy code of remote-hg until gitifyhg
works, then fix a few bugs, add a few features, and then claim
gitifyhg is superior, even though remote-hg will clearly fix those
bugs, and add those features soon enough. I don't see why nobody tried
to contact git.git, cooperate, and improve what is clearly the
original code of gitifyhg.

>>> * added many new test cases, sadly still including some xfails. Several of these (both passing and xfailing) also apply to remote-hg (i.e. the issue is also present in contrib's remote-hg)
>>
>> I ran these test-cases with remote-hg, and the same test-cases pass. I
>> only had to do minor modifications, most of the failures came from
>> subtle differences such as different strategies to sanitize authors,
>> and which branch to pick for HEAD.
>
> Yeah, that's because what I wrote was based on the situation before your recent patch series. Glad to see that git/contrib's remote-hg is improving!

No, that's not true, this test run is from v1.8.2 without the latest
fix patches, simply the compatibility ones:

========================================================================
test session starts
========================================================================
platform linux2 -- Python 2.7.3 -- pytest-2.3.4
collected 80 items

test/test_author.py ........F
test/test_clone.py ......xx.........x...x..
test/test_notes.py ..FxF
test/test_pull.py ....x..xx..
test/test_push.py ...FFFFF.FF...x....FFFFFFFFF
test/test_special_cases.py ...

=============================================================================
FAILURES ==============================================================================
/home/felipec/tmp/gitifyhg/test/helpers.py:118: assert 'totally
<bad...e used in hg>' == 'totally <unknown>'
/usr/lib/python2.7/site-packages/sh.py:309: ErrorReturnCode_128:
/home/felipec/tmp/gitifyhg/test/test_notes.py:107: assert not 'error'
in 'From hg::file:///tmp/pytest-0/test_simple_push_updates_notes_after_contentful_pull0/hg_base\n
66d7a5d..a8745f0 mas...rror: refs/notes/hg does not point to a
valid object!\nerror: refs/notes/hg-origin does not point to a valid
object!\n'
/home/felipec/tmp/gitifyhg/test/test_push.py:77: assert -1 > 0
/home/felipec/tmp/gitifyhg/test/test_push.py:86: assert -1 > 0
/home/felipec/tmp/gitifyhg/test/test_push.py:97: assert -1 > 0
/home/felipec/tmp/gitifyhg/test/test_push.py:117: assert 'default' ==
'branch_one'
/home/felipec/tmp/gitifyhg/test/test_push.py:136: assert 'default' ==
'branch one'
/home/felipec/tmp/gitifyhg/test/test_push.py:168: assert 'default' ==
'branch_one'
/home/felipec/tmp/gitifyhg/test/test_push.py:181: assert -1 > 0
/home/felipec/tmp/gitifyhg/test/test_push.py:334: assert 'anewbranch'
in 'no bookmarks set\n'
/home/felipec/tmp/gitifyhg/test/test_push.py:346: assert
'this_is_a_tag' in 'tip
0:1b3b36fc0158\n'
/home/felipec/tmp/gitifyhg/test/test_push.py:358: assert
'this_is_a_tag' in 'tip
1:7ca4e066c47f\n'
/home/felipec/tmp/gitifyhg/test/test_push.py:376: assert
'this_is_a_tag' in 'tip
2:5b8c7a98dfa4\nan_old_tag 0:af90a856304c\n'
/home/felipec/tmp/gitifyhg/test/test_push.py:391: assert
'this_is_a_tag' in 'tip
0:3c2288c2c93b\n'
/home/felipec/tmp/gitifyhg/test/test_push.py:407: assert
'this_is_a_tag' in 'tip
1:090867df4300\n'
/home/felipec/tmp/gitifyhg/test/test_push.py:420: assert 'this is a
tag' in 'tip 0:b7c5bfb3bdbc\n'
/home/felipec/tmp/gitifyhg/test/test_push.py:432: assert
'this_is_a_tag' in 'tip
1:10504139dfc4\nan_old_tag 0:b7c5bfb3bdbc\n'
/usr/lib/python2.7/site-packages/sh.py:309: ErrorReturnCode_128:
========================================================= 19 failed,
52 passed, 9 xfailed in 74.35 seconds
==========================================================

The only real bugs is that tags are not pushed, and named branches are
wrong, the rest are differences in how errors are reported. I doubt
whomever ran these tests spent much time analyzing the results.
Certainly there's no way to think that remote-hg was completely broken
and couldn't be fixed to pass these.

>>> * improved handling of hg user names (remote-hg is not able to deal with some pathological cases, failing to import commits). Sadly, mercurial allows arbitrary strings as usernames, git doesn't...
>>
>> I wouldn't call it improved. In some cases the remote-hg result is
>> better, in others gitifyhg is,
>
> I'd love to learn about cases where remote-hg's result is better in your opinion, so that I can see if the mapping in gitifyhg could be improved for those cases.

For example "no email quoting em...@example.com" results in "no email
quoting <em...@example.com>", but I already have the code that does
that in remote-hg.

> That said, this part is really quite subjective I guess. In the end, since Mercurial names can be *anything*, one can never get a "perfect" mapping. Luckily, for most real repositories out there, user names will be quite sane and remote-hg and gitifyhg will produce identical results. (Although some hg repos out there have some really weird stuff going on. Yuck.)

Indeed, which is why I advocate for an author mapping solution, so if
the user is not happy, they can manually fix the wrong authors.
Moreover, the author sanitizing should be done by 'git fast-import',
so that people don't have to implement and re-implement the same
sanitizing code all the time. Also, the mapping should be part of 'git
fast-import' so it works the same for all solutions.

>>> * failed pushes to hg are cleanly rolled back (using mq.strip() from the mq extension), instead of resulting in inconsistent internal state. This is quite important in real life, and has bitten me several times with remote-hg (and was the initial reason why I switched to gitifyhg). A typical way to reproduce this is to push to a remote repository that has commits not yet in my local clone.
>>
>> This is not an issue in remote-hg any more since now we force the
>> push. It's not nice, but there's no other way to push multiple
>> bookmarks (aka git branches) to the same branch (aka commit label).
>
> Uhh, what? The push is forced? That sounds to me like a much, much bigger issue with remote-hg than anything else ever was, from my point of view. That seems tobe akin to making "--force" default to "git push", and I don't think anybody here would consider that a good idea.

It's absolutely nothing like that; with 'git push --force' you
**loose** commits, with 'hg push --force' you don't, you simply have
another head. It's mercurial's fault that multiple bookmarks cannot be
pushed any other way.

>> I doubt these inconsistent states can happen any more, but if they do,
>
> Seriously? This is triggered quite frequently in real life. And it will very likely cause somebody to mess up a hg repository they work on. As long is this in, using remote-hg is a total no-go for me. Just consider the following scenario:
> * user A clones a hg repository into a git repository
> * user A commits some commits in the git clone
> * meanwhile, user B pushes changes to the hg repository
> * user A tries to push his changes to the hg remote
>
> The last step causes this result in gitifyhg (similar to what one gets when the remote is a git repos):
>

> With remote-hg, you just force push the change, creating a new head in the remote repo. So, yeah, failed pushes which mess up the internal state don't happen anymore. But I rather have those than potentially mess up the upstream repository like that.

Unfortunately that's per mercurial design; you can't have git-like
branches any other way.

>> the plan in remote-hg is to simply ignore those revisions, and only
>> push the ones that have git refs. I have the code for that, but I'll
>> not be pushing it to git.git for the time being.
>
> I am not quite sure what you mean here, but I'll just wait for your code and hope it'll explain itself.

Even if you end with a dangling head because of a failed push, you can
do 'hg push -r good_head', and the bad head will be ignored.

>>> * git notes are used to associate to each git commit the sha1 of the corresponding hg commit, to help users figure out that mapping
>>
>> This is a minor feature. I've had the code for this for quite some
>> time, but for the moment I think there are higher priorities.
>
> Well, I guess it depends on how you use this. In my daily work with hg repositories, that's a quite important feature.

It's more important to fix clone, fetch and push failures.

> Actually, I wonder: Are you actually using remote-hg yourself for day-to-day development? Or for any other regular activities? You seem to have quite different usage patterns from me and other people involved with gitifyhg.

No, I don't, but plenty of people use hg-git, and I'd say remote-hg is
superior, because it does the same as hg-git, and more.

>>> * internally, the marks are using the hg sha1s instead of the hg rev ids. The latter are not necessarily invariant, and using the sha1s makes it much easier to recover from semi-broken states.
>>
>> I doubt this makes any difference (except for more wasted space).
>
> I disagree. Especially since remote-hg does treat local hg repositories differently than remote ones, by not cloning them but rather working with them directly. As soon as somebody uses "hg strip" or "hg graft", etc. on that, your mark files will contain incorrect data.

That might be correct, but doing so on a public repository is not a
good idea, it causes problems even in git.

When I'm done with other tasks I will do some tests to see for sure
what are the real issues, what is merely hypothetical, and what is
likely. For the moment worrying a workflow that is actually a bad
practice is not sensible.

> With a remote hg repository, that is of course not as much of an issue, since you have a local hg clone stored inside of .git/. In theory, the user would never touch that, so nothing in there changes. In reality, though, sometimes the user still has to dig in there and modify things (e.g. to undo a "bad" push; although of course for now you "fixed" that particular problem by force pushing). Of course in an ideal world, a user never should have to do that, but as long as there are bugs, users may have to. And based on painful experience, it is *much* easier to do that when the marks use the invariant hashes instead of the transient rev ids.

How do you end up with a bad push in remote-hg?

> As for the wasted space: For a repository with 999,999 commits, the size of the "marks-hg file should increase from about 1.5 MB to about 5 MB. To use your words: Hardly worth mentioning.

It is a parenthesis (it was hardly mentioned).

>>> * Better handling of various hg errors, see e.g. [2]. More work is still needed there with both tools, though [3].
>>
>> This is literally a three lines fix, and it simply makes one error
>> nicer. Hardly worth mentioning.
>
> Perhaps not for you, but for users who just want to use remote-hg resp. gitifyhg, getting a helpful error message instead of seeing an (to them) unreadable stack trace is quite important. I guess this just shows that we have quite different goals and ideas about how to interact with users of our respective tools :-).

I have a suspicion that gitifyhg converts dates wrong, how is getting
a helpful error message on something the user would rarely (if ever)
do, more important than that (something the user always does)? We
might have different priorities, but I think mine are correct.

And don't speak about the "goals", because as I said I already have
the fix, saying we have different goals is implying that I don't want
to fix this issue, as if somehow not fixing a trivial error message
would not allow me to achieve my "goals". The priorities might be
different, but the goals are the same, and I think remote-hg is much
closer.

Also, anybody can send patches to the git mailing list, so _my_ goals,
and _my_ priorities shouldn't matter much.

>>> * The gitifyhg test suite is run after each push on Travis CI against several git / mercurial combinations [4].
>>> In particular, unlike all other remote-hg implementations I know, we explicitly promise (and test) compatibility with a specific range of Mercurial versions (not just the one the dev happens to have installed right now). This has been a frequent issue for me with the msysgit remote-hg
>>
>> I've personally checked against multiple versions of Mercurial. It's
>> possible that some error might slip by, but it would get quickly
>> noticed.
>
> Really? This sounds close to some people who say things like "I don't need a test suite, I personally run some tests every now and then on my machine."

Do you see any compatibility issues reported in the git mailing list,
or my github[1]? No? KTHXBYE. There _were_ compatibility issues, and
those got reported, and fixed, not any more.

Also, having CI testing doesn't ensure a project has better code, it's
the actual code that matters, and so far you haven't proved why
remote-hg's code would be inherently inferior. Moreover, gitifyhg
doesn't seem to have bidirectionality tests, remote-hg does, and
gitifyhg basically fails all of them, not to mention the ones that
compare the output with hg-git's, which has had far many more years of
development and testing, and their extensive tests have been harnessed
by remote-hg.

> Furthermore, you don't seem to document what versions of Mercurial are supposed to work / not work. (Indeed, as far as I can tell, remote-hg has no documentation whatsoever, another difference. Granted, the gitifyhg README is not particular great at this point, but at least it exists and tells people how to get started and where to get help).

When a user has a problem with that, we can deal with it. It's not
hard to do, and in fact it was done only one month ago in gitifyhg.
Nothing inherently different.

> In contrast, the gitifyhg README clearly states that it "requires at least Mercurial 1.9". And its setup.py refuses to install it if the Mercurial version is too old. Again, for us devs that's not very important, but for users, I think it is. In recent months, I had to provide assistance with using hg to tons of people (*sigh*), and old Mercurial versions came up in a considerable portion of those (perhaps 30-40% or so).

remote-hg certainly works on versions older than 1.9, again, I find it
annoying that you claim to know what is important for users, as if
somehow knowing that gitifyhg doesn't work with the user's version of
mercurial (e.g. 1.8) is better than remote-hg's situation; where it
*actually works*, but it's not mentioned. Yeah, mentioning that it
doesn't work is better than working, right.

>>> * Renaming a gitifyhg remote just works [5]. Doing that with remote-hg triggers a re-clone of the remote repository (if it works at all, I don't remember).
>>
>> Yeah, now you can change the alias of the remote, but you can't change
>> the remote url.
>
> That's simply wrong. You still can change the remote URL, it will just lead to the creation of a fresh separate local clone.

Exactly the same will happen with remote-hg when you change the alias.
Either creating a new local clone is "working", or it's not, you can't
have your cake and eat it at the same time. remote-hg reclones on
alias change, gitifyhg reclones on URL change. There is no advantage
here.

> In contrast, with remote-hg, renaming the remote will create a fresh local clone, while changing the remote URL will *not* do that -- instead, the changes from the new remote will be pulled into the existing local clone. Which in some cases may be exactly what you want (e.g. if a repository just moved to a different remote URL). But in other cases, it won't be, namely when the repository at the new URL contains different content. This is not that unlikely when you consider that it comment Mercurial workflow to use separate repositories for different branches.

It doesn't matter. The "extra content" that you get from the different
URL will just create that; extra content, both in the local git, and
mercurial repos. Since remote-hg will selectively push specific
revisions, that extra content will simply remain there, ignored.

>> This is not really an advantage, simply an almost
>> imperceptible different choice.
>
> These are indeed different design choices, but at least to me, they are very much perceptible :-). And we made this design choice quite consciously, after looking at how it is done in remote-hg, and not liking that.

This is not design, a few lines of code would change the behavior. And
you still haven't shown a single *important* use case where a user
would decide, you know what, yeah, I'll use gitifyhg because of that;
"I rename remote aliases a lot", "My upstream repo gets constantly
rebased" are very, very, unlikely phrases any user will ever utter.

And ultimately it doesn't matter, because if it's somehow decided that
one way is better than the other, we can switch it rather easily
(because they are not design decisions). So again, no, there's nothing
inherently superior here, merely subtle different behavior.

> Anyway: At least in my day-to-day operations, I occasionally rename a remote (very rare, but it happens). So far I never had to change a remote URL. Of course that is just me, perhaps others occasionally (well, more of than me :-) have to change remote URLs. But as I said, that's still possible in gitifyhg.

And I've changed the remote URL plenty of times, and again, changing
the alias works just fine as well.

>> I still don't see any good reason why a user might prefer gitifyhg,
>
> Well, for me it is exactly the other way around :-). That said, gitifyhg certainly also still has issues and problems, but I am confident it will even better in the coming weeks. And it certainly wouldn't be were it is now without your great work on remote-hg, and also on related improvements you got into git.git. Kudos!

You know what would be a proper way to thank? Mention where most of
the code and design came from in the README, and _at least_ mention
why you decided to work on your own rather than cooperate. When I
announced the work on remote-hg I discussed with the authors of other
solutions to try to explain why this work was needed until I
demonstrated that this was indeed superior, and I didn't even borrowed
code. I also provided a public summary when I announced it in my
blog[2].

>> even more importantly, why gitifyhg developers don't contribute to
>> remote-hg.
>
> I can only speak for myself, as a (minor) contributor to both remote-hg and gitifyhg. But here are my reasons why I prefer contributing to remote-hg over gitifyhg:
>
> 1) Apparent difference between your goals and mine / those of gitifyhg. I think this is quite visible if one looks at our exchange above. Features that are quite important to me, even crucial, are unimportant for you, or even outright rejected. In contrast, so far with the couple people working on gitifyhg we always managed to arrive at solutions that satisfy all of us. I.e. it seems our ideas and goals are much better aligned within that group.

Again, my goals are irrelevant, if you present a good case for a
patch, it will be merged by the maintainer, Junio, I'm not the
maintainer. But be prepared for argumentation from me, and others.

And no, I haven't rejected any patch, I couldn't reject a patch sent
to the git mailing list even if I wanted. I have only argued against
proposals, and if you give up when somebody argues against your ideas,
perhaps they were not so good in the first place.

But most of the "features" that you mentioned I said I ALREADY HAVE
THE PATCH FOR IT. So don't say we have different goals.

> 2) Lack of reactions on pull requests and bug reports on your github pages. Perhaps you never intended this to be used for pull requests / bug reports, but I never (until recently) saw you state that, nor can one read such a statement on your github pages.

I was not working on remote-hg. But it doesn't matter, git's
development happens in the git mailing list, and anybody can send
patches.

> 3) Not shipping gitifyhg as part of git/contrib but rather as a separate package is important to me, too. It means that we can make new releases whenever we want, not tied to git. Users can install it easily via "pip", and I also think it gets a lot more visibility this way.

% wget http://raw.github.com/felipec/git/fc/remote/hg/contrib/remote-helpers/git-remote-hg
-O ~/bin/git-remote-hg

There, you now have the latest and greatest version of remote-hg, you
don't need to wait for git.git to make a release. In fact, I'm in
contact with a few people that are having issues that are running
precisely this version; the latest and greatest.

>> Also, unlike remote-hg, which basically passes all the tests of
>> gitifyhg, gitifyhg barely passes any tests of remote-hg (three).
>
> Heh, bad, but OK (as I said, my message was based on an older version of remote-hg, and actually also on an older gitifyhg). Thank you for the report, I'll look into it as soon as I can (or somebody else might).
>
> BTW, I just pulled you hg-next branch, and run "make test" in that. The tests in test-hg-hg-git.sh actually all failed (with remote-hg). Do I need to do something special for those to work?

No, they should just work. Perhaps you have an old version of hg-git
(I have v0.4). You can check the errors with
GIT_TEST_OPTS="--verbose".

Cheers.

[1] https://github.com/felipec/git/issues
[2] http://felipec.wordpress.com/2012/11/13/git-remote-hg-bzr-2/

--
Felipe Contreras

[Felipe Contreras]

On Thu, Apr 4, 2013 at 3:07 AM, Max Horn <m...@quendi.de> wrote:
>
> On 04.04.2013, at 08:46, Felipe Contreras wrote:
>
>> On Thu, Apr 4, 2013 at 12:42 AM, Felipe Contreras
>> <felipe.c...@gmail.com> wrote:
>>> On Wed, Apr 3, 2013 at 6:25 PM, Max Horn <m...@quendi.de> wrote:
>>>> On 03.04.2013, at 03:31, Felipe Contreras wrote:
>>>
>>>>> I only learned about it recently, I've looked at the history and to me
>>>>> it seems rather chaotic, and a lot of the code was simply copied from
>>>>> git-remote-hg without comment.
>>>>
>>>> gitifyhg was scrapped and completely restarted from scratch at some point. Based largely on your git-remote-hg code. A bit more on its history can be read here:
>>>> http://archlinux.me/dusty/2013/01/06/gitifyhg-rewritten/
>>
>> Please don't CC the gitifyhg mailing list, unlike vger mailing lists
>> (or any other sane list), it doesn't accept mail from non-subscribers,
>> which makes communication with outsiders much more difficult, as
>> demonstrated by this.
>
> I changed the settings of the gitifyhg list settings to accept emails from anybody.

Cool.

> Moreover, I would appreciate if you could refrain from injecting all those snide side remarks, such as the one you just needlessly made about how moderated mailing lists are insane.

I did not say that, gitifyhg's mailing list was not _moderated_, it
*automatically* rejected all non-subscriber email without any
_moderation_; that is insane in my opinion, but a lot of mailing lists
do that.

--
Felipe Contreras

[Felipe Contreras]

On Tue, Apr 2, 2013 at 4:23 PM, Max Horn <m...@quendi.de> wrote:

> I'll try to list some of remaining differences, mostly (in my biased opinion) improvements on the gitifyhg side. Note that some of these might be outdated with felipe's recent changes, i.e. I have not yet had time to review and/or test them all. So please bear that in mind.

I've implemented a lot of these, cleaned them up, and pushed them, the
ones that will be integrated:
http://github.com/felipec/git/tree/fc/remote/hg-next

The ones that won't (at least not without more discussion):
http://github.com/felipec/git/tree/fc/remote/hg-gitifyhg-compat

> * added many new test cases, sadly still including some xfails. Several of these (both passing and xfailing) also apply to remote-hg (i.e. the issue is also present in contrib's remote-hg)

I don't think there's anything inherently better about these tests,
with the compatibility patches here are the results on v0.8 running
remote-hg:

=========================================================== test
session starts ===========================================================

platform linux2 -- Python 2.7.3 -- pytest-2.3.4
collected 80 items

test/test_author.py ........F
test/test_clone.py ......xx.........x...x..

test/test_notes.py ..Fx.
test/test_pull.py ....x..xx..
test/test_push.py ..........F...x........FF...
test/test_special_cases.py ...

============================================= 5 failed, 66 passed, 9
xfailed in 75.52 seconds =============================================

> * improved handling of hg user names (remote-hg is not able to deal with some pathological cases, failing to import commits). Sadly, mercurial allows arbitrary strings as usernames, git doesn't...

This is not true; after checking the code, remote-hg can't possibly
fail, if it does, so does gitifyhg. I guarantee it. The only
differences are cosmetic.

That being said, I'll integrate a patch that I believe produces
superior sanitation than gitifyhg's, and passes the gitifyhg test (as
you can see above) (for the most part):

https://github.com/felipec/git/commit/c0e363915eb6459233e37d5082fb2ff7c7c727b4

> * failed pushes to hg are cleanly rolled back (using mq.strip() from the mq extension), instead of resulting in inconsistent internal state. This is quite important in real life, and has bitten me several times with remote-hg (and was the initial reason why I switched to gitifyhg). A typical way to reproduce this is to push to a remote repository that has commits not yet in my local clone.

After the change to force=true, let's see if this happens any more in
remote-hg (Doubt it).

> * git notes are used to associate to each git commit the sha1 of the corresponding hg commit, to help users figure out that mapping

Easy:
https://github.com/felipec/git/commit/2294fb445f5c018a39f421cba70e4d8510c04c89

I will not integrate this for the moment, there must be a better way
to interact with transport-helper to update these.

> * internally, the marks are using the hg sha1s instead of the hg rev ids. The latter are not necessarily invariant, and using the sha1s makes it much easier to recover from semi-broken states.

I will investigate the pros and cons of this, but either way it's not
something people are going to immediately need (I doubt the
semi-broken states will happen again).

> * Better handling of various hg errors, see e.g. [2]. More work is still needed there with both tools, though [3].

No idea why something so trivial was mentioned:
https://github.com/felipec/git/commit/d12e35d23b9d26d384c3dbbce25a09720ccbceff

> * Support for creating hg tags from git (i.e. pushing light git tags to heavy hg tags)

This was already merged to git.git:
https://git.kernel.org/cgit/git/git.git/commit/?id=32f370f62177b505daf96aaf711c0249d881b6c0

(link might change)

> * The gitifyhg test suite is run after each push on Travis CI against several git / mercurial combinations [4].
> In particular, unlike all other remote-hg implementations I know, we explicitly promise (and test) compatibility with a specific range of Mercurial versions (not just the one the dev happens to have installed right now). This has been a frequent issue for me with the msysgit remote-hg

This is nice, but doesn't translate necessarily to anything tangible
for the user. remote-hg, like all git.git, has good development
practices, which minimizes the risks of regressions.

> * Renaming a gitifyhg remote just works [5]. Doing that with remote-hg triggers a re-clone of the remote repository (if it works at all, I don't remember).

Changing a remote-hg URL remote just works. Potato potato.

Cheers.

--
Felipe Contreras

[Felipe Contreras]

On Thu, Apr 4, 2013 at 10:14 AM, Felipe Contreras
<felipe.c...@gmail.com> wrote:
> On Tue, Apr 2, 2013 at 4:23 PM, Max Horn <m...@quendi.de> wrote:

>> * internally, the marks are using the hg sha1s instead of the hg rev ids. The latter are not necessarily invariant, and using the sha1s makes it much easier to recover from semi-broken states.
>
> I will investigate the pros and cons of this, but either way it's not
> something people are going to immediately need (I doubt the
> semi-broken states will happen again).

And another one bites the dust:
https://github.com/felipec/git/commits/fc/remote/hg-noteids

--
Felipe Contreras

[Max Horn]

<RANT>
While I am not really interested in exchanging any further emails or any other form of communication with Felipe, as I find his vitriolic style of communication unbearable, I feel compelled to reply to a few points. I'll probably regret this... anyway, I promise this will be my last mail in this thread. Even though I fully expect to receive a barrage of hostile and aggressive replies by Felipe. So be it, /dev/null has plenty space left.
</RANT>

OK, I'll try to keep a professional tone from now on :-).


On 04.04.2013, at 08:42, Felipe Contreras wrote:

[...]

>>>> * The gitifyhg test suite is run after each push on Travis CI against several git / mercurial combinations [4].
>>>> In particular, unlike all other remote-hg implementations I know, we explicitly promise (and test) compatibility with a specific range of Mercurial versions (not just the one the dev happens to have installed right now). This has been a frequent issue for me with the msysgit remote-hg
>>>
>>> I've personally checked against multiple versions of Mercurial. It's
>>> possible that some error might slip by, but it would get quickly
>>> noticed.
>>
>> Really? This sounds close to some people who say things like "I don't need a test suite, I personally run some tests every now and then on my machine."
>
> Do you see any compatibility issues reported in the git mailing list,
> or my github[1]? No? KTHXBYE. There _were_ compatibility issues, and
> those got reported, and fixed, not any more.

Please consider that the willingness of people to collaborate with you in any way is directly related to how you treat them. That includes bug reports. The way you acted towards Jed, who was very calmly and matter-of-factly explaining things, was IMHO completely inappropriate and unacceptable. Indeed, I should augment my list of reasons why people might not want to contribute to remote-hg by one major bullet point: You. And please, don't feel to compelled to tell us that Junio is really the maintainer of remote-hg and not you: Whether this is true or not doesn't matter for this point.

> remote-hg certainly works on versions older than 1.9, again

That's plain wrong. Indeed, remote-hg is using hg interfaces that were only introduced in 1.9. As such, I would be quite surprised if remote-hg worked with older hg versions, but I don't see why I should bother to test... Hmm, wait, I see a reason:

> , I find it
> annoying that you claim to know what is important for users, as if
> somehow knowing that gitifyhg doesn't work with the user's version of
> mercurial (e.g. 1.8) is better than remote-hg's situation; where it
> *actually works*, but it's not mentioned. Yeah, mentioning that it
> doesn't work is better than working, right.

I'll leave it to everybody to read what you wrote there, and contrast it with the following, and draw their own conclusions...

The reason why I did not write what exactly is wrong with remote-hg in hg 1.8 and older is that it is so obvious that I didn't think anybody would need handholding to verify it or find out the details :-). But since you "asked" for it:


$ hg --version
Mercurial Distributed SCM (version 1.8.4)
(see http://mercurial.selenic.com for more information)

Copyright (C) 2005-2011 Matt Mackall and others
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
$ git clone hg::foobar/ foobar.git
Cloning into 'foobar.git'...
Traceback (most recent call last):
File "/Users/mhorn/bin/git-remote-hg", line 846, in <module>
sys.exit(main(sys.argv))
File "/Users/mhorn/bin/git-remote-hg", line 819, in main
fix_path(alias, peer or repo, url)
File "/Users/mhorn/bin/git-remote-hg", line 765, in fix_path
repo_url = util.url(repo.url())
AttributeError: 'module' object has no attribute 'url'

Indeed, util.url was only added in 1.9.3. OK, let's work around that. Then local clones work. But of course in reality, most users will want to interact with a remote repository. But that's still broken:

$ git clone hg::ssh://h...@bitbucket.org/fingolfin/test-gitifyhg
Cloning into 'test-gitifyhg'...
Traceback (most recent call last):
File "/Users/mhorn/bin/git-remote-hg", line 1138, in <module>
sys.exit(main(sys.argv))
File "/Users/mhorn/bin/git-remote-hg", line 1107, in main
repo = get_repo(url, alias)
File "/Users/mhorn/bin/git-remote-hg", line 284, in get_repo
peer, dstpeer = hg.clone(myui, {}, url, local_path, update=False, pull=True)
TypeError: clone() got multiple values for keyword argument 'pull'


Right, clone() changed. And some more stuff. See <https://github.com/fingolfin/gitifyhg/commit/d3d37a7a853f3c8a1907bdaf933844128d5e7d81>.

There also was a good reason why I stopped at that point, but I don't remember the details right now. And quite frankly, I have zero incentive to even try to remember. But anyway, I don't think it's that useful to add support for 1.8, too, since one can't get back much further anyway. And upgrading to a newer Mercurial is (a) quite easy (even if you don't have root, just install it into $HOME), and (b) using a newer Mercurial version is a good idea for other reasons, too.

[...]

>
>>> Also, unlike remote-hg, which basically passes all the tests of
>>> gitifyhg, gitifyhg barely passes any tests of remote-hg (three).
>>
>> Heh, bad, but OK (as I said, my message was based on an older version of remote-hg, and actually also on an older gitifyhg). Thank you for the report, I'll look into it as soon as I can (or somebody else might).
>>
>> BTW, I just pulled you hg-next branch, and run "make test" in that. The tests in test-hg-hg-git.sh actually all failed (with remote-hg). Do I need to do something special for those to work?
>
> No, they should just work. Perhaps you have an old version of hg-git
> (I have v0.4). You can check the errors with
> GIT_TEST_OPTS="--verbose".

Yup, that was it, thanks. It would be kinda nice if the test code could check for suitable versions of mercurial and hg-git (and python, I guess) and warn the user if necessary.


As for your complaints about not getting proper credit in the gitifyhg README etc.: I agree that it is very much lacking in this regard, and will work towards rectifying this (indeed, I already suggested this to the other gitifyhg devs).

I'll also look into using sharness for gitifyhg test (which is based on the git test suite), as I also don't like the current test setup in gitifyhg, and indeed, the other gitifyhg devs agree. This would also make it easier to share and compare tests between remote-hg and gitifyhg if desired.

I won't reply to all the other stuff you wrote, as it just causes too much bile to raise up re-reading it, and I am not sure I could manage to reply in an even remotely neutral tone. So I'll refrain from attempting, as I am not interested in a fight between the two projects, or anybody for that matter. Nor do I see this as a competition where the "best" wins -- if somebody prefers remote-hg over gitifyhg, or the other way around, I don't care much, as long as at least one of the tools satisfy their needs.

Rather, all *I* am interested is using git to access a couple hg repositories that I absolutely must access. And helping several friend and colleagues who are in the precise same situation. Well, Jed and me already explained why e.g. forced push makes remote-hg an absolute no-go for me and several other people. Whether you accept this or not is irrelevant -- it sadly won't change the reality I and others have to deal with at work and elsewhere.


Cheers,
Max

[Felipe Contreras]

On Fri, Apr 5, 2013 at 4:30 PM, Max Horn <m...@quendi.de> wrote:

> On 04.04.2013, at 08:42, Felipe Contreras wrote:

> Please consider that the willingness of people to collaborate with you in any way is directly related to how you treat them. That includes bug reports. The way you acted towards Jed, who was very calmly and matter-of-factly explaining things, was IMHO completely inappropriate and unacceptable

He did not provide any bug report at all, also, he was not willing to
collaborate, as he clearly stated; he won't work git-hg bridges no
more.

> Indeed, I should augment my list of reasons why people might not want to contribute to remote-hg by one major bullet point: You. And please, don't feel to compelled to tell us that Junio is really the maintainer of remote-hg and not you: Whether this is true or not doesn't matter for this point.

Oh but it does. The only reason my remote-hg managed to land in the
mainline is because I showed it was superior to the other
remote-hg[1], and that there really was no point in trying to salvage
the other code base. Maybe this turns out to be a mistake, and the
other remote-hg manages to improve and surpass my remote-hg, but it
seems rather unlikely at this point.

Similarly, you could *show* that gitifyhg is superior than remote-hg,
why it makes sense to use that instead of trying to improve this code
base, but you don't, because you can't, because it doesn't make sense.

Ultimately this is not about people, this is about the code. A
sensible person that is not emotionally attached to any code, would
simply look at the code, and if what you claim is true (that remote-hg
has many issues), this sensible person would start cherry picking
patches from gitifyhg, send them to the git mailing list explaining
why they make sense. Eventually, remote-hg in git.git would become
essentially gitifyhg (albeit better because it would have received
more review from other git developers). Codewise this would make
sense.

And this sensible person would replace me as the go-to person for
remote-hg. And he would probably have a friendly and wonderful
personality, and you can accept him as your messiah and what not. But
this wouldn't change the fact that codewise this is the best way to
proceed.

But the fact of the matter is that either a) remote-hg is perfectly
fine, or b) this sensible person doesn't exist.

>> , I find it
>> annoying that you claim to know what is important for users, as if
>> somehow knowing that gitifyhg doesn't work with the user's version of
>> mercurial (e.g. 1.8) is better than remote-hg's situation; where it
>> *actually works*, but it's not mentioned. Yeah, mentioning that it
>> doesn't work is better than working, right.
>
> I'll leave it to everybody to read what you wrote there, and contrast it with the following, and draw their own conclusions...

Keep in mind that it's not me the one that claims remote-hg is
superior because it supports 1.8, as you seem try to portray. Rather,
it's _you_ the one that claims superiority because gitifyhg *mentions*
support until 1.9 (which remote-hg also supports).

It's not important who supports the most ancient version of mercurial
that nobody uses anyway, what is important is that *mentioning* or not
mentioning which ancient version of mercurial is supported doesn't
make one project superior to the other.

But perhaps this point is too subtle for you to understand, so there:
https://github.com/felipec/git/wiki/Git-remote-hg

Now it's mentioned that remote-hg too, supports up to version 1.9. Can
we agree now, that nobody has any advantage, either on version
support, or the mentioning of version support?

> Indeed, util.url was only added in 1.9.3. OK, let's work around that. Then local clones work. But of course in reality, most users will want to interact with a remote repository. But that's still broken:

Fixed:
https://github.com/felipec/git/commit/df0ed732b56c6c7910cc76f3e930229816e27117

And it was _you_ the one that sent the commit that broke it to the git
mailing list to be pushed.

> Right, clone() changed. And some more stuff. See <https://github.com/fingolfin/gitifyhg/commit/d3d37a7a853f3c8a1907bdaf933844128d5e7d81>.

Indeed, that might be useful, but that doesn't mean remote-hg doesn't
work with 1.8; it works perfectly fine for local repositories, and all
the tests pass.

> There also was a good reason why I stopped at that point, but I don't remember the details right now. And quite frankly, I have zero incentive to even try to remember. But anyway, I don't think it's that useful to add support for 1.8, too, since one can't get back much further anyway. And upgrading to a newer Mercurial is (a) quite easy (even if you don't have root, just install it into $HOME), and (b) using a newer Mercurial version is a good idea for other reasons, too.

If supporting 1.8 is not useful because most people have newer
versions, then mentioning support for 1.9 doesn't give any advantage.
Thanks for wasting our time on something nobody cares anyway.

>>> BTW, I just pulled you hg-next branch, and run "make test" in that. The tests in test-hg-hg-git.sh actually all failed (with remote-hg). Do I need to do something special for those to work?
>>
>> No, they should just work. Perhaps you have an old version of hg-git
>> (I have v0.4). You can check the errors with
>> GIT_TEST_OPTS="--verbose".
>
> Yup, that was it, thanks. It would be kinda nice if the test code could check for suitable versions of mercurial and hg-git (and python, I guess) and warn the user if necessary.

Yes, but how many people run these tests? And how many people have
hg-git in the first place? And how many people care about
test-hg-hg-git? I'll guess not so many, so I'm not really looking
forward to see what are the issues with older versions of hg-git.

> As for your complaints about not getting proper credit in the gitifyhg README etc.: I agree that it is very much lacking in this regard, and will work towards rectifying this (indeed, I already suggested this to the other gitifyhg devs).

Thanks.

> I'll also look into using sharness for gitifyhg test (which is based on the git test suite), as I also don't like the current test setup in gitifyhg, and indeed, the other gitifyhg devs agree. This would also make it easier to share and compare tests between remote-hg and gitifyhg if desired.

That would be really good.

> Jed and me already explained why e.g. forced push makes remote-hg an absolute no-go for me and several other people.

% git config remote-hg.force-push false

Problem solved. You don't need to work on a another project to avoid this.

> Whether you accept this or not is irrelevant -- it sadly won't change the reality I and others have to deal with at work and elsewhere.

The reality is you lack imagination. I've yet to see a single argument
that shows a problem with using a dedicated branch to push bookmarks.

> I won't reply to all the other stuff you wrote, as it just causes too much bile to raise up re-reading it, and I am not sure I could manage to reply in an even remotely neutral tone. So I'll refrain from attempting, as I am not interested in a fight between the two projects, or anybody for that matter. Nor do I see this as a competition where the "best" wins -- if somebody prefers remote-hg over gitifyhg, or the other way around, I don't care much, as long as at least one of the tools satisfy their needs.
>
> Rather, all *I* am interested is using git to access a couple hg repositories that I absolutely must access. And helping several friend and colleagues who are in the precise same situation.

So you are unable to argue why is it that gitifyhg is necessary, or
good. I take this as evidence that you are too emotionally invested to
accept the possibility that there's no good reason for the existence
of gitifyhg.

gitifyhg is not technically, or in any way, superior to remote-hg. You
should just give it up, and try to move as much code as possible to
remote-hg (is there any that will be useful?), and perhaps maintain a
branch to make remote-hg behave as you want (I already provided an
example), but on top of remote-hg. You can have your own github repo,
with your own wiki, and your own releases as you have it now. Then it
would be easier to share code, and if gitifyhg turns out to be truly
superior, it would be easier to simply merge those patches that enable
gitifyhg behavior, which you claim make all the difference in the
world.

It would be great if we could all work together in remote-hg, but
given the fact that at no point in time did you consider to "fix"
remote-hg, and that you spend considerable amounts of time claiming
that gitifyhg is superior, specially in remote-hg's own bug tracker,
rather than sending patches for remote-hg, which would prove your
point much more easily (if it was correct), I don't think you are
allowing for that possibility.

Cheers.

[1] https://github.com/msysgit/msysgit/wiki/Guide-to-git-remote-hg

--
Felipe Contreras

[Junio C Hamano]

Max Horn <m...@quendi.de> writes:

> OK, I'll try to keep a professional tone from now on :-).
>

> Please consider that the willingness of people to collaborate with
> you in any way is directly related to how you treat them. That
> includes bug reports. The way you acted towards Jed, who was very
> calmly and matter-of-factly explaining things, was IMHO completely
> inappropriate and unacceptable. Indeed, I should augment my list
> of reasons why people might not want to contribute to remote-hg by
> one major bullet point: You. And please, don't feel to compelled
> to tell us that Junio is really the maintainer of remote-hg and
> not you: Whether this is true or not doesn't matter for this
> point.

Only on this point, as the top-level maintainer. I do not have any
opinion on technical merits between the two Hg gateways myself.

A tool that is in contrib/ follows the contrib/README rule.

I do not maintain it. Maintenance is up to the person who asked to
include it there. I do ask the people who propose to add something
in contrib/ to promise that they arrange it to be maintained.

I do not even guarantee that they are the best in the breed in their
respective category. When something is added to contrib/, others can
raise objections by proposing alternatives, by arguing that tools of
the nature are better kept out of my tree, etc. When remote-hg was
added, I didn't see specific objections against it.

There is one generic objection to adding anything new in contrib/ I
have myself, though.

In early days of Git, almost all users, who might be interested in
improving their Git experience by helping to polish third-party
tools, had clones of my tree and did not hesitate to come to this
list. Back then, having a copy of an emerging third-party tool in my
tree in contrib/ was a good way to give more exposure to it, and to
give those interested in it a place to meet and join forces to
improve it. Because Git population was small, almost everybody was
here, and it was an efficient distribution mechanism.

Git is now reasonably well known and has big enough user base, and
many users, even those who are inclined to help improving their Git
experience by contributing to third-party tools, do not necessarily
have a clone of my tree. A third-party tool around Git, if it is
any good, is likely to have much much better chance to thrive as a
free-standing project with its own community, compared to those
early days.

[Felipe Contreras]

On Fri, Apr 5, 2013 at 7:28 PM, Junio C Hamano <git...@pobox.com> wrote:

> A tool that is in contrib/ follows the contrib/README rule.
>
> I do not maintain it. Maintenance is up to the person who asked to
> include it there. I do ask the people who propose to add something
> in contrib/ to promise that they arrange it to be maintained.

That's true, but I meant that you are the gatekeeper. Ultimately you
decide which patches go in. If Max, or anybody else, wants a patch
into contrib, you can get it in, even if I disagree with it. You can
also decide that somebody from gitifyhg might be a more suitable
person as a maintainer of remote-hg.

Either way, I think if things go well, remote-hg will prove it's worth
and move out of contrib and into git's core.

--
Felipe Contreras

[Philip Oakley]

From: "Felipe Contreras" <felipe.c...@gmail.com>
Sent: Saturday, April 06, 2013 1:45 AM

> On Fri, Apr 5, 2013 at 4:30 PM, Max Horn <m...@quendi.de> wrote:
>
>> On 04.04.2013, at 08:42, Felipe Contreras wrote:
>

>> Please consider [...]

> Ultimately this is not about people, this is about the code.

In the case of helper functions this is not the case.

The question would be better framed:
"Does this, or that, helper function make users (people) feel helped, or
frustrated (or somewhere in limbo)?".

I've called IT help desks and often felt frustrated, and some times I've
got one of the good girls/guys who worked with me to improve my
situation (often despite official policies). I get back to those folks
(even if they 'failed').

It's not a binary black/white issue when real users need help. It's no
good keeping with the faith (e.g. the Git ideal, the coders ideal, ..)
when the users (a mixed group) environmental doctine differs.

> A sensible person that is not emotionally attached to any code,

[I'm thinking users here, they are emotionally attached to their
original problem, and sense doesn't come into it]

> would simply look at the code,

Unfortunately, even for reasonable coders, looking at the code isn't
usually the case because of lack of time, unfamiliarity with the code,
extent of the code, availablity of the code (they may be simply running
a packaged/compiled 'app'), this is not that likely to happen. We should
be thankful when folk do look.

It's hard enough to get "good" bug reports from fellow coders (they are
only human / no more human than us) that tell us what _we_ want to know
(rather than what _they_ remember, or was important to them). ;-)


I don't use Hg, but as I read the discussion, there are
incomaptibilities between Git, and Hg. Thus neither helper can ever be
perfect. The winners will be those who solve a user need with enough
documentation and error capture to make them (their user group) feel
happy. At the moment it looks like the discussion is stratifying into
various "it worked for me" camps, each with their own problem children
repos that won't respond to parental advice, even with a --force from
social services.

Philip
[As they say back home: Between thee and me, ther's nowt so queer as
fowk, and I ain't so sure about thee]

[Felipe Contreras]

On Sat, Apr 6, 2013 at 8:09 AM, Philip Oakley <philip...@iee.org> wrote:
> From: "Felipe Contreras" <felipe.c...@gmail.com>
> Sent: Saturday, April 06, 2013 1:45 AM

>> Ultimately this is not about people, this is about the code.
>
>
> In the case of helper functions this is not the case.
>
> The question would be better framed:
> "Does this, or that, helper function make users (people) feel helped, or
> frustrated (or somewhere in limbo)?".

And that question is ultimately answered by code.

> I've called IT help desks and often felt frustrated, and some times I've got
> one of the good girls/guys who worked with me to improve my situation (often
> despite official policies). I get back to those folks (even if they
> 'failed').

That is not an issue when you don't have a situation that needs
support, when everything just works. Having one issue with bad support
is better than having 5 issues with good support.

> It's not a binary black/white issue when real users need help. It's no good
> keeping with the faith (e.g. the Git ideal, the coders ideal, ..) when the
> users (a mixed group) environmental doctine differs.

And neither Max or Jed are users, so this doesn't apply to them.

As for actual users, well show me, which remote-hg users have had bad support?

https://github.com/felipec/git/issues

>> A sensible person that is not emotionally attached to any code,
>
> [I'm thinking users here, they are emotionally attached to their original
> problem, and sense doesn't come into it]

That's irrelevant, a sensible developer would maximize the amount of
issues that do **not** hit the users.

>> would simply look at the code,
>
> Unfortunately, even for reasonable coders, looking at the code isn't usually
> the case because of lack of time, unfamiliarity with the code, extent of the
> code, availablity of the code (they may be simply running a
> packaged/compiled 'app'), this is not that likely to happen. We should be
> thankful when folk do look.

If this sensible developer doesn't have time to look at remote-hg
code, he has less time to make a possibly inferior alternative work on
par or better. Thus a sensible developer would either look at
remote-hg code, or do not develop.

> It's hard enough to get "good" bug reports from fellow coders (they are only
> human / no more human than us) that tell us what _we_ want to know (rather
> than what _they_ remember, or was important to them). ;-)

Nobody is asking for that.

> I don't use Hg, but as I read the discussion, there are incomaptibilities
> between Git, and Hg. Thus neither helper can ever be perfect. The winners
> will be those who solve a user need with enough documentation and error
> capture to make them (their user group) feel happy. At the moment it looks
> like the discussion is stratifying into various "it worked for me" camps,
> each with their own problem children repos that won't respond to parental
> advice, even with a --force from social services.

I don't think so. The discussion has been about hypothetical, not real
users. The one person that did claim was hit by a bug, had no evidence
for it nor cared, so it hardly matters. The rest is in pondering such
as if the user does this and that, and somebody else in the team does
that, and then the user does this, and the team has that policy, they
might end in a bit of a kerfuffle. Not something that has _actually_
happened.

And I'm confident that with time it will be shown that in terms of
real issues, remote-hg would have less, and be fixed faster.

Cheers.

--
Felipe Contreras

[Junio C Hamano]

Felipe Contreras <felipe.c...@gmail.com> writes:

> On Fri, Apr 5, 2013 at 7:28 PM, Junio C Hamano <git...@pobox.com> wrote:
>
>> A tool that is in contrib/ follows the contrib/README rule.
>>
>> I do not maintain it. Maintenance is up to the person who asked to
>> include it there. I do ask the people who propose to add something
>> in contrib/ to promise that they arrange it to be maintained.
>
> That's true, but I meant that you are the gatekeeper. Ultimately you
> decide which patches go in. If Max, or anybody else, wants a patch
> into contrib, you can get it in, even if I disagree with it.

In an ideal fantasy world, it could be true, but when I say "I do
not maintain it, and people who put it in contrib/ is responsible
for it", I really mean it.

I may find the behaviour/performance of the sub-maintainer of a part
in contrib/ unsatisfactory and have to take an administrative action
(e.g. to remove the problematic part out of contrib/), but I would
rather not to do that kind of thing, if possible. Instead I expect
people who have any code in my tree (even in contrib/) to act as a
responsible adult, taking and responding to constructive criticisms
well, and more importantly, nudging those whose utterance you find
are mostly noise into raising a more concrete and actionable issues
that would be useful to you as an input.

> Either way, I think if things go well, remote-hg will prove it's worth
> and move out of contrib and into git's core.

That was what you promised when we started carrying it in contrib/;
I am still hoping to see it happen when it matures.