threads / discuss / 36120

[GSOC] Git Configuration API improvements

Subject: [GSOC] Git Configuration API improvements

## tl;dr

5 messages between Mar 11, 2014 and Mar 15, 2014.

replies: 4people: 3as markdown or json

karthik nayak· Mar 11, 2014, 13:12 UTC · lore

Hello to all,

I'm Karthik Nayak, a Computer Science student from Bangalore India. I will be applying for GSOC 2014. This is my first time applying for GSOC. I have been using Git for a long time now, so it would be an ideal organisation for me to contribute to. As suggested I completed the Micro Project under the guidance of Eric Sunshine.[1]

I went through the Idea's page[2] and the mails corresponding to the topic[3]. Currently we have multiple invocation of git_config() in an individual invocation of git() which is not efficient. Also, it is hard to implement new features, as mentioned -- such as allowing a configuration to be unset.

The basic idea is to use a data structure to store the config, the first time the git_config() is called. And refer to this each and every time we invoke git_config() further. Jeff suggested [4] to use a tree or a mapping of keys to values which are stored in a string. I think that the idea of storing using the config as a tree data structure would be really advantageous.

This would allow us to easily implement modifications later, and can help to easily take care of duplicates being created on multiple usage of config set and unset. As, whenever a node's children have been deleted the node can also be automatically deleted. The tree can be a struct with values, header link and link to other configs. This way we can also create functions to work on the tree easily.

Would be great to hear your thoughts on this topic and also I plan to start creating a proposal. Nice to have any suggestions or feedback of any kind.

Thank you, Karthik

[1] : http://article.gmane.org/gmane.comp.version-control.git/243695/match=karthik+188+gmail+com [2] : http://git.github.io/SoC-2014-Ideas.html [3] : http://article.gmane.org/gmane.comp.version-control.git/243500/match=git+configuration+caching [4] : http://article.gmane.org/gmane.comp.version-control.git/243542

Matthieu Moy· Mar 11, 2014, 14:51 UTC · re: karthik nayak · lore

Re: [GSOC] Git Configuration API improvements

karthik nayak <karthik.188@gmail.com> writes:

> Currently we have multiple invocation of git_config() in an
> individual invocation of git() which is not efficient. Also, it is
> hard to implement new features,

I think efficiency is not a real concern here. Config files are small and easy to parse, so parsing them multiple times isn't really noticeable from the performance point of view.

OTOH, the extensibility is a real concern, and that should be the main motivation for the project.

-- 
Matthieu Moy
http://www-verimag.imag.fr/~moy/

karthik nayak· Mar 11, 2014, 16:19 UTC · re: Matthieu Moy · lore

Re: [GSOC] Git Configuration API improvements

On Tue, Mar 11, 2014 at 8:21 PM, Matthieu Moy <Matthieu.Moy@grenoble-inp.fr> wrote:

Show 16 quoted lines

> karthik nayak <karthik.188@gmail.com> writes:
>
>> Currently we have multiple invocation of git_config() in an
>> individual invocation of git() which is not efficient. Also, it is
>> hard to implement new features,
>
> I think efficiency is not a real concern here. Config files are small
> and easy to parse, so parsing them multiple times isn't really
> noticeable from the performance point of view.
>
> OTOH, the extensibility is a real concern, and that should be the main
> motivation for the project.
>
> --
> Matthieu Moy
> http://www-verimag.imag.fr/~moy/

Hello Matthieu,

Thanks. I understand what you mean. extensibility is the main motivation of the project, i think that by implementing the cache system we can fix the small problems (reappearing of headers while setting and unsetting configs) and also implement new features like to unset a config easily.

Jeff King· Mar 14, 2014, 04:31 UTC · re: karthik nayak · lore

Re: [GSOC] Git Configuration API improvements

On Tue, Mar 11, 2014 at 09:49:33PM +0530, karthik nayak wrote:

Show 21 quoted lines

> On Tue, Mar 11, 2014 at 8:21 PM, Matthieu Moy
> <Matthieu.Moy@grenoble-inp.fr> wrote:
> > karthik nayak <karthik.188@gmail.com> writes:
> >
> >> Currently we have multiple invocation of git_config() in an
> >> individual invocation of git() which is not efficient. Also, it is
> >> hard to implement new features,
> >
> > I think efficiency is not a real concern here. Config files are small
> > and easy to parse, so parsing them multiple times isn't really
> > noticeable from the performance point of view.
> >
> > OTOH, the extensibility is a real concern, and that should be the main
> > motivation for the project.
>
> Thanks. I understand what you mean. extensibility is the main motivation of the
> project, i think that by implementing the cache system we can fix the
> small problems
> (reappearing of headers while setting and unsetting configs) and also
> implement new features
> like to unset a config easily.

I think the most interesting part of the config idea is turning the fetching of config variables "inside out".

That is, right now we turn control over to the config code, which invokes our callbacks. So we see all variables in sequence, and pick out the ones that are interesting. We implement precedence with a "last one wins" technique where we keep overwriting a variable with subsequent config options.

This can lead to difficult ordering situations, such as when a config option _might_ be respected based on another factor (e.g., the presence of a command line option, as in the status.branch option).

It also means that it's impossible to tell after the fact whether a value was set explicitly on the command line, by config, or if we are using a system default. Most of the time this doesn't matter, but there are a few cases where we care, and we end up having to manually manage a separate flag in each case.

By the phrase "inside out" above, I mean that we could read all config, and then fetch individual values as we need them. We do need to care about efficiency here, but only insofar as we don't cause any regressions (that is, the current system is fine speed-wise, we just need to make sure that we do not create new problems by calling a slow lookup in a tight loop or anything like that).

-Peff

karthik nayak· Mar 15, 2014, 16:30 UTC · re: Jeff King · lore

Re: [GSOC] Git Configuration API improvements

On Fri, Mar 14, 2014 at 10:01 AM, Jeff King <peff@peff.net> wrote:

Show 51 quoted lines

> On Tue, Mar 11, 2014 at 09:49:33PM +0530, karthik nayak wrote:
>
>> On Tue, Mar 11, 2014 at 8:21 PM, Matthieu Moy
>> <Matthieu.Moy@grenoble-inp.fr> wrote:
>> > karthik nayak <karthik.188@gmail.com> writes:
>> >
>> >> Currently we have multiple invocation of git_config() in an
>> >> individual invocation of git() which is not efficient. Also, it is
>> >> hard to implement new features,
>> >
>> > I think efficiency is not a real concern here. Config files are small
>> > and easy to parse, so parsing them multiple times isn't really
>> > noticeable from the performance point of view.
>> >
>> > OTOH, the extensibility is a real concern, and that should be the main
>> > motivation for the project.
>>
>> Thanks. I understand what you mean. extensibility is the main motivation of the
>> project, i think that by implementing the cache system we can fix the
>> small problems
>> (reappearing of headers while setting and unsetting configs) and also
>> implement new features
>> like to unset a config easily.
>
> I think the most interesting part of the config idea is turning the
> fetching of config variables "inside out".
>
> That is, right now we turn control over to the config code, which
> invokes our callbacks. So we see all variables in sequence, and pick out
> the ones that are interesting.  We implement precedence with a "last one
> wins" technique where we keep overwriting a variable with subsequent
> config options.
>
> This can lead to difficult ordering situations, such as when a config
> option _might_ be respected based on another factor (e.g., the presence
> of a command line option, as in the status.branch option).
>
> It also means that it's impossible to tell after the fact whether a
> value was set explicitly on the command line, by config, or if we are
> using a system default. Most of the time this doesn't matter, but there
> are a few cases where we care, and we end up having to manually manage
> a separate flag in each case.
>
> By the phrase "inside out" above, I mean that we could read all config,
> and then fetch individual values as we need them. We do need to care
> about efficiency here, but only insofar as we don't cause any
> regressions (that is, the current system is fine speed-wise, we just
> need to make sure that we do not create new problems by calling a slow
> lookup in a tight loop or anything like that).
>
> -Peff

Hello Jeff, Like you said yes, this will be a complete "inside out" change. Thanks for summing it up, really helpful. Currently i am going through the code and understanding how it works currently. Simultaneously working on the proposal[1], would be great to have your feedback on that. Thanks, Karthik

[1] https://gist.github.com/KarthikNayak/98569dd34326f7e6813a