53New Unison beta-release (2.40)
- Apr 25 1:03 PMOn behalf of the Unison team, I'm delight to announce the availability of a new beta-release, version 2.40:
Source code is available now for those that want to compile it themselves. Binaries for various architectures will appear in due course.
Major kudos to Jérome Vouillon for the lion's share of the hacking on the new features!
Changes since 2.32:
* Major enhancement: Unicode support.
+ Unison should now handle unicode filenames correctly on all
+ This functionality is controlled by a new preference unicode.
+ Unicode mode is now the default on all architectures.
This may make upgrades a bit more painful (the archives
cannot be reused), but it is the right default.
* Partial transfer of directories. If an error occurs while
transferring a directory, the part transferred so far is copied
into place (and the archives are updated accordingly). The
"maxerrors" preference controls how many transfer error Unison
will accept before stopping the transfer of a directory (by
default, only one). This makes it possible to transfer most of a
directory even if there are some errors. Currently, only the first
error is reported by the GUIs.
Also, allow partial transfer of a directory when there was an
error deep inside this directory during update detection. At the
moment, this is only activated with the text and GTK UIs, which
have been modified so that they show that the transfer is going to
be partial and so that they can display all errors.
* Improvement to the code for resuming directory transfers:
+ if a file was not correctly transferred (or the source has
been modified since, with unchanged size), Unison performs a
new transfer rather than failing
+ spurious files are deleted (this can happen if a file is
deleted on the source replica before resuming the transfer;
not deleting the file would result in it reappearing on the
* Experimental streaming protocol for transferring file contents
(can be disabled by setting the directive "stream" to false): file
contents is transfered asynchronously (without waiting for a
response from the destination after each chunk sent) rather than
using the synchronous RPC mechanism. As a consequence:
+ Unison now transfers the contents of a single file at a time
(Unison used to transfer several contents simultaneously in
order to hide the connection latency.)
+ the transfer of large files uses the full available bandwidth
and is not slowed done due to the connection latency anymore
+ we get performance improvement for small files as well by
scheduling many files simultaneously (as scheduling a file
for transfer consume little ressource: it does not mean
allocating a large buffer anymore)
* Changes to the internal implementation of the rsync algorithm:
+ use longer blocks for large files (the size of a block is the
square root of the size of the file for large files);
+ transmit less checksum information per block (we still have
less than one chance in a hundred million of transferring a
file incorrectly, and Unison will catch any transfer error
when fingerprinting the whole file)
+ avoid transfer overhead (which was 4 bytes per block)
For a 1G file, the first optimization saves a factor 50 on the
amount of data transferred from the target to the source (blocks
are 32768 bytes rather than just 700 bytes). The two other
optimizations save another factor of 2 (from 24 bytes per block
down to 10).
* Implemented an on-disk file fingerprint cache to speed-up update
detection after a crash: this way, Unison does not have do
recompute all the file fingerprints from scratch.
+ When Unison detects that the archive case-sensitivity mode
does not match the current settings, it populates the
fingerprint cache using the archive contents. This way,
changing the case-sensitivity mode should be reasonably fast.
* New preferences "noupdate=root", "nodeletion=root",
"nocreation=root" that prevent Unison from performing files
updates, deletions or creations on the given root. Also 'partial'
versions of 'noupdate', 'nodeletion' and 'nocreation'
* Limit the number of simultaneous external copy program ("copymax"
* New "links" preference. When set to false, Unison will report an
error on symlinks during update detection. (This is the default
when one host is running Windows but not Cygwin.) This is better
than failing during propagation.
* Added a preference "halfduplex" to force half-duplex communication
with the server. This may be useful on unreliable links (as a more
efficient alternative to "maxthreads = 1").
* Renamed preference "pretendwin" to "ignoreinodenumbers" (an alias
is kept for backwards compatibility).
* Ignore one-second differences when synchronizing modification
time. (Technically, this is an incompatible archive format change,
but it is backward compatible. To trigger a problem, a user would
have to synchronize modification times on a filesystem with a
two-second granularity and then downgrade to a previous version of
Unison, which does not work well in such a case. Thus, it does not
seem worthwhile to increment the archive format number, which
would impact all users.)
* Do not keep many files simultaneously opened anymore when the
rsync algorithm is in use.
* Add "ignorearchives" preference to ignore existing archives (to
avoid forcing users to delete them manually, in situations where
one archive has gotten deleted or corrupted).
* Added "BelowPath" patterns, that match a path as well as all paths
below (convenient to use with nodeletion,update,creationpartial
* Added a "fat" preference that makes Unison use the right options
when one of the replicas is on a FAT filesystem.
* Allow "prefer/force=newer" even when not synchronizing
modification times. (The reconciler will not be aware of the
modification time of unchanged files, so the synchronization
choices of Unison can be different from when "times=true", but the
behavior remains sane: changed files with the most recent
modification time will be propagated.)
* Mac OS
+ fixed rsync bug which could result in an "index out of
bounds" error when transferring resource forks.
+ Fixed bug which made Unison ignore finder information and
resource fork when compiled to 64bit on Mac OSX.
+ should now be 64 bit clean (the Growl framework is not up to
+ Made the bridge between Objective C and Ocaml code GC
friendly (it was allocating ML values and putting them in an
array which was not registered with the GC)
+ use darker grey arrows (patch contributed by Eric Y. Kow)
* GTK user interface
+ assistant for creating profiles
+ profile editor
+ pop up a summary window when the replicas are not fully
synchronized after transport
+ display estimated remaining time and transfer rate on the
+ allow simultaneous selection of several items
+ Do not reload the preference file before a new update
detection if it is unchanged
+ disabled scrolling to the first unfinished item during
transport. It goes way too fast when lot of small files are
synchronized, and it makes it impossible to browse the file
list during transport.
+ take into account the "height" preference again
+ the internal list of selected reconciler item was not always
in sync with what was displayed (GTK bug?); workaround
+ Do not display "Looking for change" messages during
propagation (when checking the targe is unchanged) but only
during update detection
+ Apply patch to fix some crashes in the OSX GUI, thanks to
* Text UI
+ During update detection, display status by updating a single
line rather than generating a new line of output every so
often. Should be less confusing.
+ Fastcheck is now the default under Windows. People mostly use
NTFS nowadays and the Unicode API provides an equivalent to
inode numbers for this filesystem.
+ Only use long UNC path for accessing replicas (as '..' is not
handled with this format of paths, but can be useful)
+ Windows text UI: now put the console into UTF-8 output mode.
This is the right thing to do when in Unicode mode, and is no
worse than what we had previously otherwise (the console use
some esoteric encoding by default). This only works when
using a Unicode font instead of the default raster font.
+ Don't get the home directory from environment variable HOME
under Windows (except for Cygwin binaries): we don't want the
behavior of Unison to depends on whether it is run from a
Cygwin shell (where HOME is set) or in any other way (where
HOME is usually not set).
* Miscellaneous fixes and improvements
+ Compare filenames up to decomposition in case sensitive mode
when one host is running MacOSX and the unicode preference is
set to true.
+ Rsync: somewhat faster compressor
+ Made a server waiting on a socket more resilient to
unexpected lost connections from the client.
+ Small patch to property setting code suggested by Ulrich
+ Several fixes to the change transfer functions (both the
internal ones and external transfers using rsync). In
particular, limit the number of simultaneous transfer using
an rsync (as the rsync algorithm can use a large amount of
memory when processing huge files)
+ Keep track of which file contents are being transferred, and
delay the transfer of a file when another file with the same
contents is currently being transferred. This way, the second
transfer can be skipped and replaced by a local copy.
+ Experimental update detection optimization: do not read the
contents of unchanged directories
+ When a file transfer fails, turn off fastcheck for this file
on the next sync.
+ Fixed bug with case insensitive mode on a case sensitive
o if file "a/a" is created on one replica and directory
"A" is created on the other, the file failed to be
synchronized the first time Unison is run afterwards, as
Unison uses the wrong path "a/a" (if Unison is run
again, the directories are in the archive, so the right
path is used);
o if file "a" appears on one replica and file "A" appears
on the other with different contents, Unison was unable
to synchronize them.
+ Improved error reporting when the destination is updated
during synchronization: Unison now tells which file has been
updated, and how.
+ Limit the length of temporary file names
+ Case sensitivity information put in the archive (in a
backward compatible way) and checked when the archive is
+ Got rid of the 16mb marshalling limit by marshalling to a
+ Resume copy of partially transferred files.
Changes since 2.31:
* Small user interface changes
+ Small change to text UI "scanning..." messages, to print just
directories (hopefully making it clearer that individual
files are not necessarily being fingerprinted).
* Minor fixes and improvements:
+ Ignore one hour differences when deciding whether a file may
have been updated. This avoids slow update detection after
daylight saving time changes under Windows. This makes Unison
slightly more likely to miss an update, but it should be safe
+ Fix a small bug that was affecting mainly windows users. We
need to commit the archives at the end of the sync even if
there are no updates to propagate because some files (in
fact, if we've just switched to DST on windows, a LOT of
files) might have new modtimes in the archive. (Changed the
text UI only. It's less clear where to change the GUI.)
+ Don't delete the temp file when a transfer fails due to a
fingerprint mismatch (so that we can have a look and see
why!) We've also added more debugging code togive more
informative error messages when we encounter the dreaded and
longstanding "assert failed during file transfer" bug
+ Incorrect paths ("path" directive) now result in an error
update item rather than a fatal error.
+ Create parent directories (with correct permissions) during
transport for paths which point to non-existent locations in
the destination replica.
Changes since 2.27:
* If Unison is interrupted during a directory transfer, it will now
leave the partially transferred directory intact in a temporary
location. (This maintains the invariant that new files/directories
are transferred either completely or not at all.) The next time
Unison is run, it will continue filling in this temporary
directory, skipping transferring files that it finds are already
* We've added experimental support for invoking an external file
transfer tool for whole-file copies instead of Unison's built-in
transfer protocol. Three new preferences have been added:
+ copyprog is a string giving the name (and command-line
switches, if needed) of an external program that can be used
to copy large files efficiently. By default, rsync is
invoked, but other tools such as scp can be used instead by
changing the value of this preference. (Although this is not
its primary purpose, rsync is actually a pretty fast way of
copying files that don't already exist on the receiving
host.) For files that do already exist on (but that have been
changed in one replica), Unison will always use its built-in
implementation of the rsync algorithm.
+ Added a "copyprogrest" preference, so that we can give
different command lines for invoking the external copy
utility depending on whether a partially transferred file
already exists or not. (Rsync doesn't seem to care about
this, but other utilities may.)
+ copythreshold is an integer (-1 by default), indicating above
what filesize (in megabytes) Unison should use the external
copying utility specified by copyprog. Specifying 0 will
cause ALL copies to use the external program; a negative
number will prevent any files from using it. (Default is -1.)
Thanks to Alan Schmitt for a huge amount of hacking and to an
anonymous sponsor for suggesting and underwriting this extension.
* Small improvements:
+ Added a new preference, dontchmod. By default, Unison uses
the chmod system call to set the permission bits of files
after it has copied them. But in some circumstances (and
under some operating systems), the chmod call always fails.
Setting this preference completely prevents Unison from ever
+ Don't ignore files that look like backup files if the
backuplocation preference is set to central
+ Shortened the names of several preferences. The old names are
also still supported, for backwards compatibility, but they
do not appear in the documentation.
+ Lots of little documentation tidying. (In particular,
preferences are separated into Basic and Advanced! This
should hopefully make Unison a little more approachable for
+ Unison can sometimes fail to transfer a file, giving the
unhelpful message "Destination updated during
synchronization" even though the file has not been changed.
This can be caused by programs that change either the file's
contents or the file's extended attributes without changing
its modification time. It's not clear what is the best fix
for this - it is not Unison's fault, but it makes Unison's
behavior puzzling - but at least Unison can be more helpful
about suggesting a workaround (running once with fastcheck
set to false). The failure message has been changed to give
+ Further improvements to the OS X GUI (thanks to Alan Schmitt
and Craig Federighi).
* Very preliminary support for triggering Unison from an external
filesystem-watching utility. The current implementation is very
simple, not efficient, and almost completely untested--not ready
for real users. But if someone wants to help improve it (e.g., by
writing a filesystem watcher for your favorite OS), please make
On the Unison side, the new behavior is very simple:
+ use the text UI
+ start Unison with the command-line flag "-repeat FOO", where
FOO is name of a file where Unison should look for
notifications of changes
+ when it starts up, Unison will read the whole contents of
this file (on both hosts), which should be a
newline-separated list of paths (relative to the root of the
synchronization) and synchronize just these paths, as if it
had been started with the "-path=xxx" option for each one of
+ when it finishes, it will sleep for a few seconds and then
examine the watchfile again; if anything has been added, it
will read the new paths, synchronize them, and go back to
+ that's it!
To use this to drive Unison "incrementally," just start it in this
mode and start up a tool (on each host) to watch for new changes
to the filesystem and append the appropriate paths to the
watchfile. Hopefully such tools should not be too hard to write.
* Bug fixes:
+ Fixed a bug that was causing new files to be created with
permissions 0x600 instead of using a reasonable default (like
0x644), if the 'perms' flag was set to 0. (Bug reported by
+ Follow maxthreads preference when transferring directories.