Re: Lost Repository Was: Removing git-annex Repo
-BEGIN PGP SIGNED MESSAGE- Hash: SHA512 Am Mo den 5. Dez 2011 um 17:19 schrieb Joey Hess: Klaus Ethgen wrote: Can you check out the git-annex branch and run git-log on uuid.log, and see what the most recent change to it looked like? It is an update. After that I revert this update and the next time it will purged again. It sort of sounds as if you are checking out the git-annex branch, manually editing and committing a file, and seeing git-annex revert that change. Well, after the Bug happens several times I started to do the following after git-annex removed the occurrences: git checkout git-annex git cherry-pick id # This id is the first time I fixed it by hand git add trust.log uuid.log git commit -m Correcting update git checkout master The patch looks somethink like: diff --git a/trust.log b/trust.log index xxx..xxx 100644 --- a/trust.log +++ b/trust.log @@ -1 +1,2 @@ ---- 1 timestamp=1322867090.765867s +---- 0 timestamp=1322867165.394761s diff --git a/uuid.log b/uuid.log index xxx..xxx 100644 --- a/uuid.log +++ b/uuid.log @@ -1,2 +1,3 @@ ----- Backup timestamp=1322866827.929813s ---- Master timestamp=1322866770.445515s +---- Backup timestamp=1322866827.929813s +---- Clone timestamp=1322867722.827595s Which is the revert of this two files. That's expected behavior actually.. http://git-annex.branchable.com/internals/ explains why. Sorry, no it don't. I do not want to modify the git-annex branch as I know it is internal. But the situation gives me no other choice than to revert it every and every time. Regards Klaus - -- Klaus Ethgen http://www.ethgen.ch/ pub 4096R/4E20AF1C 2011-05-16 Klaus Ethgen kl...@ethgen.de Fingerprint: 85D4 CA42 952C 949B 1753 62B3 79D0 B06F 4E20 AF1C -BEGIN PGP SIGNATURE- Version: GnuPG v1.4.11 (GNU/Linux) iQGcBAEBCgAGBQJO3PVpAAoJEKZ8CrGAGfasAPcL/2G12mXQWQcNaDO6tEdL7hkJ CM5IA8xnxN4cOIWry3YbfDePwQ4Q3/rPnCEj/epGP7QIYB+tatPjJCCz+2ivprc7 GyZtih+C8cwYpad/T/QKEEAM2txMR2uy2kkGy43aFaCN2YRC/2KDmK5ePfxgcRTJ W+U5VyNu8Aury73WzNGc41e8R/Uple8QZz/r9fvP5c23MtNB83229cjMNBauw4Q6 IHuM0tBNDSY22rZ0MG7WRFgtzgPOZjGsShMVn1TJFpelTheOsCtc0GVjkbwVaGrF IND2Vo23kBFBlc6vy7g99lra7qAoAZxptfGbqZaMKEWXYVmAMiCB+KiBYE5CcH58 2e4sRbPj2xkVITD9RNlWrH7e/amBpf1w5a+i/gNmHMAGYvX0vYS2abwz+FdwqoQ2 ALhycZFig3kmYlDSw+64lS98j9TwAXdZ2IoBCwnJHRZrbrOLhcfBpmncZRW9TRa6 bW56MR5zZUzMpq8L4xtacmxKputo1WGsNhWXftpOGA== =uWye -END PGP SIGNATURE- ___ vcs-home mailing list vcs-home@lists.madduck.net http://lists.madduck.net/listinfo/vcs-home
Re: New integration branch
Adam Spiers wrote: 9c87f2352214175de307efedb8fd93889a26afbc Can you give an example of when this is needed? I can't remember but I definitely saw it happen at least once :-/ My worry is that, since that really shouldn't happen AFIACS, you were actually seeing a bug. Either that or it's a corner case I have not identified. 602f26714254f3c65389b7665d15d1d5d0e227a9 mr is quite typically (I know, not by you) run inside the repository. Which would leave the user in an apparently empty directory after mr update if an mr update deleted and remade the whole repository. I don't like that; I don't think things in mr should be deleting repositories in general; mr doesn't even delete a repo that has deleted = true, it only warns the user about it. Hmm, that's a fair point, although the only alternative is to change the contents of the directory rather than the directory itself - similarly to how `git checkout' does, for instance. I'll see if I can get around to doing that. Perhaps some of the effort could be reused for implementing download_diff (diff against the archive file). I think you could just use rsync :) cf3388f443b9d7afe6dc7d8a2159b45fb01ab4e4 This is a slow way to make machine-parsable info available -- the similar mr list takes 8 seconds here, since it has to run 169 shells. That's ok when you're just running mr, but I would not like to use a command that depended on that information. Sure, that's why I used a simple on-disk cache: https://github.com/aspiers/kitenet-mr/commit/b60acb2e767b91ca6d406198d7eea1b3f73ad2bf It works fine. I could get more sophisticated and allow per-user configuration of the cache invalidation strategy, e.g. so that it would automatically rebuild the cache when ~/.mrconfig et al. are changed, but manual rebuilds aren't a great hardship. In fact I could even rebuild the cache every time mr runs! If a machine-parseable list of repositories is needed, I think it'd be better to have a perl function that emits it in one go. I don't see how that's possible without ignoring the `skip', `deleted', and `include' parameters. The include parameter is not a big problem, it's unlikely to require more than one shell process, which will be relatively fast. It's not clear to me what should be done about skip and deleted. skip in particular can behave in weird ways, when something like hours_since is used. To handle that all the skips would need to be tested, which would be less work than mr list but still verging on expensive. Depending on the application, it might be better to just dump all the defined repositories including skipped and deleted ones; if the consumer than runs mr in a skipped/deleted repo, mr will do something sane after all. -- see shy jo signature.asc Description: Digital signature ___ vcs-home mailing list vcs-home@lists.madduck.net http://lists.madduck.net/listinfo/vcs-home
Re: alternative mechanisms for including config
Adam Spiers wrote: This may be a good time to discuss the design of the `include' parameter. When you were deciding what its value should be, I guess there were at least three possibilities: (1) a chunk of shell-code which returns the actual shell-code to include (2) a chunk of shell-code which returns a list of names of files to include (3) a delimited list of files to include You went with (1). One advantage of this is the ability to dynamically generate code to include. But this could also be achieved with (2), by generating the files to include and then returning the names of the generated files. Also, with (1), if the shell-code has an issue it's harder to debug because there's no containing file (and line number and surrounding lines) to refer to. The main advantage of (3) is that you don't have to execute any shell code at all. This would facilitate implementation of your suggestion of writing a Perl function to emit the repo list, although there's still the problem of the `skip' parameter, and I suspect too many users are already relying on the dynamic nature of `include' for (3) to be feasible. But might it be worth implementing (2) alongside the existing (1), via a new `includefiles' special parameter? I've made mr show the included line content in error messages now. The speed hit of running that one shell command is minor. It doesn't seem worth bothering users with deprecating the current include, and needless complication to have a separate way with a list of files. -- see shy jo signature.asc Description: Digital signature ___ vcs-home mailing list vcs-home@lists.madduck.net http://lists.madduck.net/listinfo/vcs-home
Re: New integration branch
On Mon, Dec 5, 2011 at 5:38 PM, Joey Hess j...@kitenet.net wrote: Adam Spiers wrote: 9c87f2352214175de307efedb8fd93889a26afbc Can you give an example of when this is needed? I can't remember but I definitely saw it happen at least once :-/ My worry is that, since that really shouldn't happen AFIACS, you were actually seeing a bug. Either that or it's a corner case I have not identified. It's on my TODO list to try to reproduce; I'll let you know if I manage to. 602f26714254f3c65389b7665d15d1d5d0e227a9 mr is quite typically (I know, not by you) run inside the repository. Which would leave the user in an apparently empty directory after mr update if an mr update deleted and remade the whole repository. I don't like that; I don't think things in mr should be deleting repositories in general; mr doesn't even delete a repo that has deleted = true, it only warns the user about it. Hmm, that's a fair point, although the only alternative is to change the contents of the directory rather than the directory itself - similarly to how `git checkout' does, for instance. I'll see if I can get around to doing that. Perhaps some of the effort could be reused for implementing download_diff (diff against the archive file). I think you could just use rsync :) Yeah, that sounds worth trying. If a machine-parseable list of repositories is needed, I think it'd be better to have a perl function that emits it in one go. I don't see how that's possible without ignoring the `skip', `deleted', and `include' parameters. The include parameter is not a big problem, it's unlikely to require more than one shell process, which will be relatively fast. It's not clear to me what should be done about skip and deleted. skip in particular can behave in weird ways, when something like hours_since is used. To handle that all the skips would need to be tested, which would be less work than mr list but still verging on expensive. Depending on the application, it might be better to just dump all the defined repositories including skipped and deleted ones; if the consumer than runs mr in a skipped/deleted repo, mr will do something sane after all. Skipper functions like hours_since could (and probably should) decide not to skip if MR_ACTION is set to a read-only action such list - arguably diff and status too, although that's a matter of personal taste. But maybe we should step back a bit. Currently we know that a full mr list is not particularly fast, but has anyone actually profiled it to find out where most of the time is being spent? If we're only guessing then we might have it completely wrong ... ___ vcs-home mailing list vcs-home@lists.madduck.net http://lists.madduck.net/listinfo/vcs-home
Re: New integration branch
Joey Hess wrote: It could well not be. mr list -j 10 runs in the same time as mr list -j 1, suggesting the overhead is in something else than actually running the shell. Whoops, bad benchmark, -j comes before action. Anyway, yes, without any calls to system(), mr list takes just 0.35 seconds. Those calls are: 169 mr list: running set -e; # actual list command 118 mr skip: running vcs test # 55 mr list: running skip test set -e; 50 mr deleted: running vcs test (Note that the vcs test is split between skip and deleted, but if those features are removed, the actual list command would trigger the same test, so those don't add overhead.) Moving the git_test etc into perl code would be one way to speed it up for the common case. Adding a special case optimisation to avoid the shell for true and false brings mr list down from 8.50 to 1.81 seconds. The remaining time is here spent running skip tests, I have a lot. Probably looking at sub-1-second times for most people. -- see shy jo signature.asc Description: Digital signature ___ vcs-home mailing list vcs-home@lists.madduck.net http://lists.madduck.net/listinfo/vcs-home
Re: New integration branch
Joey Hess wrote: Moving the git_test etc into perl code would be one way to speed it up for the common case. Adding a special case optimisation to avoid the shell for true and false brings mr list down from 8.50 to 1.81 seconds. The remaining time is here spent running skip tests, I have a lot. Probably looking at sub-1-second times for most people. These optimisations are now in place. joey@gnu:~/src/d-itime mr -q list 1.14user 2.17system 0:05.12elapsed 64%CPU (0avgtext+0avgdata 26368maxresident)k 0inputs+0outputs (0major+269034minor)pagefaults 0swaps joey@gnu:~/src/d-itime ~/src/mr/mr -q list 0.38user 0.02system 0:00.44elapsed 91%CPU (0avgtext+0avgdata 26640maxresident)k 0inputs+0outputs (0major+6429minor)pagefaults 0swaps joey@gnu:~time mr -q list 1.67user 3.86system 0:08.75elapsed 63%CPU (0avgtext+0avgdata 26720maxresident)k 0inputs+0outputs (0major+464487minor)pagefaults 0swaps joey@gnu:~time ~/src/mr/mr -q list 0.56user 0.60system 0:01.78elapsed 65%CPU (0avgtext+0avgdata 26800maxresident)k 0inputs+0outputs (0major+84959minor)pagefaults 0swaps -- see shy jo signature.asc Description: Digital signature ___ vcs-home mailing list vcs-home@lists.madduck.net http://lists.madduck.net/listinfo/vcs-home