The only fully supported thing is to have the main repository on one disk,
and .git/annex on another. Only commands that move data in/out of the annex
will need to copy it across devices.
There is only partial support for putting arbitrary subdirectories of
.git/annex on different devices. For one thing, but this can require more
copies to be done. For example, when .git/annex/tmp is on one device, and
.git/annex/journal on another, every journal write involves a call to
mv(1). Also, there are a few places that make hard links between various
subdirectories of .git/annex with createLink, that are not handled.
In the common case without cross-device, the new moveFile is actually
faster than renameFile, avoiding an unncessary stat to check that a file
(not a directory) is being moved. Of course if a cross-device move is
needed, it is as slow as mv(1) of the data.
The bug was that with --json, output lines were sometimes doubled. For
example, git annex init --json would output two lines, despite only running
one thing. Adding to the weirdness, this only occurred when the output
was redirected to a pipe or a file.
Strace showed two processes outputting the same buffered output.
The second process was this writer process (only needed to work around
bug #624389):
_ <- forkProcess $ do
hPutStr toh $ unlines paths
hClose toh
exitSuccess
The doubled output occurs when this process exits, and ghc flushes the
inherited stdout buffer. Why only when piping? I don't know, but ghc may
be behaving differently when stdout is not a terminal.
While this is quite possibly a ghc bug, there is a nice fix in git-annex.
Explicitly flushing after each chunk of json is output works around the
problem, and as a side effect, json is streamed rather than being output
all at the end when performing an expensive operaition.
However, note that this means all uses of putStr in git-annex must be
explicitly flushed. The others were, already.
It would be nice if command-specific options were supported. The first
difficulty is that which command is being called is not known until after
getopt; but that could be worked around by finding the first non-dashed
parameter. Storing the settings without putting them in the annex monad is
the next difficulty; it could perhaps be handled by making the seek stage
pass applicable settings into the start stage (and from there on to perform
as needed). But that still leaves a problem, what data type to use to
represent the options between getopt and seek?
Left out the backend usage graph for now, and bad/temp directory sizes
are only displayed when present. Also, disk usage is returned as a string
with units, which I can see changing later.
This is actually tricky, 45bbf210a1 added
the escaping because it's needed for rsync that does go over ssh.
So I had to detect whether the remote's rsync url will use ssh or not,
and vary the escaping.
Before, a merge was first calculated, by running various actions that
called git and built up a list of lines, which were at the end sent
to git update-index. This necessarily used space proportional to the size
of the diff between the trees being merged.
Now, lines are streamed into git update-index from each of the actions in
turn.
Runtime size of git-annex merge when merging 50000 location log files
drops from around 100 mb to a constant 4 mb.
Presumably it runs quite a lot faster, too.
Avoids doing auto-merging in commands that don't need fully current
information from the git-annex branch. In particular, git annex add no
longer needs to auto-merge. Affected commands: Anything that doesn't
look up data from the branch, but does write a change to it.
It might seem counterintuitive that we can change a value without first
making sure we have the current value. This optimisation works because
these two sequences are equivilant:
1. pull from remote
2. union merge
3. read file from branch
4. modify file and write to branch
vs.
1. read file from branch
2. modify file and write to branch
3. pull from remote
4. union merge
After either sequence, the git-annex branch contains the same logical content
for the modified file. (Possibly with lines in a different order or
additional old lines of course).
More accurately, it was supported already when map uses git-annex-shell,
but not when it does not.
Note that the user name cannot be shell escaped using git-annex's current
approach for shell escaping. I tried and some shells like dash cannot
cd ~'joey'. Rest of directory is still shell escaped, not for security but
in case a directory has a space or other weird character.
This is my own damn fault for not making UUID a real type, and then relying
on the type checker to ensure my refactoring was correct -- which it wasn't!
I should probably add code to clean up bogus entries in the uuid.log, but
right now I want to get the fix out there to prevent people experiencing
this bug.
I should also make UUID a real data type.
Thanks Valentin Haenel for a test case showing how non-fast-forward merges
could result in an ongoing pull/merge/push cycle.
While the git-annex branch is fast-forwarded, git-annex's index file is still
updated using the union merge strategy as before. There's no other way to
update the index that would be any faster.
It is possible that a union merge and a fast-forward result in different file
contents: Files should have the same lines, but a union merge may change
their order. If this happens, the next commit made to the git-annex branch
will have some unnecessary changes to line orders, but the consistency
of data should be preserved.
Note that when the journal contains changes, a fast-forward is never attempted,
which is fine, because committing those changes would be vanishingly unlikely
to leave the git-annex branch at a commit that already exists in one of
the remotes.
The real difficulty is handling the case where multiple remotes have all
changed. git-annex does find the best (ie, newest) one and fast forwards
to it. If the remotes are diverged, no fast-forward is done at all. It would
be possible to pick one, fast forward to it, and make a merge commit to
the rest, I see no benefit to adding that complexity.
Determining the best of N changed remotes requires N*2+1 calls to git-log, but
these are fast git-log calls, and N is typically small. Also, typically
some or all of the remote refs will be the same, and git-log is not called to
compare those. In the real world I expect this will almost always add only
1 git-log call to the merge process. (Which already makes N anyway.)
To get old behavior, add a .gitattributes containing: * annex.backend=WORM
I feel that SHA256 is a better default for most people, as long as their
systems are fast enough that checksumming their files isn't a problem.
git-annex should default to preserving the integrity of data as well as git
does. Checksum backends also work better with editing files via
unlock/lock.
I considered just using SHA1, but since that hash is believed to be somewhat
near to being broken, and git-annex deals with large files which would be a
perfect exploit medium, I decided to go to a SHA-2 hash.
SHA512 is annoyingly long when displayed, and git-annex displays it in a
few places (and notably it is shown in ls -l), so I picked the shorter
hash. Considered SHA224 as it's even shorter, but feel it's a bit weird.
I expect git-annex will use SHA-3 at some point in the future, but
probably not soon!
Note that systems without a sha256sum (or sha256) program will fall back to
defaulting to SHA1.
Checks location log information, and file contents.
Does not check that numcopies is satisfied, as .gitattributes information
about numcopies is not available in a bare repository. In practice, that
should not be a problem, since fsck is also run in a checkout and will
check numcopies there.
Specifically, disabled trying to update the git-annex branch on the remote,
since that data is never used by operations that act on such remotes.
Also, when copying content to such a remote, skip committing the presence
information changes to its git-annex branch. Leaving it in the journal there
is ok: Any command run on the remote that needs the info will flush the
journal.
This may partially solve this bug:
http://git-annex.branchable.com/bugs/fails_to_handle_lot_of_files/
Although I still see unreaped git processes piling up when doing a copy --to.
* This version of git-annex only works with git 1.7.7 and newer.
The breakage with old versions is subtle, and affects
annex.numcopies .gitattributes settings, so be sure to upgrade git
to 1.7.7. (Debian package now depends on that version.)
* Don't pass absolute paths to git show-attr, as it started following
symlinks when that's done in 1.7.7. Instead, use relative paths,
which show-attr only handles 100% correctly in 1.7.7. Closes: #645046
Unfortunatly I can find no way to work with the old and new gits, as
the old had bugs that require absolute paths, while the new doesn't like
them at all. And the behavior of git show-attr in 1.7.7. is the same as
eg, git add of an absolute path to a symlink, so seems entirely
intentional and not likely to change.
* git-annex now asks git-annex-shell to verify that it's operating in
the expected repository.
* Note that this git-annex will not interoperate with remotes using
older versions of git-annex-shell.
The reason for this check is to avoid git-annex getting confused about
what remote repository actually contains a value. It's a prerequisite for
supporting git insteadOf aliases.
* New or changed repository descriptions in uuid.log now have a timestamp,
which is used to ensure the newest description is used when the uuid.log
has been merged.
* Note that older versions of git-annex will display the timestamp as part
of the repository description, which is ugly but otherwise harmless.
This yields a second or so speedup in unused, find, etc. Seems that even
when the ByteString is immediately split and then converted to Strings,
it's faster.
I may try to push ByteStrings out into more of git-annex gradually,
although I suspect most of the time-critical parts are already covered
now, and many of the rest rely on libraries that only support Strings.
Added Git.ByteString which replaces Git IO methods with ones using lazy
ByteStrings. This can be more efficient when large quantities of data are
being read from git.
In Git.LsTree, parse git ls-tree output more efficiently, thanks
to ByteString. This benchmarks 25% faster, in a benchmark that includes
(probably predominately) the run time for git ls-tree itself.
In real world numbers, this makes git annex unused 2 seconds faster for
each branch it needs to check, in my usual large repo.
Using Sets is the right thing; they have constant size lookup like my
SizeList, and logn insertation, which beats nub to death.
Runs faster than --fast mode did before, and gives accurate counts.
13 seconds total runtime with a warm cache in a repository with 40 thousand
keys.
find: Rather than only showing files whose contents are present, when used
with --exclude --copies or --in, displays all files that match the
specified conditions.
Note that this is a behavior change for find --exclude! Old behavior
can be gotten with find --in . --exclude=...
get, drop: Added --auto option, which decides whether to get/drop content
as needed to work toward the configured numcopies.
The problem with bundling it up in optimize was that I then found I wanted
to run an optmize that did not drop files, only got them. Considered adding
a --only-get switch to it, but that seemed wrong. Instead, let's make
existing subcommands optionally smarter.
Note that the only actual difference between drop and drop --auto is that
the latter does not even try to drop a file if it knows of not enough
copies, and does not print any error messages about files it was unable to
drop.
It might be nice to make get avoid asking git for attributes when not in
auto mode. For now it always asks for attributes.
First, this ensures that git annex addurl, when run repeatedly with the
same url, doesn't create duplicate files, which it did before when it
fell back to the longer filename.
Secondly, the file part of an url is frequently not very descriptive on its
own.
The uri scheme, auth, and port is intentionally left out, as clutter.
This includes a generic JSONStream library built on top of Text.JSON
(somewhat hackishly).
It would be possible to stream out a single json document describing
all actions, but it's probably better for consumers if they can expect
one json document per line, so I did it that way instead.
Output from external programs used for transferring files is not
currently hidden when outputting json, which probably makes it not very
useful there. This may be dealt with if there is demand for json
output for --get or --move to be parsable.
The version, status, and find subcommands have hand-crafted output and
don't do json. The whereis subcommand needs to be modified to produce
useful json.
Using a single strictness annotation, in just the right place.
Tried several others, none of which helped and some of which potentially
hurt. This is only the second time I've really had to deal with this in
a year of using haskell, which is, I suppose not that bad.
Statting files returned by dirContents to see if they exist and are regular
files seems pretty useless. This code was originally part of fsck, and
perhaps the idea then was to avoid things returned by dirContents that were
not files. But it's certianly not needed in the current use cases for
getKeysPresent.
when a git repository is first being created. Clones will automatically
notice that git-annex is in use and automatically perform a basic
initalization. It's still recommended to run "git annex init" in any
clones, to describe them.
The tricky part about this is that to generate a key, the file must be
present already. Worked around by adding (back) an URL key type, which
is used for addurl --fast.
This was more complex than would be expected. unannex has to use git commit -a
since it's removing files from git; git commit filelist won't do.
Allow commands to be added to the Git queue that have no associated files,
and run such commands once.
This allows eg, `git-annex -c annex.rsync-options=-6 get file`
The overridden git configs are not passed on to git plumbing commands
that are run. Perhaps someone will find a need to do that, but I don't yet
and it would require storing more state to know what config settings
have been overridden and need to be passed on.
Could result in bad location log data for keys that contain [&:%] in their
names. (A workaround for this problem is to run git annex fsck.)
`git annex unused --from remote` could also run into the broken code.
Giulio Eulisse reported that on OSX, bad free space numbers were being
shown. It thought he had negative free space.
While the documentation is not clear, especially across OS's, it seems
likely that statfs uses unsigned long. It doesn't make sense for any
numbers to be negative.
This is substantially slower than using make, does not build or install
documentation, does not run the test suite, and is not particularly
recommended, but could be useful to some.
Rebenchmarked v2 vs v3, and v3 is now actually faster. Yes, storing data
in git, using git as a filesystem is actually faster than just using the
filesystem. If you do it just right. :)
All commands that often have to read a lot of information from
the git-annex branch should now be nearly as fast as before
the branch was introduced.
Before fsck was taking approximatly 3 hours, now it's running in 8 minutes.
The code is very nasty. It should be rewritten to read the header line
from git cat-file, and then read the specified number of bytes of content.
Since the logs have just been moved into the git-annex branch, don't need
to worry about backwards compatability with old versions of git-annex that
would fail to parse location logs with extra fields tacked on.
This is a new git subcommand, that does a generic union merge operation
between two refs, storing the result in a branch. It operates efficiently
without touching the working tree. It does need to write out a temporary
index file, and may need to write out some other temp files as well.
This could be useful for anything that stores data in a branch,
and needs to merge changes into that branch without actually checking the
branch out. Since conflict handling can't be done without a working copy,
the merge type is always a union merge, which is fine for data stored in
log format (as git-annex does), or in non-conflicting files
(as pristine-tar does).
This probably belongs in git proper, but it will live in git-annex for now.
---
Plan is to move .git-annex/ to a git-annex branch, and use git-union-merge
to handle merging changes when pulling from remotes.
Some preliminary benchmarking using real .git-annex/ data indicates
that it's quite fast, except for the "git add" call, which is as slow
as "git add" tends to be with a big index.
cp is still used when copying file from repos on the same filesystem, since
--reflink=auto can make it significantly faster on filesystems such as
btrfs.
Directory special remotes still use cp, not rsync. It's not clear what
tmp file should be used when rsyncing to such a remote.
get not honoring --from has surprised me a few times, so least surprise
suggests it should just behave like copy --from. This leaves the difference
between get and copy being that copy always requires the remote to copy
from, while get will decide whether to get a file from a key/value store or
a remote.
Avoid git reset here too, so I no longer need to care that it's much more
expensive than seems wise (but I asked the git list about that anyway).
It's not necessary to reset the staged file content from the index, as
the `git add` of the the symlink will replace it anyway.
`git commit` of unlocked files is still slow, since git still has to shove
their entire content into the index, only to have it be thrown away. So it's
still better to use `git annex add`
This takes advantage of the debug logging done by missingh, and I added
my own debug messages for executeFile calls. There are still some other
low-level ways git-annex runs stuff that are not shown by debugging,
but this gets most of it easily.
In particular, munge key filenames to comply with the IA's filename limits,
disable encryption, support their nonstandard way of creating buckets, and
allow x-amz-* headers to be specified in initremote to set item metadata.
Still TODO: initremote does not handle multiword metadata headers right.
Releasing before I have quite finished the code. Got a little caught
up in Anathem references. Time for a walk and then a tiny bit more coding
and possibly testing.
When it's stalled, there are 3 processes:
git annex
git ls-files
git check-attr
git-annex stalls trying to write to git check-attr, which stalls trying to
write to stdout (read by git-annex).
git ls-files does not seem to be involved directly; I've seen the stall when
it was still streaming out the file list, and after it had exited and
zombified.
The read and write are supposed to be handled by two different threads,
which pipeBoth forks off, thus avoiding deadlock. But it does deadlock.
(Certian signals unblock the deadlock for a while, then it stalls again.)
So, this is another case of WTF is the ghc IO manager doing today?
I avoid the issue by converting the writer to a separate process.
Possibly this was caused by some change in ghc 7 -- I'm offline and cannot
verify now, but I'm sure I used to be able to run git annex drop w/o it
hanging! And the code does not seem to have changed, except for commit
c1dc407941, which I tried reverting without
success. In fact, I reverted all the way back to 0.20110316 and still
saw the stall.
Update: Minimal test case:
import System.Cmd.Utils
main = do
as <- checkAttr "blah" $ map show [1..100000]
sequence $ map (putStrLn . show) as
checkAttr attr files = do
(_, s) <- pipeBoth "git" params $ unlines files
return $ lines s
where
params = ["check-attr", attr, "--stdin"]
Bug filed on ghc in debian, #624389
Fully tested and working, including resuming and encryption. (Though not
resuming when sending *with* encryption; gpg doesn't produce identical
output each time.)
Uses same layout as the directory special remote and the .git/annex/objects/
directory.
This was a real PITA to fix, since location logs can be staged in
both the current repo, as well as in local remote's repos, in
which case the cwd will not be in the repo. And git add needs different
params in both cases, when absolute paths are not used.
In passing, git annex fsck now stages location log fixes.
The test suite will not be run if it cannot be compiled.
It may be possible later to split off the quickcheck using tests into
a separate program and keep most of the tests using just hunit.
The remaining leaks are in hS3. The leak with encryption was worked around
by the use of the temp file. (And was probably originally caused by
gpgCipherHandle sparking a thread which kept a reference to the start
of the byte string.)
* Update Debian build dependencies for ghc 7.
* Debian package is now built with S3 support. Thanks Joachim Breitner for
making this possible, also thanks Greg Heartsfield for working to improve
the hS3 library for git-annex.
Also hid a conflicting new symbol from Control.Monad.State
This was a most surprising leak. It occurred in the process that is forked
off to feed data to gpg. That process was passed a lazy ByteString of
input, and ghc seemed to not GC the ByteString as it was lazily read
and consumed, so memory slowly leaked as the file was read and passed
through gpg to bup.
To fix it, I simply changed the feeder to take an IO action that returns
the lazy bytestring, and fed the result directly to hPut.
AFAICS, this should change nothing WRT buffering. But somehow it makes
ghc's GC do the right thing. Probably I triggered some weakness in ghc's
GC (version 6.12.1).
(Note that S3 still has this leak, and others too. Fixing it will involve
another dance with the type system.)
Update: One theory I have is that this has something to do with
the forking of the feeder process. Perhaps, when the ByteString
is produced before the fork, ghc decides it need to hold a pointer
to the start of it, for some reason -- maybe it doesn't realize that
it is only used in the forked process.
Stalls were caused by code that did approximatly:
content' <- liftIO $ withEncryptedContent cipher content return
store content'
The return evaluated without actually reading content from S3,
and so the cleanup code began waiting on gpg to exit before
gpg could send all its data.
Fixing it involved moving the `store` type action into the IO monad:
liftIO $ withEncryptedContent cipher content store
Which was a bit of a pain to do, thank you type system, but
avoids the problem as now the whole content is consumed, and
stored, before cleanup.
