IRC_Meetings: ghc-metting-2008-07-23.log

[17:01] JaffaCake:
hi all!
[17:01] BSP_:
afternoon!
[17:02] JaffaCake:
so, welcome to the #ghc meeting
[17:02] JaffaCake:
I thought today we could talk about version control systems
[17:03] JaffaCake:
we're still having issues with darcs, particularly on Windows
[17:03] BSP_:
have you investigated upgrading to the darcs2 format at all?
[17:03] JaffaCake:
I've tried an experimental conversion, but we haven't used it in anger yet
[17:04] JaffaCake:
if we stick with darcs, we'll certainly have to make that switch, becasue the darcs2 format fixes many of the bugs
[17:05] TMD:
Is there a solid idea of which darcs + windows bugs are left after a switch?
[17:05] dcoutts:
my (humble) opinion is that we should try darcs2 before abandoning darcs completely
[17:05] JaffaCake:
the main problem with switch VCs is that we also have all the library repos
[17:05] JaffaCake:
s/switch/switching
[17:05] Igloo:
We should also work out when a good time to switch (either to darcs2 or something else) is
[17:05] JaffaCake:
when we have the fewest active branches, presumably
[17:05] Igloo:
I think that just before we fork the 6.10 branch is probably best
[17:06] nominolo:
i have a lot of problems with darcs
[17:06] JaffaCake:
I believe tailor can recreate branches, even for a darcs2 conversion
[17:06] Igloo:
As most of the 6.10 work should be done by then, so VC issues aren't as important
[17:06] nominolo:
case-sensitivity issues, looping darcs on ./darcs-all, no bisect
[17:07] Igloo:
nominolo: Have you had looping problems since the bandwidth to monk was turned up?
[17:07] JaffaCake:
people have lost a lot of time to darcs
[17:07] nominolo:
Igloo: yes
[17:07] BSP_:
the merge bugs are what are killing me..
[17:07] nominolo:
yesterday
[17:07] • Igloo 
doesn't understand nominolo's problems
[17:07] nominolo:
right, having to avoid confilcts is a PTIA
[17:07] nominolo:
*PITA
[17:07] mainland (n=mainland@eecs-60-dhcp10.eecs.harvard.edu) joined the chat room.
[17:07] JaffaCake:
yes, definitely
[17:08] JaffaCake:
took me a whole day to merge my parallel GC branch
[17:08] JaffaCake:
and there were only 3 minor conflicts
[17:08] nominolo:
ouch
[17:08] BSP_:
wow
[17:08] JaffaCake:
just that one of the conflicts was in an early patch, so I had to replay the whole history
[17:09] nominolo:
i tried to convert the ghc repo to darcs2 format a few days ago.  except for warnings about lossy conversion (conflicts) i didn't find any problems
[17:09] JaffaCake:
can we trust darcs2 enough?
[17:09] BSP_:
i did see some complaints about performance regressions in the darcs bug tracker from the GHC team
[17:09] BSP_:
for the darcs2 format, that is
[17:09] JaffaCake:
darcs2 is definitely slower
[17:10] JaffaCake:
I did a local get yesterday, and it takes forever
[17:10] JaffaCake:
no idea why
[17:10] nominolo:
oh dear
[17:11] JaffaCake:
I assume git is the main alternative
[17:12] Deewiant:
mercurial would be nicer for windows users
[17:12] BSP_:
is it possible to do an incremental switch to darcs2, where we can synchronise patches between the old and new formats? i guess not, but i'm not massively well informed
[17:12] JaffaCake:
unless anyone wants to argue for mercurial or bzr?
[17:12] nominolo:
i have personally tried git and can say that is has enough features to be useful, and the #git channel seemed very helpful
[17:12] JaffaCake:
BSP_: I don't think so, no
[17:12] matthew-_:
I read through the mercurial and bzr tutorials over the last couple of days. They seem pretty evenly matched
[17:12] Deewiant:
bzr used to be slow, don't know about nowadays
[17:13] nominolo:
i think it is more complicated in some cases
[17:13] JaffaCake:
I couldn't see how to do things like unrecord and amend-record in mercurial
[17:13] JaffaCake:
except for the most recent commit
[17:13] Deewiant:
I've never had a need for that but I think that's what the mqueue extension is for
[17:14] JaffaCake:
it's very useful to be able to modify your history
[17:14] JaffaCake:
before sharing it, that is
[17:14] matthew-_:
could you explain why that is better than just adding an extra revision that fixes the mistake?
[17:15] dcoutts:
matthew-_: it makes you look like an uber hacker who designs perfectly from day one ;-)
[17:15] JaffaCake:
that could cause spurious conflicts later
[17:15] Deewiant:
yes, http://www.selenic.com/mercurial/wiki/index.cgi/MqExtension seems to be it
[17:15] lambdabot:
Title: MqExtension - Mercurial
[17:16] ttt-- (n=ubuntu@78-23-117-225.access.telenet.be) left the chat room. (Read error: 110 (Connection timed out))
[17:16] JaffaCake:
thanks Deewiant
[17:17] Deewiant:
but you can't mess around with anything in "prehistory" in Mercurial, which I think is a philosophical difference with git?
[17:17] nominolo:
rewriting too early isn't exactly safe in Git either
[17:17] Deewiant:
i.e. once you're done with the queues and you've made a proper commit, it's final
[17:17] nominolo:
at least i encountered some bugs there
[17:17] Deewiant:
(except for rollback which can undo the latest)
[17:17] JaffaCake:
matthew-_: so maybe I make more mistakes than average :) but I often find that I forgot to include something in a patch.  If I make two separate patches, then it's harder for somone else to know that the patches should go together.
[17:18] matthew-_:
JaffaCake: but with branch per bug dev model then I tend to treat branches as atomic units
[17:18] nominolo:
Deewiant: if you want to do things like bisect it's nice to have every version be correct
[17:18] nominolo:
ie, buildable
[17:18] Deewiant:
true
[17:19] matthew-_:
I guess it depends on how often you merge selected revisions rather than branches, or do cherry picking
[17:19] JaffaCake:
we do a lot of cherry picking
[17:19] nominolo:
oh, i meant to tell that to matthew-_
[17:20] JaffaCake:
but cherry picking isn't always easy with darcs either, particularly when you want to cherry pick a patch without its dependencies
[17:20] JaffaCake:
that works much better in git
[17:21] Deewiant:
bos wrote the Mercurial book so I suppose he could answer related questions fairly well :-)
[17:21] JaffaCake:
yep, true
[17:22] nominolo:
@seen bos
[17:22] lambdabot:
bos is in #ghc and #haskell. I don't know when bos last spoke.
[17:22] bos:
you can turn a committed change back into a patch if you  need to
[17:22] BSP_:
do we have any information on the relative adoption of bzr, hg and git?
[17:22] JaffaCake:
so if we were to switch, what happens to the library repos?
[17:22] BSP_:
because what i'm most concerned about is that the system will be actively maintained for some time to come
[17:22] Deewiant:
git > hg > bzr from what I understand
[17:22] nominolo:
BSP_: according to popcon, git is far ahead
[17:22] Deewiant:
I think bzr's quite rare, actually
[17:22] bos:
git has about 2x the number of users of hg, and only one significant project uses bzr.
[17:23] JaffaCake:
that rules out bzr, I think
[17:23] matthew-_:
quoting from: http://www.lshift.net/blog/2008/04/30/choosing-a-new-version-control-system
[17:23] lambdabot:
Title: LShift Ltd. » Choosing a new version control system, http://tinyurl.com/5tjklc incompatible encoding
[17:23] matthew-_:
"Git is used most famously by the Linux kernel, Bazaar by Ubuntu’s Launchpad development centre, and Mercurial by the Java and Mozilla projects"
[17:23] JaffaCake:
bos: so what's the situation with file/directory renames in hg?
[17:23] bos:
JaffaCake: should all work fine.
[17:24] JaffaCake:
but is it based on heuristics like git, or native like darcs?
[17:24] bos:
it's not based on heuristics.
[17:24] JaffaCake:
so it maintains accurate history of renames?
[17:24] bos:
yes.
[17:24] JaffaCake:
ok, thanks
[17:25] JaffaCake:
I'll try to break it later :)
[17:25] bos:
you have to tell it that you're renaming something of course.
[17:25] JaffaCake:
sure
[17:25] nominolo:
mercurial also seems more windows-friendly than git
[17:25] Deewiant:
yes, that's my main reason for recommending it
[17:25] bos:
it's also much easier to learn.
[17:25] nominolo:
git has git rename, too
[17:25] nominolo:
you only can't rename an empty directory
[17:26] Deewiant:
you can't even track an empty directory in Mercurial
[17:26] nominolo:
and there's bitbucket, hg also has something similar to github
[17:26] Deewiant:
which is perhaps somewhat annoying but not significantly so
[17:26] bos:
right. neither git nor mercurial track directories explicitly.
[17:26] JaffaCake:
git doesn't track renaming history, it tries to reconstruct it
[17:27] bos:
correct.
[17:27] JaffaCake:
it's quite easy to construct an example that goes wrong
[17:27] JaffaCake:
I did it on my first try
[17:27] nominolo:
oh, ok
[17:27] Deewiant:
good to know, I've always been wary of git in that regard :-)
[17:27] JaffaCake:
but adding some garbage to the files made it work again, because it detected the rename
[17:28] nominolo:
how does mercurial do merges?
[17:28] • bos 
`ap` shower
[17:28] bos:
three-way, interactive if needed.
[17:28] Deewiant:
with an external proggy.
[17:28] nominolo:
ok. same as git then
[17:29] nominolo:
i guess mercurial has something like bisect as well?
[17:29] JaffaCake:
good support for renaming would push me towards hg, but I'd really like to be able to fiddle with the history
[17:29] dbueno (n=dbueno@cpe-69-202-89-145.twcny.res.rr.com) left the chat room. (Read error: 110 (Connection timed out))
[17:29] BSP_:
in what ways is git windows unfriendly? in that it requires cygwin? but we need that for the GHC build anyway..
[17:30] MyCatVerbs:
BSP_: it's reputed to be very slow on Windows.
[17:30] JaffaCake:
git might be more windows-friendly than darcs :)
[17:30] Igloo:
Is bisect like darcs trackdown? If so I'm not convinced that it is useful
[17:30] MyCatVerbs:
BSP_: naturally the author isn't particularly interested in optimizing that particular case. ;P
[17:30] nominolo:
Igloo: yes
[17:30] nominolo:
Igloo: it uses binary search
[17:30] BSP_:
MyCatVerbs: right ;-) but wikipedia claims: "However, the recent rewriting of many shell commands in C have resulted in significant speed improvement on Windows"
[17:31] JaffaCake:
MyCatVerbs: I heard rumours it was getting faster, they rewrote more of it in C
[17:31] Deewiant:
JaffaCake: ah, there's also http://www.selenic.com/mercurial/wiki/index.cgi/TransplantExtension apparently
[17:31] lambdabot:
Title: TransplantExtension - Mercurial, http://tinyurl.com/2mdrc6
[17:31] nominolo:
well, Git isn't very windows-friendly by culture alone.  i don't think that will change soon
[17:31] MyCatVerbs:
BSP_, JaffaCake: oh, cool. I guess that was only to be expected, in the longer run.
[17:32] MyCatVerbs:
Actually, come to think of it, if they used to use shell scripts everywhere, no wonder it was slow on Windows.
[17:32] JaffaCake:
Deewiant: extensions are all very well, but they don't make for a coherent UI
[17:32] nominolo:
ah, so hg has a rebase command
[17:32] Deewiant:
I think Mercurial extensions are really easy to use, actually
[17:33] Deewiant:
just enable it in .hgrc and then the corresponding "hg foo" command works.
[17:33] MyCatVerbs:
What with the much-heavier processes and difficulty of supporting fork() and exec() efficiently on Windows - Hell, pretty much every single line that passes through the 'terp has to do both.
[17:33] nominolo:
most of git commands are extensions, too
[17:33] JaffaCake:
ok, fair enough
[17:34] nominolo:
their are just shell scripts on top of the core commands
[17:34] BSP_:
if git and hg are fairly comparable feature wise, whats the speed like?
[17:34] JaffaCake:
comparable :)
[17:34] nominolo:
hg is slightly slower
[17:34] BSP_:
git seems to have a reputation as the fastest
[17:34] nominolo:
but, mostly comparabel
[17:34] Deewiant:
I think it varies per operation
[17:34] Deewiant:
depending on what blog's benchmark you look at, of course :-P
[17:35] nominolo:
it would probably only really matter if ghc reaches 1M SLOC
[17:35] nominolo:
or more
[17:35] matthew-_:
well Linus wanted git to be as fast as bitkeeper was/is. So the inital designs were very heavily optimised
[17:35] nominolo:
which i hope will never happen
[17:35] JaffaCake:
we can't be that far off 1M if you count everything
[17:35] nominolo:
bos: how well does mercurial handle sub-modules?
[17:36] nominolo:
JaffaCake: well, the evil mangler will hopefully go
[17:36] MyCatVerbs:
BSP_: for most code bases, they're both fast enough that you don't care about the speed, so worry about feature sets instead. :)
[17:36] nominolo:
bos: for example for bisect/trackdown we need to link ghc and the core-libs, otherwise it's useless
[17:37] nominolo:
(since different library versions don't build with certain ghc versions and vice versa)
[17:37] BSP_:
nominolo: can git handle that?
[17:38] nominolo:
not nicely (yet), but you can write some scripts
[17:38] Igloo:
I'm not sure if we can change the libraries. We'd have to get agreement from hugs etc first
[17:39] JaffaCake:
yeah
[17:39] nominolo:
oh. hm.
[17:39] JaffaCake:
and Cabal
[17:40] nominolo:
ok.  so where are we now?
[17:40] JaffaCake:
we'd probably be stuck with needing to use two VCSs
[17:41] nominolo:
is there a way to have a reliable bridge?
[17:41] BSP_:
what problems can you forsee that causing?
[17:41] nominolo:
using, say, tailor?
[17:41] Igloo:
I don't think bridges are plausible
[17:42] JaffaCake:
BSP_: no technical problems, just hard to use two different tools
[17:42] Igloo:
The only problem that 2 VCSs would cause is that we'd keep typing the wrong thing
[17:42] matthew-_:
context sensitive wrappers?
[17:42] JaffaCake:
yep
[17:42] Igloo:
Maybe, if the VCSs' features map well enough
[17:43] • JaffaCake 
notes that hg is the easiest to type, followed by git, then darcs
[17:43] thoughtpolice:
nominolo: there's some mingw versions of git for windows that I had some good experiences with
[17:43] bos:
hg has partial submodule support, which i'm working on.
[17:43] thoughtpolice:
the main catch was that I think you had to run git-gc a little more frequently since garbage could build up easier
[17:43] Igloo:
JaffaCake: But I already have e.g. p aliased to (roughly) darcs pull -p, so that's easiest for me!
[17:44] nominolo:
JaffaCake: i always type "git" when i want to type "get".. so that's a con ;)
[17:44] bos:
Igloo: dons tells me you've been looking into organisations like SFC and SPI for haskell.org help.
[17:44] bos:
Igloo: i've done quite a bit of work with the SFC and SFLC, and i'm very happy with them. i'll be meeting various of their number during the week here.
[17:45] bos:
Igloo: let me know if you want anything said.
[17:45] Igloo:
bos: OK, thanks. Hopefully I'll be in touch soon
[17:46] • bos 
off to breakfast
[17:47] BSP_:
right, so what is the consensus on actually moving off darcs? are we going to do it?
[17:47] JaffaCake:
so another problem with staying with darcs is that annotate still doesn't work (in the sense that it takes several minutes to give a result)
[17:48] JaffaCake:
I use the git repository for annotate
[17:48] nominolo:
yes, and getting/pulling doesn't seem faster with v2
[17:48] nominolo:
JaffaCake: you have your own git repo?
[17:48] JaffaCake:
I'm using yours :)
[17:48] nominolo:
JaffaCake: oh, that's outdated
[17:48] JaffaCake:
the one from darcs.haskell.org
[17:48] nominolo:
because the update hook doesn't run
[17:49] JaffaCake:
right, but I'm usually not interested in recent history
[17:49] nominolo:
ah, i see
[17:49] JaffaCake:
I can get that easily enough from darcs
[17:49] JaffaCake:
with darcs changes | grep
[17:49] nominolo:
ok.  so i haven't tried mercurial, but i would be fine with git
[17:50] nominolo:
from what i hear hg would be fine, too
[17:50] JaffaCake:
anyway we can use the new git conversion tool that's much faster, right?
[17:50] nominolo:
probably, i didn't try to build it
[17:50] BSP_:
have we got a hg conversion of ghc anywhere?
[17:50] nominolo:
it requires patching darcs head
[17:50] JaffaCake:
BSP_: not that I'm aware of
[17:51] JaffaCake:
so yes, looks like we need to do an eval of hg too
[17:51] JaffaCake:
at this stage I'm tending towards switching... is anyone strongly opposed?
[17:53] JaffaCake:
I guess not!
[17:54] dcoutts:
so what's the decision? move to
[17:54] nominolo:
*crickets chirping*
[17:54] • dcoutts 
has been busy announcing the OpenSPARC project
[17:54] JaffaCake:
dcoutts: we need to look at hg too
[17:54] Igloo:
So I think that for each contender we need someone to convert the GHC repo and translate a list of darcs commands so we can try it out?
[17:55] JaffaCake:
Igloo: right
[17:55] dcoutts:
JaffaCake: so the decision is to switch from darcs1, and now we evaluate a short list of options?
[17:55] JaffaCake:
I think that would be the best plan, yes
[17:56] JaffaCake:
I don't think switching to darcs2 will buy us enough
[17:56] JaffaCake:
we'll still have performance problems and Windows bugs
[17:57] dcoutts:
unless someone fixes them
[17:57] dcoutts:
sigh
[17:57] MyCatVerbs:
Hang on, GHC still uses darcs1?
[17:57] nominolo:
which is unlikely to happen soon
[17:57] JaffaCake:
MyCatVerbs: yes
[17:58] MyCatVerbs:
Damn.
[17:58] JaffaCake:
dcoutts: I can see the Windows bugs being fixed, but not the performance issues
[17:58] MyCatVerbs:
Out of curiosity, have you tried darcs2 against ghc?
[17:58] Igloo:
Are you talking about "whatsnew is slow" or "conflict resolution is exponential"?
[17:58] • MyCatVerbs 
winces at the mention of the latter.
[17:58] JaffaCake:
Igloo: the first kind
[17:58] dcoutts:
JaffaCake: aye, probably so
[17:59] JaffaCake:
Igloo: annotate is too slow to use, and everything else is just slow
[17:59] dons:
http://www.reddit.com/r/programming/comments/6t3pa/Sun_Microsystems_funding_Haskell_on_multicore/ announce is out.
[17:59] lambdabot:
Title: Sun Microsystems funding Haskell on multicore OpenSPARC! : programming, http://tinyurl.com/5p7vjj
[17:59] matthew-_:
ironically, the continual development of GHC is likely to make darcs a more viable choice ;)
[18:00] JaffaCake:
Igloo: so is conflict resolution still exponential in darcs2?
[18:00] MarcWebe1 (n=marc@pD9E0A63E.dip.t-dialin.net) joined the chat room.
[18:00] JaffaCake:
with the darcs2 format, I mean
[18:00] MyCatVerbs:
JaffaCake: no, that bug was squished hard. It was specifically a large part of the raison d'etre for darcs2.
[18:01] nominolo:
matthew-_: how?
[18:01] Igloo:
Yeah, I understand that it's not a problem with the new format, although i haven't tried it myself
[18:02] matthew-_:
nominolo: it'll make darcs faster, due to better optimisations. Albeit I suspect too slowly
[18:02] nominolo:
matthew-_: that'll be maybe 5% / year.  in the best case
[18:02] matthew-_:
quite ;)
[18:03] BSP_:
nominolo: and that's assuming its compute bound...
[18:03] JaffaCake:
ok, I have to go, thanks folks!
[18:03] • dcoutts 
points out http://haskell.org/opensparc/
[18:03] JaffaCake:
dcoutts: nice
[18:03] nominolo:
with pictures!
[18:03] BSP_:
does anyone want to volunteer for the hg conversion? :)
[18:04] dcoutts:
any student ghc hackers interested in working on GHC's native code gen for SPARC ?
[18:04] dcoutts:
nominolo: of course :-)
[18:04] nominolo:
too bad i won't have time
[18:04] dcoutts:
:-(
[18:04] dcoutts:
doesn't all have to be in one go
[18:05] nominolo:
i'll have to talk to simon thompson then
[18:10] bringert (n=bringert@3-1-5-7a.gva.gbg.bostream.se) joined the chat room.
[18:10] bringert (n=bringert@3-1-5-7a.gva.gbg.bostream.se) left the chat room. (Client Quit)
[18:15] benny99 (n=bebenny@p5486F483.dip.t-dialin.net) joined the chat room.
[18:15] benny99 (n=bebenny@p5486F483.dip.t-dialin.net) left the chat room. (Read error: 104 (Connection reset by peer))
[18:16] TMD (n=tmdubui@144.51.26.41) left the chat room.
[18:16] MarcWeber (n=marc@pD9E0994B.dip.t-dialin.net) left the chat room. (Read error: 110 (Connection timed out))
[18:16] nominolo:
is someone responsible for putting the log somewhere?
[18:17] dejones (n=donnie@74-131-36-199.dhcp.insightbb.com) joined the chat room.
[18:19] dolio (n=dolio@216.68.186.59) left the chat room. (Read error: 104 (Connection reset by peer))
[18:19] dolio (n=dolio@216.68.186.59) joined the chat room.
[18:20] BSP_:
nominolo: igloo did the last one (http://hackage.haskell.org/trac/ghc/wiki/IRC_Meetings) but feel free :)
[18:20] lambdabot:
Title: IRC_Meetings - GHC - Trac
[18:21] malcolmw:
personally, I would be very sorry to see darcs abandoned by ghc, but then I don't actually trip over any of its foibles on a regular basis
[18:23] Igloo:
Oh, yes, please put up a log on that page if you have one
[18:23] Igloo:
malcolmw: That's basically my opinion, too
[18:23] mainland:
just out of curiosity---and please don't bludgeon me here---but what about switching to a non-distributed version control system?
[18:24] Igloo:
That's a no-go IMO
[18:24] thoughtpolice:
no offline commits seem to be the biggest issue in that case
[18:24] thoughtpolice:
imo, anyway
[18:24] mainland:
but one can always use something like darcs offline
[18:24] mainland:
do a subversion checkout, then a darcs init...
[18:24] bos:
bleugh.
[18:26] BSP_:
you could use something like svk, to build a distributed VCS on top of SVN
[18:26] BSP_:
but i'd kind of prefer something more mainstream :)
[18:27] mainland:
you can use whatever tool you want on top of SVN
[18:27] thoughtpolice:
git-svn would work too
[18:27] mainland:
darcs, got, hg, whatever
[18:27] Igloo:
mainland: And what's the advantage?
[18:27] thoughtpolice:
but regardless the cost-benefit seems too unbalanced
[18:27] thoughtpolice:
you can reap just as many benefits from using solely distributed version control if you ask me
[18:28] mainland:
sure, if they worked well
[18:28] thoughtpolice:
they do, darcs in particular just has problems
[18:28] bos:
it would be very silly to switch from a DVCS to SVN.
[18:29] thoughtpolice:
that's nothing against darcs
[18:29] thoughtpolice:
for small stuff it works well, and it's working on a hard problem
[18:29] bos:
even the original authors of SVN use hg for most of their revision control work now.
[18:29] nominolo:
BSP_: ok, your log format is nicer
[18:29] mainland:
will people really not have just as many problems, though perhaps different, with any other DVCS?
[18:29] nominolo:
no
[18:29] bos:
no.
[18:30] nominolo:
the problems with darcs are mostly performance
[18:30] thoughtpolice:
darcs has inherent issues that inhibit it's flexibility and performance
[18:30] nominolo:
and some missing features
[18:30] thoughtpolice:
especially for large stuff
[18:30] Svrog (n=ivan@d5153048F.access.telenet.be) left the chat room. ("Ex-Chat")
[18:31] nominolo:
darcs is nice and simple for smaller projects
[18:32] thoughtpolice:
imo if you want a vcs that can sustain 'high load,' darcs isn't it; not yet, anyway.
[18:32] nominolo:
BSP_: can you upload your log, please?
[18:33] nominolo:
i'll try to find out why mine didn't generate a log file
[18:33] BSP_:
nominolo: i'm trying to work out how to get colloquy to generate a plain text log file :)
[18:33] nominolo:
heh, me too
[18:34] waern (i=53915df2@gateway/web/ajax/mibbit.com/x-70ab84da77ee7d90)  left the chat room. ("http://www.mibbit.com ajax IRC Client")
[18:34] BSP_:
actually i didn't even turn on the logging so this is fruitless