Review Request 116461: KConfigSkeleton: avoid calling reparseConfiguration() immediately after creation.

[Matthew John Dawson]

> On Feb. 28, 2014, 3:41 p.m., Matthew John Dawson wrote:
> > While I'm fine with the idea behind this optimization, I worry that this implementation could create situations were a configuration change is not picked up by the system.  For instance, what happens if the user doesn't immediately call readConfig?  This could create some subtle bugs in downstream code.
> > 
> > I had two ideas for how this optimization could be implemented:
> > 1) Lazy load the KConfig object the first time it is used.  Then, in readConfig, the call to be reparseConfiguration could be avoided if the KConfig is created due to its call.  This would retain the current behaviour, while ensuring readConfig reads in the most configuration.  Other uses of the KConfig will have to ensure the KConfig object has already been created, and if the user calls one of those functions before readConfig, they will still double read the configuration.  But since this is just status quo, I'm not too worried.
> > 2) Alternately, create a set of construction functions, like make_shared, that imitate the creation of a KConfigSkeleton subclass, and then reading the configuration through readConfig.  Internally, it can use a private readConfig function to ensure the configuration is no re-read.  This would require changes to applications to avoid the extra configuration call, unfortunately.
> > 
> > I saw RR #115964, and I assume that some of the reductions to the readings of oxygenrc are caused by the sharing the KConfig between some KConfigSkeleton's?  If so, I'd suggest implementing solution 1 for the general case, and then making a special construction function to handle shared KConfig's.  I don't want to avoid calling reparseConfiguration without some warning around this, as it may again cause some surprises.  A new appropriately named function shouldn't be too bad though, as opposed to changing the behaviour of the constructor.
> 
> David Faure wrote:
>     I've been thinking about this too, but what good is a KConfigSkeleton if you don't call readSettings() on it? You can't read any settings from it then, so all you can do is a) keep it around for later or b) use it purely for writing out a new config file. Use case b) is no problem, so I think we're talking about use case a). Yes in theory an app could see a behavior change in that the config file is loaded from disk at the time of creating the skeleton rather than at the time of calling readConfig the first time. But this is why I'm making this change in 5.0 and not in 4.13. I think it's an acceptable behavior (matching KConfig's behavior more closely - it parses in the ctor, not in readEntry) and I doubt it affects many apps, since all I see everywhere is singletons - i.e. the KConfigSkeleton is created on demand at the time where it's first needed, therefore the ctor is immediately followed by readConfig.
>     
>     My alternative idea (let's call it "3" to complete your list) was to pass a flag to the KConfig constructor to tell it "don't parse now", and setting that flag from the KConfigSkeleton constructor. Then readConfig can keep always calling reparseConfiguration(). This would work, right?
>     
>     Your suggestion 1 is somewhat equivalent, but since one of the ctors for KCoreConfigSkeleton takes a KSharedConfig::Ptr as input, it's not applicable, we can't delay the creation of the kconfig within KCoreConfigSkeleton since it's created beforehand by the application.
>     Your suggestion 2 requires changing all the apps, which lowers the benefits of the optimization.

I agree, it is a weird use case, and the software should probably be adjusted.  However, if an app does rely on that, it is very hard for the author of the software to notice the change, even with the port to 5.  If I just looked at the functions names, I'd expect readConfig to read the file all the time.  Following the principle of least surprise, I'd like to avoid readConfig ever not reading the file.

I'm fine with your alternate idea.  I prefer that over my first idea, as it effectively does the same thing while being less invasive.

For my second suggestion, I realize its downsides.  I just like following the principle of least surprise.  If your alternate idea is implemented, I believe that would cover most cases.  Suggestion 2 can then be implemented, and its related constructor could be marked deprecated.  This would allow for existing programs to continue working, while allowing developers to change their code to take advantage of the optimization.

As I stated earlier, I'm not sure about who KDE wants to handle source compatibility with kdelibs.  I also wouldn't mind just removing the second constructor, forcing all users to upgrade their code.  Since the function is a drop in replacement, it wouldn't be that hard for developers to upgrade.


- Matthew John


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/116461/#review51374
-----------------------------------------------------------


On Feb. 27, 2014, 4:18 p.m., David Faure wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://git.reviewboard.kde.org/r/116461/
> -----------------------------------------------------------
> 
> (Updated Feb. 27, 2014, 4:18 p.m.)
> 
> 
> Review request for KDE Frameworks and Matthew John Dawson.
> 
> 
> Repository: kconfig
> 
> 
> Description
> -------
> 
> KConfigSkeleton: avoid calling reparseConfiguration() immediately after creation.
> 
> KConfig already parses the config files from disk in the constructor,
> which is necessary for non-KConfigXT users. However when using KConfigXT
> the first thing one has to do after creation is to call readConfig(),
> which should therefore not call reparseConfiguration the first time.
> 
> strace -e open kate |& grep -v NOENT | grep oxygenrc | wc -l
>   went from 4 to 1 --> bingo, goal reached!
>   (and when looking for kdeglobals, from 10 to 7)
> 
> 
> Diffs
> -----
> 
>   src/core/kcoreconfigskeleton.cpp 9c5fb4a80d500e81b483b749a137ad5f2c99a55f 
>   src/core/kcoreconfigskeleton_p.h 0b020ed3493186e699d872ddc7a9f9294d797a95 
> 
> Diff: https://git.reviewboard.kde.org/r/116461/diff/
> 
> 
> Testing
> -------
> 
> (see commit log) + unittests in kconfig still pass.
> 
> 
> Thanks,
> 
> David Faure
> 
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/kde-frameworks-devel/attachments/20140301/b4416206/attachment-0001.html>