GSettings again

[Geert Janssens]

Op 24-02-13 16:01, John Ralls schreef:
> On Feb 24, 2013, at 2:08 AM, Geert Janssens <janssens-geert at telenet.be> wrote:
>
>> GnuCash generally has long release cycles. 2.0.0 to 2.2.0 was roughly 3 years. 2.2.0 to 2.4.0 about 4 years and before 2.6.0 will be release another 3 years will have passed.
>>
>> This got me thinking about GConf/GSettings again. Currently the migration is not planned for 2.6. Which means it will make it's appearance at best in 2.8. And that would mean we'd have to carry GConf around for at least another 3 years.
>>
>> It is worse actually. To maintain a good user experience we will have to make all efforts to migrate all user settings from GConf to GSettings the first time they run GnuCash with GSettings support. GConf (as of 2.31) luckily comes with a tool that can do this migration. Using that tool also means we still require GConf to be around as long as we expect users to want to migrate. As our policy is that we support migration from one major release series to the first next major series (from 2.2.x to 2.4.x, from 2.4.x to 2.6.x) that would mean we will need to keep GConf around at least one major release cycle next to GSettings. So if we only introduce GSettings in 2.8, we will only be able to drop GConf in 3.0/2.10. That means at least 6 more years looking at bugs like [1] that are inherently tied to GConf's internals.
>>
>> I'm not looking forward to that. Which inspired me to instead look at what would be required to migrate to GSettings  already in 2.6.
>>
>> Most part is pretty simple:
>> - Make sure there is a complete schema for all settings keys used. We were missing a couple, but I have already added those in my local repo.
>> - Convert the schema from GConf format to GSettings format. There is a tool for that. The end result probably needs a little tweaking, but most of the work is done automatically
>> - Use the GSettings api instead of the GConf one in code. This is pretty much straight-forward find and replace stuff.
>>
>> With that GnuCash is GSettings ready. The trickier part is converting the existing preferences from GConf to GSettings. Again there is infrastructure to help us.  There is a tool (part of GConf) that can do the migration for us. This tool needs a config file as input to know which keys from the old system match the keys in the new one.
>> Again this config file will be fairly easy to write and is well documented.
>>
>> On linux systems, that's where our job would end. The conversion tool is supposed to be run automatically by the desktop environment each time a user logs in. I haven't verified if this works, but it does look like it. I have updated my linux system several times since GSettings was introduced and I never had the feeling I lost any preferences in those migrations. Several programs did migrate to GSettings in this time frame.
>>
>> On to Windows and OS X. On those two platforms we don't have the support of the environment to take care of this for us. So we'll have to run the conversion tool ourselves at application launch time. The documentation states explicitly that it doesn't harm to run the tool more than once even. It keeps track by itself whether or not the conversion has happened In principle, again not that much work.
>>
>> On to the caveats:
>> - To use the migration tool, GConf must be 2.31 or more recent. That in itself does raise a few issues.
>> - The first being this is not available on current Debian stable (squeeze). It is on Debian testing. If they continue their trend of the last couple of releases the new stable (wheezy) should be around the corner while GnuCash 2.6 is only planned for November this year. So I don't expect Debian squeeze to backport GnuCash 2.6. But even if they do, it should either also backport GConf 2.31 or more recent or worst case the user's preferences won't be migrated. This would be inconvenient but not dramatical.
>> - Our Windows build currently ships GConf 2.22, the last version available in binary format on the gnome's ftp server. So we'd have to build a more recent version ourselves (which I'm currently working on).
>>
>> Other than these, I don't see any issues that would prevent this from being implemented rather sooner than later.
>>
>> So why do I bring all this up ? Mainly to discuss timing. We more or less aimed around this weekend to release our first 2.5 series development snapshot. I don't think I can finish the GSettings work this weekend. So we have three options:
>>
>> 1. postpone 2.5.0 a little more
>> 2. release 2.5.0 without GSettings and include it in 2.5.1 (which means an exception on the feature freeze)
>> 3. don't do GSettings at all in the 2.5/2.6 cycle (not my preference at all)
>>
>> Can I hear your opinions on this please ?
> There's a problem with GConf > 2.28 on OSX: It won't talk to dbus. I think that's due to the change from dbus-glib to
> gdbus and a bug in the latter, but I haven't been able to figure it out.
>
> On Windows for some reason GConf doesn't use dbus at all. I'd like to make OSX do that too, which would mean that I could use  a more recent GConf for the transition.
Well, a little further in my experiment I have learned that the Windows 
build is in a similar situation. It used 2.22 until now. Attempting to 
build 2.31 will pull in a dbus dependency. I don't feel like going that 
route just to be able to migrate from GConf to GSettings. So for now I 
skipped this part and am considering writing a simple script that does 
the same migration. Unlike on linux, GnuCash is the only user of the 
gconf db. So it won't harm to read the gconf files directly and feeding 
the contents to gsettings.
> But in line with reducing our dependency upon GLib, how about getting rid of both and using a plain ini file? Or if you think our schema is too deep for that, a single xml preferences file? Do that many Linux users change their Gnucash prefs from outside of Gnucash?
I see some benefit in sticking with a GLib based preferences solution 
for now.
- At some point the whole preferences system will need some careful 
attention. We have preferences littered in GConf, a meta resource file, 
in the data file,... And for some preferences not too much thought has 
gone into where they should be stored. But this is a large work. Not 
something to start a couple of days before feature freeze.
- On the other hand GConf is a nuisance in the GnuCash code. I think you 
will agree from your experience in the Quarz build.
- GSettings is very close to a drop in replacement for GConf. So it can 
be done in a short time frame.
- It has one major compelling advantage over GConf: it uses native 
backends on all platforms we care about. On linux that is dbus, on 
Windows that's the registry. On OS X I don't know exactly, but it's 
mentioned explicitly that there is a native backend on OS X as well. So 
no more orbit/dbus related misery on OS X and Windows.
- Further on, as a GLib component, it integrates well with Gtk and 
friends (direct binding of GSettings keys to GObject properties). I 
realize we talked about GLib reduction, but at the end of the discussion 
that was currently only aimed for at the engine level. The GUI will 
certainly stay Gtk for a while, so for me it makes sense to take 
advantage of the tool that go well with it.

In the future this may change if we effectively choose to move to 
another GUI toolkit or someone really takes on refactoring the 
preferences system. That is definitely not short-term though.

As a side note: our gconf logic is currently stored in core-utils, which 
at some point may have made some sense. If we want to get the engine 
GLib free, that should change. There is only one consumer of GConf data 
in GnuCash that's *not* in the GUI code: the xml backend. This can 
probably easily be factored out though. When that's done, the gconf code 
that's currently in core-utils can move to for example app-utils, which 
is our support library for all GUI modules.

Geert