Here we list planned and imagined improvements to Unison. Ones that we
regard as most important are marked with more *s. (Unfortunately, since
Unison is no longer under active development [though it is still heavily
used by its original developers], the presence of a suggestion in this file
is not promise that anybody is going to implement it!)
See the file BUGS.txt for a list of currently open bugs.
###########################################################################
SOON
====
**** Document: root, fspath, path (local/not)
**** Nice code cleanup trick: Add a phantom type param to Pref (and Pred?)
that prevents mutation from outside the module where the preference is
defined (by exposing it with a weak type).
**** The third assertion in Remote.fill_buffer failed for me (BCP) during a transfer!
**** Remaining problem with merging code:
- create two directories, each containing a .txt file
- sync so they are equal
- change the file so that one region is in conflict and another
region has changes that can be propagated correctly
- sync
- now we should be able to change the second region in just one file,
sync again, and see the change propagate; instead, it conflicts
- diagnosis: the merge stuff is not correctly updating the archive in
the event of a partial reconciliation
**** When deleting a directory, we should *not* skip over Unison temp files
in the process of listing children
*** make the ETA bar show which file is actually transferring bytes at the
moment
*** Fix the pred module to understand negation and delete XXXnot predicates
*** Web
- Add a "supported platforms" page mentioning system-specific stuff
- Add an installation instructions page, removing it from the manual
*** See if we can get rid of some Osx.XXX stuff (e.g. ressLength!?)
*** Add the following to the Problems FAQ:
--- In unison-hackers@y..., "Matt Swift" <swift@a...> wrote:
> I just posted a msg to cygwin@c... detailing some very strange
> behavior of chmod when a file's owner is also the file's group. It
I was right about the crucial circumstances of owner = group. Moral:
do not let user=group under Cygwin. I know it causes a problem when
you make unison use the full permissions model on Cygwin systems; I
think this may also explain similar problems I had using the default
unison behavior (which treats Cygwin files as read-only or read-write
only) -- though there are several possible causes of like failures to
syncrhonize permissions.
The answer is obvious, following from the basic handling of permissions
in Cygwin (in NT permissions mode), but I didn't see it. Users and
groups to Windows are the same kind of object (SID), and permissions on
a file or directory are represented as a list of (any number of) SIDs
paired with permissions such as read, write, execute (and quite a few
more). When you try to map this to the Unix model of user and group,
when the user and group happen to be the same, the user-permissions and
the group-permissions are operating on the same underlying Windows
object, and so they cannot be different. I think the user-permissions
prevail.
For example, if you try to sync a Unix file with permissions rw-r--r--
with a Cygwin file with permissions rw-rw-r-- whose owner happens to be
the same as the group, unison will report success, but the actual
permissions will not be changed. Moreover, during the next sync,
unison will by default propogate the Cygwin file back to the Unix file,
so that the degenerate permissions under Cygwin will migrate to the
Unix system unless you are careful to prevent unison from doing it.
(When you are trying to sync some 75,000 email and font files, this all
is more than a little exasperating!)
---
Further important advice if you are going to synchronize Cygwin
filesystems with unison's full Unix permissions model (and perhaps it
is also important even with unison's default behavior):
Background: the flags "ntsec" or "ntea" in the CYGWIN environment
variable signals Cygwin's libraries to use the richer NT permissions
model rather than a simplified Win95-98 model. "ntsec" requires an
NTFS filesystem, "ntea" will work with FAT filesystems. I use
"ntsec".
If unison does not have CYGWIN set appropriately in its environment,
some chmod calls will not do the expected thing, even though they
return with success. This will result in the file coming up again in
the next synchronization, and unison will then by default propagate the
(wrong) permissions from the Cygwin file back to the Unix system. (The
first chmod apparently succeeded, so unison records the new permissions
in its archive; the second time, when the file does not match the
archive, it seems to unison that the Cygwin file has been changed.)
If you run unison from the bash command line, you will most likely not
have a problem, since CYGWIN is probably set appropriately and exported
in the .bat script that launches bash. Likewise, when the Cygwin
filesystem is the remote one, Cygwin's sshd is by default set up (by
/usr/bin/ssh-host-config) to establish and export an appropriate value
of CYGWIN to ssh clients.
If you launch unison directly from a Windows shortcut, however, you
must set CYGWIN in your Windows environment variables. This is
certainly a convenient way to launch unison either with a particular
profile or generically. The instructions for setting up Cygwin and the
discussions of the CYGWIN envariable in the user manual never mention
any need to put CYGWIN in the Windows envariables, however. (I'm
writing them to suggest they do.)
>From the unison standpoint, the code which chooses to use the full
permissions model on Cygwin hosts (right now I have it hacked simply to
always use full permissions, by commenting out a line) perhaps ought to
confirm that "ntsec" or "ntea" is in the CYGWIN envariable and issue a
big warning that permissions may not be properly synchronized if
neither value is there.
** add '<meta http-equiv="expires" content="0">'
to the head section of all the unison web pages.
** Peter Selinger has built an SHA256 implementation that should be usable
as a drop-in replacement for MD5, if we ever need to do that
BUILDING AND INSTALLING
=======================
** 'make install' could be improved (and documented)
1. Typing "make install' after a "make" should simply install the
program that was made, not attempt to do a remake with different options.
===> Doesn't it???
2. "make install' should try to install as /usr/local/bin/unison, not
~/bin/, especially considering that ~/bin is the wrong place to do the
install under OSX (it should be ~/Apps or ~/Apps/bin)
** document the dynamically linked version, as some user already reported
that it works fine. Also, try to make the statistics window work with
this version. [This is "under windows," I think.]
should strip symbols from binary files in 'make exportnative'
DOCUMENTATION
=============
** Put a little more order on the flags and preferences -- e.g.,
organize them into "basic preferences", "advanced preferences,"
"expert preferences," etc. Requires hacking the Uarg module.
** Add something to docs about how to use 'rootalias'. Include an
explanation of the semantics, a couple of examples, and a suggestion
for how to debug what it's doing by turning on appropriate debugging
flags. (And maybe we should actually make the debug output there a
bit more verbose?)
** Misc:
- document good trick: use -1 switch to ssh if the paths are set up wrong
on the remote host
- should say whether trailing slashes are ok for paths; should say
that leading slashes are illegal.
===> check
- not so clear what you have to do with a Regex to match a directory
and all its subfiles: foo or foo/ or foo/.* ?
===> the first. document it. (Does foo/ match foo? I don't think so.
Document, one way or the other.)
- what happens when files are included whose parent dirs are
excluded? (With Regex? With multiple Path and Name?)
===> document
- the documentation is very good, but i couldn't find a description of how
to respond to the prompts in the textual ui. is that explained
somewhere? a few typos i noticed: "with t fast", "nison", "off of".
** what happens when we ssh through loopback and sync the same
directory?
===> Needs to be thought about. In particular, what is the name of the
archive in this case? Could they ever be exactly the same?
===> Try it and see.
SMALL FUNCTIONALITY IMPROVEMENTS
================================
**** The archive should indicate whether it is case-dependant or not.
(This is important for correctness -- if the case-insensitive flag is
set differently on different runs, things can get very confused!)
**** Use LargeFile (submodule of Unix) instead of standard file commands,
to avoid problems with huge files
DONE
*** Delete old backups mechanism and, instead, extend new one to cover its
functionality
- put backups in same dir as files by default
- otherwise, put them in a central place if one is given
- Update.incrVersionsOfBackups should not be externally visible
*** there's an HFS+ aware version of rsync called rsyncx. It should be
relatively easy to import that functionality into unison.
*** Consider altering the socket method, so the server accepts connections
only on a particular address? This would be very useful, because many people
tunnel unison over an OpenVPN Link, and this software works with virtual
devices and additional IP addresses on it. If unison would accept
connections only on the virtual device, the security would be enhanced,
because the OpenVPN key should be unavailable for the black hats.
*** unison -help doesn't go to stdout so it's hard to pipe it into less
===> Probably *all* output should go to stdout, not stderr (but maybe
we need a switch to recover the current behavior)
*** for the MSVC version of unison, we should deal with the nonstandard
semantics regarding read-only files.
===> What does that mean??
*** If a root resides on a `host' with an ever and unpredictably changing
host name (like a public login cluster with dozens of machines and a
shared file system), listing each possible host name for this root is
not feasible. The ability of specifing patterns in rootaliases would
help a lot in this case. I'm thinking of something like this:
rootalias = //.*//afs/cern.ch/user/n/nagya ->
//cern.ch//afs/cern.ch/user/n/nagya [NAGY Andras <nagya@inf.elte.hu>,
March 12]
===> We definitely ought to do something about this problem -- it's
increasingly common. Not sure if this is the right proposal, but
something.
*** Currently, if a file changes on either side between the initial update
detection and the time when the transport module tries to propagate
changes, the transport is aborted. But if the change occurred on the
replica that is being used as the source for the transfer (which will
be the common case!), then there is no reason to abort -- we should
just propagate the newest version.
*** When unison notices lock files in the archive directory, it should
offer to delete them *for* the user, rather than forcing the user to
delete them manually.
*** improve error reporting when Unison is started with different versions of
client and server
*** A switch to delete files before replication. It's not something I
would have considered doing, and in normal replication, there have
already been pointed out good reasons why Unison works the way it
does, but Roman makes a good reason for why this is useful in CD-RW
backups, and why this could be useful on a general to do list. And
this is certainly *generic*, which my point is not (as it only applies
to the Microsoft Windows NTFS situation).
*** A switch to include NTFS ACE/ACL file permissions to be copied when
copying from one NTFS location to another NTFS location. As I
mentioned this is less generic, but of fundamental usefullness in
Windows usage, as NTFS permissions are absolutely essential in many
backup/replication situations in Windows systems. Robocopy has the
/SEC switch, but Unison is a far better tool, and I was hoping in that
light that Unison could implement the rights/permissions stuff also.
*** There is no command-line argument to tell Unison where the .unison
directory is; Unison finds it in the environment or not at all. I was
able to workaround this with a symbolic link to put .unison where it was
expected, but it seems like an easy option to add.
*** The other is possibly a bit more difficult, but more useful as well. There
is a brief window of vulnerability between when the local server is started
and when the remote client connects to it. (It's no longer than that
because Unison won't take more than one connection at a time.) I can
tolerate it, but the window could be eliminated entirely by allowing socket
connections to require a nonce.
** Would be nice to transfer directories "incrementally" rather than
atomically (i.e., if Unison is interrupted during the transfer of a
directory, the partially-transferred directory should persist). Is
this allowed by the specification? (If so, then it should just become
the default behavior.)
===> BCP and William Lovas have discussed how to do this, but it is
not all that straightforward.
** we should reload the current preference file (if it's changed, at least)
when we restart
** [A good idea for the ssh prompt issue...] I'm not sure why you would
need a C implementation; you could do the same thing in CAML that expect
does: allocate a PTY, start up ssh on that, and interact with it. On
Windows, you can probably do the same with the Win32 console API,
although I don't see why such an improvement needs to work uniformly
across all platforms to be useful. [Note that allocating PTYs is not
very portable, but we could at least try allocating one and see if
something useful comes back...]
** An idea for the interface to the external merge functionality:
created a general mechanism for invoking external functionality...
- in profile, declare a command of the form
key M = external "merge ##1 ##2 ###" --> overwriting originals
(concrete syntax open to discussion!). Main parts are
- what key to bind it to in the UI(s)
- the command line to start up
- variables (##1 and ##2) for the local and remote files
(the remote file will automatically be copied to a local temp
file, if this variable is used)
- a variable (###) for a temporary output file
- an indication of what to do with this output file
(or maybe this could be automatic)
- (should also indicate which machine(s) to run the command on?)
** small additions to merge functionality:
- if the external merge program *deletes* one of the files it is given,
Unison should interpret this as "Copy the other file onto this location
(instead of merging)". This will allow some other interesting
functionality, e.g. external programs that may decide to keep both
versions by moving one of them out of the way (mh-rename).
- the invocation of the external 'diff' program should be selectable
using the same conventions as the 'merge' program
- would be nice to be able to invoke DIFFERENT merge programs
depending on paths
** We should document other available merge tools, e.g.,
idiff [BCP has a copy of the code for idiff that Norman sent.]
** Allow 'default.prf' in place of 'default' for profile names
** [dlux@dlux.hu, Feb 2002] For some apps (e.g., some mail readers?),
putting temp files in the same directory as the file we're about to
overwrite is bad/dangerous. Some alternatives that we could
consider...
- Add a configuration option for temporary directory and notice the
user about the volume restrictions in the docs and then if the user
does not consider it, then we use a non-atomic (copy + unlink)
rename. In an ideal environment (where the user consider this
restriction), it makes possible to sync a maildir folder while it is
online!
- An even better solution: One more temporary file step. If the user
sets the temporary directory, then we synchronize the files to that
directory, and if the file is downloaded/uploaded fully, then we move
it to a tempfile into the target directory (with .unison.tmp
extension) and then rename it into the final name.
** Suggestion for extending merge functionality
- add a new kind of preference -- a conditional stringlist preference
- in the preference file, each value looks like either
prefname = string
or
prefname = string WHEN Path PPPPP
prefname = string WHEN Name XXXXX
prefname = string WHEN Regex XXXXX
- when we look up such a preference, we provide a current path, and it
returns the one that matches the current path, if any
** Would be good to (optionally) change the semantics of the "backup"
functionality, so that Unison would not insist on making a *full*
backup of the whole replica, but just do so lazily. (I.e., it would
not make backups when files get put into the archive, but only when
they actually get changed.)
** Would also be nice to allow the backup preference to be set
differently on different hosts -- so that all the backups could be
kept on one side (if there is no space on the other side, e.g.). The
obvious way to do this is to add a switch like '-suppressbackupsonroot
BLAH' but this feels a bit ad hoc. It would be nicer to decide, in
general, which preferences can sensibly have different settings on
different roots (e.g., the location of the archive dir, ...) and
provide a general mechanism for setting them per-host.
** ~/foo seems to work on the command line but not in root = ~/foo in the
config file.
--
Similarly: It seems that when one specifies logfile = foobar
in the preferences file, then unison assumes that it is relative to the
current directory. Since neither ~ nor $HOME are understood in the
preference file, this is an inconvenience, because it forces the user to
remember to run unison from the root directory.
===> Would be nice to support ~ internally
** giving a -path preference whose parent dir doesn't exist currently causes
Unison to abort with a fatal error. Would be better if it just
signalled an error for that file.
** no spec for escaping regexp chars; spaces? newlines? tabs? others?
mechanism for getting the list of files from another program (plugin)?
===> needs to be documented (look at rx.ml)
** seems not to recognise ignores when they are inside a path that has
just been added.
===> Jamey claims that if we add a new directory, some of whose children
are ignored, then when this new dir is propagated, also the ignored
stuff gets copied (if this is true, then it's probably a bug in
update.ml)
* When loading archives (not just when dumping them), one should check that
they have the same checksum.
* [July 2002, S. Garfinkel] Maybe we should turn the 'time' option on by
default. We might need to help people a little on the upgrading,
though. When you did a sync with time=false, then a sync with
time=true, you get a zillion conflicts...
==> This is probably a good idea, but I'm a little scared of all the
messages we'd get from upgrading users
* Maybe we should write debugging and tracing information to stdout
instead of stderr?
* URI pathname syntax
Why is the following command wrong?
unison -servercmd `which unison` /usr/local ssh://labrador/usr/local
It took me three tries and careful reading of the documentation to
figure it out. I don't have any good suggestions here, other than
that I think the whole issue of relative vs absolute pathnames needs
serious thought. I think the current interfaces do not work very
well. One possibility that I will float is that you invent a special
character string to refer to the root of synchronization.
E.g., interpret ~ as $HOME in roots.
--
Also: we should add the file:// syntax to URIs...
file://C:/Necula (C:/Necula on the local file system)
file:////share/subdir (//share/subdir as from the point of view of
the local file system)
unison://host///share/subdir
--
Should local roots in a profile be canonized?
Right now, we can have a relative root in the profile. This
is going to be a problem if unison is started in a different
directory.
* At the moment, if Unison is interrupted during a non-atomic operation
on the file system, the user has to clean things up manually, following
the instructions in the the recovery log. We should do that for them.
(This is actually a bit tricky, since we need to be careful about what
might happen if unison crashes during recovery, etc. The best way to
accomplish this would be to write a general logging/recovery facility
in OCaml.)
* Dealing with ACLs: Maybe this is what we should do actually. We could
specify a user (and similarly a group) to unison. It would be
interpreted in a special way: if a file is owned by this user, unison
will rather consider that the owner of the file is undefined. So, when
a file owned by an unkown user is synchronized, the file owner is set
to the default user. Then, on the next synchronizations, unison will
consider that the owner has not been propagated and try again. [Should
be easy once the reconciler is made more modular]
* The -terse preference should suppress more (in fact, almost all)
messages in the text ui. See Dale Worley's message for a detailed
proposal.
Make sure that no filesystem check is missing in the transport agent.
===> What does this mean?
Would be nice to have the Unison log file relative to my home directory,
like this
logfile = ~/.unision/log
or
logfile = $HOME/.unision/log
(We should do this for *all* files that the user specifies.)
It would be nice if Unison could have the "power" to copy write-protected
files, maybe as an option.
Update checking over NFS might be *much* faster if we use only relative
pathnames (absolute paths may require an RPC per level!?)
On one server (Saul), Unison seems to use HUGE amounts of memory (250Mb
resident), while on my laptop it's much less. WTF?
[Ben Wong, Aug 2002] Why not make unison fall back to addversionno if it
would otherwise bomb out with an incorrect version number? That way I
wouldn't have to educate people on how to use Unison at my site; it'd
"just work".
The -sortbysize is nice, but what I would really like is a -limitbysize.
When I'm connected over a modem line, I would like not to transfer the
larger files that need synchronization. That can wait until I am
connected via a faster connection. What I presently do is allow unison
to run in -sortbysize mode, and abort once I have all my little, more
important files. -limitbysize should simply filter the list of transfer
to only those that are below the threshold size. The syntax is
obvious... It should be -limitbysize xxx, where xxx is the size
(preferably in kb, but bytes will do as well).
Maybe we should use getcwd for canonizing roots under Unix. For some
systems (Linux, for instance), getcwd succeeds even when some parent
directory is not readable.
[From Yan Seiner]
Can unison modify the (*nix) environment to show the
ip/name/some_other_id of the system making the connection? This would
help tremendously.
For example, vtun does this:
---
root 6319 0.0 0.6 1984 852 ? S< Aug27 0:37 vtund[s]:
bgsludge tun tun10
root 6324 0.0 0.6 1984 852 ? S< Aug27 2:00 vtund[s]:
cardinal tun tun0
root 17001 0.0 0.6 1984 848 ? S< Aug27 0:05 vtund[s]:
wtseller tun tun11
root 20100 0.0 0.6 1984 852 ? S< Aug28 0:02 vtund[s]:
cardridg tun tun1
----
So I know I have four sessions, to each named machine, and I know
immediately who is connected and who is not. If I have to kill a
session, I don't kill the wrong one.
add a switch '-logerrors' that makes unison log error messages to a
separate file in addition to the standard logfile
Dale Worley's suggestion for relocating archives:
> You're right: it's not all that tricky. So would you be happy if you
> could run unison in a special mode like this
> unison -relocate //old-host1//path1 //old-host2//path2 \
> //new-host1//path1 //new-host2//path2
> (where all the hosts and paths are normalized) and it would move the
> archives for you on both machines?
Actually, I think that what you want is for the user to specify the
old paths in *normalized* form and the new paths in *non-normalized*
form. That is, unison uses the old paths literally as provided by the
user, but it applies the usual normalization algorithm to the new
paths.
This may sound strange, but I think that it's the Right Thing:
- There is no guarantee that the normalization algorithm, applied to
the old paths as the user used to specify them, normalizes to the
the normalized paths that are recorded in the archive. Indeed,
there may no longer be *any* path which normalizes to the recorded
paths.
- The user can extract the normalized old paths from the second line
of the archive files. This is clumsy, but reliable. And we don't
intend the user to relocate an archive very often.
- But for the new paths, you want to normalize what the user supplies,
because he doesn't know in advance how Unison is going to normalize
the new paths, and may well specify them incorrectly. That would
leave him with a relocated archive that he might not be able to use
at all.
You might want to put quotes around the pathnames in the second line
of the archive, since MS-Windows directory names can contain spaces,
etc.
For safety...
- Add a preference 'maxdelete' taking an integer parameter, default 100
(or perhaps even less -- keeping it fairly small will help naive users
avoid shooting themselves in the foot). A negative number means
skip this check (i.e., infinity).
- When the transport subsystem gets control (i.e., just after the user
says 'go' to the user interface, when not running in batch mode)
it first checks the number of files that are going to be deleted
(including all the contents of any directories that are marked for
deletion). If it is more than maxdelete (and maxdelete is
positive), then...
- If we're in batch mode (batch=true), we halt without doing
anything.
- If we're not in batch mode, we display a warning message and
make the user confirm. (If they do *not* confirm, it would be
nice to dump them back into the user interface again, but this
would require a little rewriting of our control flow.)
- Would also be nice to include a display in the UI someplace that says
how many files are to be deleted/changed/created plus how many bytes
to be transferred, and a warning signal (display in red or something)
if these exceed the current setting of maxdelete.
Might be nice to provide an option that says "if you're propagating a
newly created directory and something goes wrong with something inside
it, just ignore the file that failed and keep going with the rest of
the directory." [We probably don't want to continue in all cases (for
instance, when the disk is full)]
Would be nice to be able to run unison in a special mode like this
unison -relocate //old-host1//path1 //old-host2//path2 \
//new-host1//path1 //new-host2//path2
(where all the hosts and paths are canonized) and have it move the
archives for you on both machines?
It would be nice if unison had a tool by which it could regenerate all
the MD5 sums and compare them to what it has stored, then produce a list
of files that are different. I obviously cannot count on file size and
date in this case; those may not have changed but the contents may be
corrupt.
If the connection to the server goes away and then comes back up, it
would be nice if Unison would transparently re-establish it (at least,
when this makes sense!)
If we synchronize a path whose parent doesn't exist in one replica, we'll
fail. Might be nicer to create the parent path if needed.
maybe put backup files somewhere other than in the replica (e.g. in
$HOME/tmp, or controlled by preference)
Better documentation of the -backups flag, and a way to expire old backups
Add a preference that makes the reconciler ignore prefs-only differences
between files (not updating the archive, though -- just suppressing
the difference -- will this slow things down too much?? Maybe it needs
to happen in the update detector, before things are transmitted across
the network.)
Perhaps we should interpret both / and the local separator as path
separators, i.e., under Windows / and \, under Mac / and :, and under
Unix just /. For Windows this will be fine, since / is not allowed in
filenames.
Maybe have an option to tell do not transfer toto.dvi if toto.tex exists (or
toto.ps if toto.dvi): something like
Ignore .dvi If .tex
===> This is not a good idea -- would give different ignore results on
the two machines. But maybe a variant would work:
- Have an option to execute a command if a given file exist like
Execute rm core If core
Execute make clean If Makefile
We should put in a preference that forces Unison to do really safe update
detection (with fingerprinting), even on Unix systems. (Maybe just for
some paths?)
Maybe we should never emit a conflict for modtimes; instead, we just
propagate the largest one.
[John Langford] Some code for (at least partially) handling large files
can be found in 64bit_ops.c in:
http://www-2.cs.cmu.edu/~jcl/programs/sync_file.tar.gz
Make sure you pay attention to the compile line as it is important.
USER INTERFACE
==============
** In menu Actions
- show Diff applies to the current line, while
- revert to unision's recommandation applies to all lines
Should be clearer and/or homogeneous behavior.
I would also like to have "revert to unision's recommandation" for the
current line.
** in gtk ui, display green checkmark next to finished items even if their
direction indicates a conflict; do not list such items as "skipped" at
the end
** In both UIs, show how many bytes/files were successfully transferred
at the end
** Should support auto-termination of the graphical UI (switch-controlled)
* Unison starts in the usual way and checks for changes
* If there are no conflicts, it proceeds without waiting for confirmation
* If there *are* conflicts, it waits for instructions, just like now
* In either case, when it's finished transferring the changes, it quits
* [Matthew Swift] in the GTK gui at least,
display the total MB or #files or whatever it is that the ticking
%-meter is referring to when it goes from 0 to 100. it is useful to
know how big the xfer is going to be before starting it (might induce me
to choose "sort by size", or abandon and choose a smaller subset, etc.).
Also, esp. since the gui is single-threaded and unresponsive, i would
like to know what size of a synch that I am for example 50% or 22%
through. I know that an ETA and other things we're used to from many
downloading apps would require quite a bit of code, but it would help a
lot just to display whatever constant is represented by 100%.
* [BCP] Error reporting for per-file problems during updating leaves
something to be desired. In particular, there's no indication even of
which host the problem occurred on. (I added something that includes
"root 1" or "root 2", but I'm not sure that's better than nothing.) If
there are errors on both hosts, only one will be reported. If there
are lots of errors in a subdir, only the first will be reported.
Recon.propagateUpdates would be a starting point for changes.
* [Jamey Leifer]
Would be nice if both UIs had a "revert to Unison's proposal" button...
* [Jamey Leifer]
[graphic ui, wishlist] The documentation topics aren't searchable. As
a result "unison -doc running | less" is still indispensable if one
wants to find anything. I suggest adding a box
"search in this topic: ---"
which is always available in the doc viewer. It would be nice to
support keyboard shortcuts in the "less" style, namely "/", "n", and
"N" (i.e. search, next, previous) to avoid too much clicking.
[graphic ui, wishlist] Ditto as far as searchability for diff reports.
* Would be nice to have a keystroke in the UI that means 'add the current
directory to the set of ignore patterns.'
* In the text UI, during the transport phase, print each file being
transferred on *one* line, with an arrow to indicate which way (and
dropping the explicit indication of which host from and to). The
logfile should be more explicit.
* The unison gui currently displays a percentage completion in the lower right
corner. I would find it comforting if it would also display an effective
bandwidth there, i.e., how many bits per second are flowing through the
transport layer? I make this request because owing to a hardware
catastrophe, I have just started using Unison through the phone lines, and
it seems to do nothing for a long period of time. I don't know whether
to blame the cheap modem, the cheap ISP, or whether Unison simply isn't
telling me that bits are flowing through the wire. (netstat -tn
suggests not much is happening, but I don't know if the results can
be trusted.)
* Would it be hard to add "tool tips" to the buttons in the UI?
==> Look for "tooltip" in examples/testgtk.ml.
The easiest way is with a toolbar, but you can also add tooltips to any
widget (cf lines 867 and after).
* > On a line, I would like to have a description of the action to be taken in
> clear words: (e.g. will erase file on local or will copy from local to
> remote, etc.)
This might be a good use for "tool tips," if I knew how to make them work
using lablGTK.
* After clicking "Create new profile" in the initial profile window and
giving a name for the new profile, it is confusing to get dumped back
into the profile window again and have to explicitly select the new
profile. Would be better to skip this step and go straight into
filling in its fields.
* Another usability issue in the text UI: , and < should mean the same to
unison. It would be nice if both had the same representation on-screen
(ie, show a "<" even if I typed a ","). Similarly for . and >.
* The menu help for left/right arrow both said `transfer local to local'.
Not helpful. The items in question are pathnames, which you might not
have to abbreviate. To save space one might consider replacing any
common prefix, and also short prefixes that look like they might be
automounter goo, with an ellipsis. Then show, e.g., 20 chars. I'd
also be willing to name paths in my profile, e.g., replica flatcoat =
/home/cellar/nr replica cellar = /m/cellar60/nr This would be
especially attractive if my short names were meaningful on the command
line.
* In the GTK user interface, it would be nice to be able to put up a window
displaying the contents of the log file (and add log messages to it
dynamically as we're working). Be careful, though: the log could get
large and we don't want this to be too slow.
* Could there be an option between -ui text and -ui graphic that when combine
with -batch and -auto would start in text mode, but pop up an interactive
graphic window when real conflicts happens.
* [Jamey Leifer] I think "unison -doc" should be mapped to "unison
-doc topics" and the error message for the former eliminated.
* [Jamey Leifer] Typing "unison" results in the Profiles box
("Select an existing profile..."). I think the help topics should be
available here.
* [Jamey Leifer] The file list is confusing since the paths
are sometime relative to the root and sometimes relative to the
previous path:
Mail/drafts/3
inbox/5538
5539
5540
I now understand that the indentation is significant, but it's not
that clear. A further confusion is that there's varying amounts of
indentation depending on the depth of the enclosing path:
foo/1
2
boo/goo/loo/1
3
4
This is really hard to parse since the fonts are variable width.
I would prefer to read the former as:
Mail/drafts/3
inbox/5538
5539
5540
(with the indentation actually showing the relationship) though this
may take too much horizontal space. Alternatively, one could choose a
Windows-style display:
|-Mail/drafts/3
|-inbox/5538
|- 5539
|- 5540
Unison's gui offers an `Actions' menu with a variety of features
regarding preferences. I would love to see an action with the following
semantics: if the two files differ only in their modification time,
prefer the older modification time.
===> This would be easy to add, but I am beginning to worry that we are
getting too many funny little switches like this. We should think
about them all together and make sure they make sense.
I'm watching it sync a very large file that I don't want anyway, and I'm in
a hurry. I'd like a way to say "forget that file, I don't care about it, go
on to the next one you have to sync". Doesn't sound hard...?
[Perdita Stevens, Perdita.Stevens@dcs.ed.ac.uk, Mar 14 2002]
===> It's not trivial (involves some subtle stuff about our RPC
implementation and the single-thread nature of the GUI), but might
not be impossible either.
"Quit" during synchronization should abort all current operations (so
that temporary files are deleted) before exiting.
===> Again, requires some careful thinking about how this would work
with the RPC layer.
It would be nice to have a command in the GUI that would allow a single
path within the replica to be selected from a file dialog and
synchronized.
The scroll bar is not usable during transport: every time a line changes
in the list, the display jumps to that line; if many small files are
transfered, it makes browsing in the list quite impossible...
[From Manuel Serrano] Would be nice to put the arrows in different
directions in different colors, so that, e.g., you could quickly scan the
list of changes and make sure that they are all in the same direction
===> We tried this, but we couldn't find color combinations that did not
seem confusing. (Two different shades of green? Three? ...) If we
really want this, probably the best is to put in some preferences for the
user to control the colors of all the arrows individually.
Under Windows, convert filename to Unicode before printing them.
Text mode user interface should be brought up to date with graphical
interface (it should prompt for profile selection, creation, root
entry, etc.; command characters should be the same; ...)
Since the manual is pretty big, it would be nice if the on-line version
were accessible through cascading menus, allowing direct access to
individual subsections. It would also be nice if it were formatted a
bit more attractively, using proportional-width fonts, etc. (Does GTK
have something like an RTF widget?)
If I have a change I look at the detail window. It would be nice to be
able to click on one of the lines there instead of pressing one of <-
or ->. For one thing in the detail window the relative position of the
two files is up and down and translating that to <- or -> is somewhat
unintuitive.
Also, it would be nice to highlight in the detailed window the
elements that have changed.
Make it possible to select a bunch of conflicts at the same time and
override them all together
The UI window should display the current roots somewhere.
There should be a -geometry command-line interface, following the usual X
conventions.
put in a command-line option that makes fatal errors exit right away
without displaying anything in the graphical UI (for debugging)
Use the CTree widget to display the list of files
Add the ability to close and open directories in the UI.
it would be nice to give a visual indication of which files are
particularly big, so that the user can tell where the transfer
operations may get slowed down. Maybe a "size bar" showing the log
of the size (perhaps also color coded).
===> less urgent now because we can re-sort the update items by size
Would it be hard to allow long-running transfers to be aborted?
For instance, the key "/" aborts the transmission of the selected file
OR:
Allow the user to terminate individual operations by clicking a
"cancel" button. (This is not completely straightforward because
the whole program is single-threaded. But it should be possible for
the low-level transport code in remote.ml to realize that the
operation has been aborted, clean up, and raise an exception.)
It would be nice if the initial 'usage' message were not so long. Maybe
we could split options into 'novice' and 'expert' ones, and only print
the novice ones (with an indication how to obtain the full expert
printout).
> Show diff should behave as an emacs view-mode buffer and quit on a single
> 'q' in the window, or better quit even without focus be sent to the diff
> window...
The UI for the diff functionality needs some polishing. (Also, it should
be merged with the new "merge" functionality.)
consider separating switches into 'ordinary' and 'expert' categories,
documented in separate sections
would be nice to be able to "Proceed" just the selected line
might be nice if the GUI would beep when finished syncing (needs to be
switch-selectable and off by default, naturally). Is this easy with
LablGTK?
It would be nice to be able to STOP the GUI in the middle of propagating
changes.
TIDYING
=======
* Go through the sources and make all fatal and transient error messages
as informative as possible
More documentation (especially in the interface files) is always nice.
In particular, there isn't enough documentation of the big picture.
It isn't clear how to fit together archives, servers, paths, roots,
update detection, reconciliation, conflict resolution, or the user
interface...
Ocamlexc v1.0, the uncaught exceptions analyzer for Objective Caml is now
available from Pessaux's home page. It would be fun to run it over the
Unison sources and see if it reveals any problems.
LARGER EXTENSIONS
=================
Fast update checking would be cool... Some resources:
FAM (used in Enlightenment)
dnotify (linux 2.4)
BSD kqueue
the "VFS stacking layer" implemented by a guy at Columbia
[From JMS]
Some update detection speed improvement suggestions:
- Read the FFS (Fast Filesystem) paper for hints
- change the working directory instead of using absolute paths; this
avoids calls to the evil iname(?) facility in the kernel
- work breadth-first instead of depth first, to keep things in the
kernel cache
Rewrite recon.ml in a more modular way. Probably, have for each property
a function taking the previous file state and the state on each
replicas, and returning in what the synchronization operation should be
(nothing, left, right, conflict); a combinator then merge the results.
It would be good to have a graphical interface allowing management and
editing of profiles, ignore patterns, etc. Or, less ambitiously, just
have UI options for all command-line options (killServer)
How about a facility so that you can specify more than one pair of
file systems for a single invocation of Unison? This would be like
calling Unison multiple times, except that it would ask all the
questions at once. Better yet, we could actually deal with the
multi-replica case. (The latter is pretty hard.)
What about invoking some user-specified operation on each file as it
is transferred? Or in each directory where things have changed?
(This will require some careful design work.)
Sync with archived directories (in tar / zip / gz format) would be
nice. Seems a bit awkward to implement, though: at the moment there
are a lot of functions all over the place that investigate and
modify the file system, and these would all have to be replaced with
a layer that transparently parses, etc., etc.
Consider using other authentication services (e.g. Kerberos) instead
of / in addition to ssh.
What happens when we synchronize, then decide to ignore some existing file
What happens to the entry in the archive? If mirroring, it may be
large, we probably want to delete it from the archive.
File level synchronization (bookmarks, mailboxes)
It might be nice to implement an (optional) safety check that detects
aliasing within a replica due to followed links (or hard links) and
complains if it finds any. This should not be *too* expensive, since
we already know all the inode numbers. (Even if it *is* expensive, it
might be useful to allow users to do this occasionally, if they are
paranoid.)
WINDOWS ISSUES
==============
Suggestion from Arnaud:
I have been using XP for a while and despite all the problems I have, there
is a very nice feature: being able to mount remote folders (nothing new), to
work with them offline and synchronize them. Really useful.
--
A good way to simulate this with Unison would be to package it as a shell
extension. From the desktop by clicking on the right button the user selects
"create new Unison mount point" and answers a few trivial question. And the
rest is done in the background. There are a lot of examples of shell
extensions and there is a really good book for O'Reilly about it.
--
A good project for a student :-)
--
PS: see http://www.simplythebest.net/shellenh.html for some examples.
when typing ctrl-c in windows (dos-window in win98SE) when
unison is asking for conflicting updates there araises following
message (sorry for my bad translation to english):
"This program is closes because of a non-valid action. Contact the
manufactura if the error remains".
NTFS seems to have two ways of setting a file read-only!
Comments from Karl Moerder:
Tonight I made some files read-only on my desktop at home. I did this by
setting global read and execute permissions (from the security tab of
properties). I ran Unison and it didn't notice the change. I then set
the permissions back to full control and then selected the read-only box
(from the general tab of properties). I ran Unison again and it noticed
and pushed the perms change to the server.
I understand that Windows is a bit squirrely here, but how do you decide
which permissions to look at? It seems like perhaps the ones on the
security tab would be more natural. (?)
--
I get similar results with both bits (they both cause read-only
behavior).
I believe that the origin of the two modes of setting is that the first
set is the old way of doing Windows protection (probably the interface
provided on FAT file systems) and the new way is the more Unix like way
(added for NTFS file systems). The new way has rwxdpo bits for each
group (and there can be several groups).
