1 bup: It backs things up
2 =======================
4 bup is a program that backs things up. It's short for "backup." Can you
5 believe that nobody else has named an open source program "bup" after all
8 Despite its unassuming name, bup is pretty cool. To give you an idea of
9 just how cool it is, I wrote you this poem:
12 What rhymes with awesome?
14 But that's irrelevant.
16 Hmm. Did that help? Maybe prose is more useful after all.
19 Reasons bup is awesome
20 ----------------------
22 bup has a few advantages over other backup software:
24 - It uses a rolling checksum algorithm (similar to rsync) to split large
25 files into chunks. The most useful result of this is you can backup huge
26 virtual machine (VM) disk images, databases, and XML files incrementally,
27 even though they're typically all in one huge file, and not use tons of
28 disk space for multiple versions.
30 - It uses the packfile format from git (the open source version control
31 system), so you can access the stored data even if you don't like bup's
34 - Unlike git, it writes packfiles *directly* (instead of having a separate
35 garbage collection / repacking stage) so it's fast even with gratuitously
36 huge amounts of data. bup's improved index formats also allow you to
37 track far more filenames than git (millions) and keep track of far more
38 objects (hundreds or thousands of gigabytes).
40 - Data is "automagically" shared between incremental backups without having
41 to know which backup is based on which other one - even if the backups
42 are made from two different computers that don't even know about each
43 other. You just tell bup to back stuff up, and it saves only the minimum
44 amount of data needed.
46 - You can back up directly to a remote bup server, without needing tons of
47 temporary disk space on the computer being backed up. And if your backup
48 is interrupted halfway through, the next run will pick up where you left
49 off. And it's easy to set up a bup server: just install bup on any
50 machine where you have ssh access.
52 - Bup can use "par2" redundancy to recover corrupted backups even if your
53 disk has undetected bad sectors.
55 - Even when a backup is incremental, you don't have to worry about
56 restoring the full backup, then each of the incrementals in turn; an
57 incremental backup *acts* as if it's a full backup, it just takes less
60 - You can mount your bup repository as a FUSE filesystem and access the
61 content that way, and even export it over Samba.
63 - It's written in python (with some C parts to make it faster) so it's easy
64 for you to extend and maintain.
67 Reasons you might want to avoid bup
68 -----------------------------------
70 - This is a very early version. Therefore it will most probably not work
71 for you, but we don't know why. It is also missing some
72 probably-critical features.
74 - It requires python >= 2.5, a C compiler, and an installed git version >=
77 - It currently only works on Linux, MacOS X >= 10.4,
78 NetBSD, Solaris, or Windows (with Cygwin). Patches to support
79 other platforms are welcome.
89 - Check out the bup source code using git:
91 git clone git://github.com/bup/bup
93 - Install the needed python libraries (including the development
96 On Debian/Ubuntu this is usually sufficient (run as root):
98 apt-get install python2.6-dev python-fuse
99 apt-get install python-pyxattr python-pylibacl
100 apt-get install linux-libc-dev
102 Substitute python2.5-dev if you have an older system. Alternately,
103 on newer Debian/Ubuntu versions, you can try this:
105 apt-get build-dep bup
107 On CentOS (for CentOS 6, at least), this should be sufficient (run
110 yum groupinstall "Development Tools"
111 yum install python python-dev
112 yum install fuse-python pyxattr pylibacl
113 yum install perl-Time-HiRes
115 In addition to the default CentOS repositories, you may need to add
116 RPMForge (for fuse-python) and EPEL (for pyxattr and pylibacl).
118 - Build the python module and symlinks:
126 (The tests should pass. If they don't pass for you, stop here and send
129 - You can install bup via "make install", and override the default
130 destination with DESTDIR and PREFIX.
132 Files are normally installed to "$DESTDIR/$PREFIX" where DESTDIR is
133 empty by default, and PREFIX is set to /usr. So if you wanted to
134 install bup to /opt/bup, you might do something like this:
136 make install DESTDIR=/opt/bup PREFIX=''
142 Binary packages of bup are known to be built for the following OSes:
145 http://packages.debian.org/search?searchon=names&keywords=bup
147 http://packages.ubuntu.com/search?searchon=names&keywords=bup
148 - pkgsrc (NetBSD, Dragonfly, and others)
149 http://pkgsrc.se/sysutils/bup
150 http://cvsweb.netbsd.org/bsdweb.cgi/pkgsrc/sysutils/bup/
156 - Try making a local backup as a tar file:
158 tar -cvf - /etc | bup split -n local-etc -vv
160 - Try restoring your backup tarball:
162 bup join local-etc | tar -tf -
164 - Look at how much disk space your backup took:
168 - Make another backup (which should be mostly identical to the last one;
169 notice that you don't have to *specify* that this backup is incremental,
170 it just saves space automatically):
172 tar -cvf - /etc | bup split -n local-etc -vv
174 - Look how little extra space your second backup used on top of the first:
178 - Restore your old backup again (the ~1 is git notation for "one older than
181 bup join local-etc~1 | tar -tf -
183 - Get a list of your previous backups:
185 GIT_DIR=~/.bup git log local-etc
187 - Make a backup on a remote server (which must already have the 'bup' command
188 somewhere in the server's PATH (see /etc/profile, etc/environment,
189 ~/.profile, or ~/.bashrc), and be accessible via ssh.
190 Make sure to replace SERVERNAME with the actual hostname of your server):
192 tar -cvf - /etc | bup split -r SERVERNAME: -n local-etc -vv
194 - Try restoring the remote backup tarball:
196 bup join -r SERVERNAME: local-etc | tar -tf -
198 - Try using the new (slightly experimental) 'bup index' and 'bup save'
199 style backups, which bypass 'tar' but have some missing features (see
200 "Things that are stupid" below):
203 bup save -n local-etc /etc
205 - Do it again and see how fast an incremental backup can be:
208 bup save -n local-etc /etc
210 (You can also use the "-r SERVERNAME:" option to 'bup save', just like
211 with 'bup split' and 'bup join'. The index itself is always local,
212 so you don't need -r there.)
214 That's all there is to it!
220 - FreeBSD's default 'make' command doesn't like bup's Makefile. In order to
221 compile the code, run tests and install bup, you need to install GNU Make
222 from the port named 'gmake' and use its executable instead in the commands
223 seen above. (i.e. 'gmake test' runs bup's test suite)
225 - Python's development headers are automatically installed with the 'python'
226 port so there's no need to install them separately.
228 - To use the 'bup fuse' command, you need to install the fuse kernel module
229 from the 'fusefs-kmod' port in the 'sysutils' section and the libraries from
230 the port named 'py-fusefs' in the 'devel' section.
232 - The 'par2' command can be found in the port named 'par2cmdline'.
234 - In order to compile the documentation, you need pandoc which can be found in
235 the port named 'hs-pandoc' in the 'textproc' section.
238 Notes on NetBSD/pkgsrc
239 ----------------------
241 - See pkgsrc/sysutils/bup, which should be the most recent stable
242 release and includes man pages. It also has a reasonable set of
243 dependencies (git, par2, py-fuse-bindings).
245 - The "fuse-python" package referred to is hard to locate, and is a
246 separate tarball for the python language binding distributed by the
247 fuse project on sourceforge. It is available as
248 pkgsrc/filesystems/py-fuse-bindings and on NetBSD 5, "bup fuse"
251 - "bup fuse" presents every directory/file as inode 0. The directory
252 traversal code ("fts") in NetBSD's libc will interpret this as a
253 cycle and error out, so "ls -R" and "find" will not work.
255 - It is not clear if extended attribute and POSIX acl support does
264 bup stores its data in a git-formatted repository. Unfortunately, git
265 itself doesn't actually behave very well for bup's use case (huge numbers of
266 files, files with huge sizes, retaining file permissions/ownership are
267 important), so we mostly don't use git's *code* except for a few helper
268 programs. For example, bup has its own git packfile writer written in
271 Basically, 'bup split' reads the data on stdin (or from files specified on
272 the command line), breaks it into chunks using a rolling checksum (similar to
273 rsync), and saves those chunks into a new git packfile. There is one git
276 When deciding whether to write a particular chunk into the new packfile, bup
277 first checks all the other packfiles that exist to see if they already have that
278 chunk. If they do, the chunk is skipped.
280 git packs come in two parts: the pack itself (*.pack) and the index (*.idx).
281 The index is pretty small, and contains a list of all the objects in the
282 pack. Thus, when generating a remote backup, we don't have to have a copy
283 of the packfiles from the remote server: the local end just downloads a copy
284 of the server's *index* files, and compares objects against those when
285 generating the new pack, which it sends directly to the server.
287 The "-n" option to 'bup split' and 'bup save' is the name of the backup you
288 want to create, but it's actually implemented as a git branch. So you can
289 do cute things like checkout a particular branch using git, and receive a
290 bunch of chunk files corresponding to the file you split.
292 If you use '-b' or '-t' or '-c' instead of '-n', bup split will output a
293 list of blobs, a tree containing that list of blobs, or a commit containing
294 that tree, respectively, to stdout. You can use this to construct your own
295 scripts that do something with those values.
299 'bup index' walks through your filesystem and updates a file (whose name is,
300 by default, ~/.bup/bupindex) to contain the name, attributes, and an
301 optional git SHA1 (blob id) of each file and directory.
303 'bup save' basically just runs the equivalent of 'bup split' a whole bunch
304 of times, once per file in the index, and assembles a git tree
305 that contains all the resulting objects. Among other things, that makes
306 'git diff' much more useful (compared to splitting a tarball, which is
307 essentially a big binary blob). However, since bup splits large files into
308 smaller chunks, the resulting tree structure doesn't *exactly* correspond to
309 what git itself would have stored. Also, the tree format used by 'bup save'
310 will probably change in the future to support storing file ownership, more
311 complex file permissions, and so on.
313 If a file has previously been written by 'bup save', then its git blob/tree
314 id is stored in the index. This lets 'bup save' avoid reading that file to
315 produce future incremental backups, which means it can go *very* fast unless
316 a lot of files have changed.
319 Things that are stupid for now but which we'll fix later
320 --------------------------------------------------------
322 Help with any of these problems, or others, is very welcome. Join the
323 mailing list (see below) if you'd like to help.
325 - 'bup save' and 'bup restore' have immature metadata support.
327 On the plus side, they actually do have support now, but it's new,
328 and not remotely as well tested as tar/rsync/whatever's. If you'd
329 like to help test, please do (see t/compare-trees for one
332 In addition, at the moment, if any strip or graft-style options
333 are specified to 'bup save', then no metadata will be written for
334 the root directory. That's obviously less than ideal.
336 - bup is overly optimistic about mmap. Right now bup just assumes
337 that it can mmap as large a block as it likes, and that mmap will
338 never fail. Yeah, right... If nothing else, this has failed on
339 32-bit architectures (and 31-bit is even worse -- looking at you,
342 To fix this, we might just implement a FakeMmap[1] class that uses
343 normal file IO and handles all of the mmap methods[2] that bup
344 actually calls. Then we'd swap in one of those whenever mmap
347 This would also require implementing some of the methods needed to
348 support "[]" array access, probably at a minimum __getitem__,
349 __setitem__, and __setslice__ [3].
351 [1] http://comments.gmane.org/gmane.comp.sysutils.backup.bup/613
352 [2] http://docs.python.org/2/library/mmap.html
353 [3] http://docs.python.org/2/reference/datamodel.html#emulating-container-types
355 - 'bup index' is slower than it should be.
357 It's still rather fast: it can iterate through all the filenames on my
358 600,000 file filesystem in a few seconds. But it still needs to rewrite
359 the entire index file just to add a single filename, which is pretty
360 nasty; it should just leave the new files in a second "extra index" file
363 - bup could use inotify for *really* efficient incremental backups.
365 You could even have your system doing "continuous" backups: whenever a
366 file changes, we immediately send an image of it to the server. We could
367 give the continuous-backup process a really low CPU and I/O priority so
368 you wouldn't even know it was running.
370 - bup currently has no features that prune away *old* backups.
372 Because of the way the packfile system works, backups become "entangled"
373 in weird ways and it's not actually possible to delete one pack
374 (corresponding approximately to one backup) without risking screwing up
377 git itself has lots of ways of optimizing this sort of thing, but its
378 methods aren't really applicable here; bup packfiles are just too huge.
379 We'll have to do it in a totally different way. There are lots of
380 options. For now: make sure you've got lots of disk space :)
382 - bup has never been tested on anything but Linux, MacOS, and Windows+Cygwin.
384 There's nothing that makes it *inherently* non-portable, though, so
385 that's mostly a matter of someone putting in some effort. (For a
386 "native" Windows port, the most annoying thing is the absence of ssh in
387 a default Windows installation.)
389 - bup needs better documentation.
391 According to a recent article about git in Linux Weekly News
392 (https://lwn.net/Articles/380983/), "it's a bit short on examples and
393 a user guide would be nice." Documentation is the sort of thing that
394 will never be great unless someone from outside contributes it (since
395 the developers can never remember which parts are hard to understand).
397 - bup is "relatively speedy" and has "pretty good" compression.
399 ...according to the same LWN article. Clearly neither of those is good
400 enough. We should have awe-inspiring speed and crazy-good compression.
401 Must work on that. Writing more parts in C might help with the speed.
405 Actually, that's not stupid, but you might consider it a limitation.
406 There are a bunch of Linux GUI backup programs; someday I expect someone
407 will adapt one of them to use bup.
413 bup has an extensive set of man pages. Try using 'bup help' to get
414 started, or use 'bup help SUBCOMMAND' for any bup subcommand (like split,
415 join, index, save, etc.) to get details on that command.
417 For further technical details, please see ./DESIGN.
423 bup is a work in progress and there are many ways it can still be improved.
424 If you'd like to contribute patches, ideas, or bug reports, please join the
427 You can find the mailing list archives here:
429 http://groups.google.com/group/bup-list
431 and you can subscribe by sending a message to:
433 bup-list+subscribe@googlegroups.com
435 Please see <a href="HACKING">./HACKING</a> for
436 additional information, i.e. how to submit patches (hint - no pull
437 requests), how we handle branches, etc.