[SoC] egit: pre-proposal, problem recognition

[SoC] egit: pre-proposal, problem recognition

[Marek Zawirski]

Hello Git Team,

I'm a student interested participating in egit/jgit project for this 
year GSoC. Here I'm trying to investigate what can and need to be done 
for egit plugin and jgit library -> I'm asking you kindly for some 
comments or answers. That's somehow pre-proposal, as I'd like to know on 
which part should I do more research before preparing more detailed 
proposal.

[my git-awareness status]
To this time, I've used mainly SVN as version control. Last days were my 
first days with git, and I'm excited :) I've been doing some intensive 
learning: now I feel quite confident as a git user, also have done 
overview of git's internals (.git/ structure, data structures: DAG of 
objects, branches, tags, index etc.) and workflow of main operations. 
I've also downloaded and tested egit plugin on my Eclipse instance, and 
looked over jgit and egit code (nice!).
[/my git-awareness status]

Let's start from some clean-up:
http://git.or.cz/gitwiki/EclipsePluginWishlist - isn't this site a 
little outdated? It appears that tasks: Commit, Switch Branch (without 
dirty workspace/merge support) are already done, aren't they?
Some operations look for me as rather straight-forward to implement - 
not appropriate for a whole GSoC project (or do I underestimate?). 
Namely: Create/Delete Branch, Create/Delete Tag, Gitignore, Checkout(?).
GSoC2008 site states, that most wanted tasks are (that's reasonable) 
Push and Merge implementation. Fetch operation also looks to be not 
implemented yet? Do you agree that is also one of most important 
operation not yet supported? It seems that Pull and Clone are tasks for 
future, because of Merge or Fetch dependencies.

So what can I do:
I think that 1-2 from 3 main jgit+egit tasks (Push, Fetch, Merge) appear 
to be well-sized for a GSoC project. IMO, there is also need for a lot 
of smaller improvements in egit itself: preferences, preference page, 
user-friendliness related: more icons and information, help in dialogs, 
wizards, menus etc. I'm thinking of taking 1 main task - implementing 
first jgit part, then egit part. And if I finish it before time, I could 
work on set of such smaller improvements in egit itself, to make it 
easier/nicer for user.
I'm particulary interested in protocol implementation (Fetch or Push 
operation) as main-task, but Merge also looks interesting. What is more, 
I perceive Merge and Fetch operations as most wanted, because being on 
project critical-path, blocking implementation of other tasks. Merge 
seems also to be used very often itself.

Current implementation status:
There is currently no Eclipse (GUI, workspace logic) code directly for 
each of main tasks.
What about jgit? I haven't looked at details of implementation yet 
(sorry, I'm starting), but...
For Fetch or Push: I'll need to implement protocol(s) for sure: git, ssh 
protocols first as wiki-recommends - wisely, I think. What about support 
from jgit internals: do I need to add some git structures implementation 
to accomplish Fetch or Push: like object packing, needed objects 
tracking? Or is almost everything available, so I just need to use this 
structures and concentrate on upload/receive operation and protocol 
implementation? In case of Fetch, I guess that Create Branch (easy?) 
implementation is also needed. Is local Fetch or Push already supported 
in jgit? If not, I'd start from this. I've got also some doubts 
concerning SSH vs git access protocols. Are git-receive-pack and 
git-upload-pack used by both of these access protocols? If so, it make 
things easier:)
For Merge: I'll need to implement 3-way merge algorithm. Wiki says that 
one can look at Eclipse, if it is already implemented there. IMHO it's 
better to implement it at jgit level, independently of egit, basing on 
original git implementation, don't you think (I've seen that there are 
some efforts to implement git plugin for NetMeans, basing on jgit 
library)? I wonder how laborious this task is, i.e. as for Fetch/Push: 
does translation of this algorithm written in C to Java requires 
significant changes or improvements to existing jgit code?

What do you think about importance of each task and what is your feeling 
about time needed to implement each?


At the end, few words about me, what you could expect:
I'm a BSc/MSc undergraduate of a Computer Science course at Poznan 
University of Technology (Poland). My main areas of interests are 
distributed systems and software engineering, especially quality of 
software. I'm quite experienced Java developer; and C is my "second" 
language, so I can analyze existing git implementation. I've finished 
Eclipse Summer School workshop some time ago - on Eclipse plugins 
development, so I've got some knowledge on Eclipse internals. I'm ready 
to implement protocol or algorithm when needed.
I've got some experiences with JSch (bad experience...) and Trilead SSH 
(much more postive) - BSD-licensed SSH implementations for Java.

To this time I was working on many smaller projects, and one bigger (>1 
year, BSc and grids related), mostly in small teams, sometimes 
distributed. It seems to be great experience if I can join git community:)

Thank you for your attention, I'd appreciate any comments, expectations 
or info (answers)! If you want I can move to private discussion.

PS BTW: I've wonder what "Checkpoint project" in Eclipse Team menu 
stands for?

-- 
Marek Zawirski [zawir]
marek.zawirski@gmail.com

Re: [SoC] egit: pre-proposal, problem recognition

[Shawn O. Pearce]

Marek Zawirski <marek.zawirski@gmail.com> wrote:
> 
> Let's start from some clean-up:
> http://git.or.cz/gitwiki/EclipsePluginWishlist - isn't this site a 
> little outdated? It appears that tasks: Commit, Switch Branch (without 
> dirty workspace/merge support) are already done, aren't they?

Yea, I think you may be right.  I haven't been too current on egit
lately, but I think these did get implemented.  Sometimes the wiki
is out of date.  Happens because we would rather be working on code
than updating the wiki.  :)

> Some operations look for me as rather straight-forward to implement - 
> not appropriate for a whole GSoC project (or do I underestimate?). 
> Namely: Create/Delete Branch, Create/Delete Tag, Gitignore, Checkout(?).

I think you are right.  Doing all of them might be a GSoC project,
but a lot of it is just busy work and not that interesting of a
challenge.  Sure, it has to be done by someone, and everyone will
appreciate the effort once its complete, especially if its done well.
But doing just one of them as a GSoC project would be too small.

> GSoC2008 site states, that most wanted tasks are (that's reasonable) 
> Push and Merge implementation. Fetch operation also looks to be not 
> implemented yet?

Fetch and clone are in progress.  Robin Rosenburg (the current egit
maintainer) has it just about finished and will probably contribute
it soon.  But I think its going to cause merge conflicts with my own
revwalk work.  So that's why fetch isn't on the GSoC 2008 ideas list.

> So what can I do:
> I think that 1-2 from 3 main jgit+egit tasks (Push, Fetch, Merge) appear 
> to be well-sized for a GSoC project. IMO, there is also need for a lot 
> of smaller improvements in egit itself: preferences, preference page, 
> user-friendliness related: more icons and information, help in dialogs, 
> wizards, menus etc. I'm thinking of taking 1 main task - implementing 
> first jgit part, then egit part. And if I finish it before time, I could 
> work on set of such smaller improvements in egit itself, to make it 
> easier/nicer for user.

Yes, I agree with that, except no fetch, as it will hopefully merge
before GSoC summer officially starts.  A lot of the polish you are
talking about leaving to the end of the project is critical for any
good GUI.  I would not mind seeing it worked once the major features
are completed.

I did want to start on push work myself.  But if a student comes
along and does it for me, then awesome. :)

> operation) as main-task, but Merge also looks interesting. What is more, 
> I perceive Merge and Fetch operations as most wanted, because being on 
> project critical-path, blocking implementation of other tasks. Merge 
> seems also to be used very often itself.

Yes.  Lots of interesting git features (cherry-pick, revert, rebase)
are heavily based upon diff/apply and merge.  diff/apply can be
implemented in terms of merge in many cases, and we do that a lot
in C Git. So we do really need a working merge implementation.
 
> Current implementation status:
> There is currently no Eclipse (GUI, workspace logic) code directly for 
> each of main tasks.
> What about jgit? I haven't looked at details of implementation yet 
> (sorry, I'm starting), but...

Yea, aside from fetch there is no support yet in jgit for any of
the main tasks you were talking about, and the fetch part hasn't
merged to the public egit tree yet.

> For Fetch or Push: I'll need to implement protocol(s) for sure: git, ssh 
> protocols first as wiki-recommends - wisely, I think. What about support 
> from jgit internals: do I need to add some git structures implementation 
> to accomplish Fetch or Push: like object packing, needed objects 
> tracking?

The object packing (aka git-pack-objects in Java) is required
for push.  Object tracking to feed the packer is almost finished
in my revwalk branch, which also has not merged yet.  I started
that work to improve the performance and feature capabilities of
the History view, but its also to provide the object enumeration
functions required for the packfile generator, which is needed for
push to be implemented.

> Or is almost everything available, so I just need to use this 
> structures and concentrate on upload/receive operation and protocol 
> implementation?

Some of it is there.  The packer isn't, and the packer is a key
part of the protocol.  Its what is actually doing the data transfer.
The command part of the protocol is actually fairly simple, and is
even easier on the push side (the fetch side is more complex).

> I've got also some doubts 
> concerning SSH vs git access protocols. Are git-receive-pack and 
> git-upload-pack used by both of these access protocols? If so, it make 
> things easier:)

Yes.  If you dig into the C implementation what we are actually doing for
fetch over SSH is:

	ssh user@hostname git-upload-pack path/to/project.git

and for push:

	ssh user@hostname git-receive-pack path/to/project.git

Now the git:// native protocol just opens a TCP connection to a
server running git-daemon, sends the service name and directory
(which is the last args you see above to SSH), and then the daemon
executes that program.  After that point the SSH vs. native Git
are identical.  The major reason for SSH wrapping is its obvious
security benefits.  The major reason for the native Git protocol
is efficient anonymous publication, especially for popular open
source projects.

> For Merge: I'll need to implement 3-way merge algorithm. Wiki says that 
> one can look at Eclipse, if it is already implemented there. IMHO it's 
> better to implement it at jgit level, independently of egit, basing on 
> original git implementation, don't you think (I've seen that there are 
> some efforts to implement git plugin for NetMeans, basing on jgit 
> library)? I wonder how laborious this task is, i.e. as for Fetch/Push: 
> does translation of this algorithm written in C to Java requires 
> significant changes or improvements to existing jgit code?

Yea, I wrote that statement about using merge code from Eclipse,
but lately I have been thinking that is a bad idea.  Another person
has stated to talk about using jgit to build a Git NetBeans plugin,
and in another case there may be some idea of compiling jgit to
CLR and using it for a Git Mono plugin.

There is currently little-to-no merge support in jgit.  All of it
needs to be ported in, or obtained from a suitable Java library that
we can embed and ship as part of the package.  C Git for example
embeds and ships libxdiff for most of its diff/merge code.
 
> What do you think about importance of each task and what is your feeling 
> about time needed to implement each?

I think merge is a huge task, especially if you have to do your own
file level merge implementation by porting in say libxdiff to Java.
Focusing on merge may take the better part (or all!) of a GSoC
summer, but if you finished early then I would suggest doing the
much needed UI additions like you discussed above.
 
> I've got some experiences with JSch (bad experience...) and Trilead SSH 
> (much more postive) - BSD-licensed SSH implementations for Java.

Interesting that JSch wasn't a good experience.  Eclipse has just made it
an official part of the core platform, even without the IDE and its CVS
plugin.  I was sort of hoping that given it is now fully integrated into
Eclipse, and its key management is just part of the workspace, that we
could take advantage of that in egit.
 
> To this time I was working on many smaller projects, and one bigger (>1 
> year, BSc and grids related), mostly in small teams, sometimes 
> distributed. It seems to be great experience if I can join git community:)

You can join any time.  GSoC is just a good execuse.  :)

> PS BTW: I've wonder what "Checkpoint project" in Eclipse Team menu 
> stands for?

That's my fault.  A long time ago egit stored the workspace back
out as trees, rather than using the standard .git/index file.
It was slow and cost a lot of time, so I tried to put that activity
onto a low priority background job.  The job caused more problems,
so Robin got rid of it.  But the menu option to allow the user to
force a checkpoint is still there.

A checkpoint lets egit construct a tree object for the current
files, so a commit is faster as all of the blobs are already in the
object store.  But with the index and add (aka stage) now in egit,
and the commit dialog, that is less of an issue as the tree gets
created from the index, which you already staged files into.

In short, that menu option needs to come out at some point.

-- 
Shawn.

Re: [SoC] egit: pre-proposal, problem recognition

[Robin Rosenberg]

Den Sunday 23 March 2008 07.52.41 skrev Shawn O. Pearce:
> Marek Zawirski <marek.zawirski@gmail.com> wrote:

Welcome Marek.

> > Let's start from some clean-up:
> > http://git.or.cz/gitwiki/EclipsePluginWishlist - isn't this site a
> > little outdated? It appears that tasks: Commit, Switch Branch (without
> > dirty workspace/merge support) are already done, aren't they?
>
> Yea, I think you may be right.  I haven't been too current on egit
> lately, but I think these did get implemented.  Sometimes the wiki
> is out of date.  Happens because we would rather be working on code
> than updating the wiki.  :)

Deleting some stuff isn't that hard once you realize it needs to be done. 
Done.

> > Some operations look for me as rather straight-forward to implement -
> > not appropriate for a whole GSoC project (or do I underestimate?).
> > Namely: Create/Delete Branch, Create/Delete Tag, Gitignore, Checkout(?).
>
> I think you are right.  Doing all of them might be a GSoC project,
> but a lot of it is just busy work and not that interesting of a
> challenge.  Sure, it has to be done by someone, and everyone will
> appreciate the effort once its complete, especially if its done well.
> But doing just one of them as a GSoC project would be too small.
I agree it's not really the challenge a GSoc project should be. Leave for 
newcomers that wants something simple to chew. We can do that a patch
at a time.

> > GSoC2008 site states, that most wanted tasks are (that's reasonable)
> > Push and Merge implementation. Fetch operation also looks to be not
> > implemented yet?
>
> Fetch and clone are in progress.  Robin Rosenburg (the current egit
> maintainer) has it just about finished and will probably contribute
> it soon.  But I think its going to cause merge conflicts with my own
> revwalk work.  So that's why fetch isn't on the GSoC 2008 ideas list.
Yes, fetch (and clone) over git and git+ssh works on my machine. I did not go 
into the http based version which is a different story altoghether so it is 
still open.

> > So what can I do:
> > I think that 1-2 from 3 main jgit+egit tasks (Push, Fetch, Merge) appear
> > to be well-sized for a GSoC project. IMO, there is also need for a lot
> > of smaller improvements in egit itself: preferences, preference page,
> > user-friendliness related: more icons and information, help in dialogs,
> > wizards, menus etc. I'm thinking of taking 1 main task - implementing
> > first jgit part, then egit part. And if I finish it before time, I could
> > work on set of such smaller improvements in egit itself, to make it
> > easier/nicer for user.
>
> Yes, I agree with that, except no fetch, as it will hopefully merge
> before GSoC summer officially starts.  A lot of the polish you are
> talking about leaving to the end of the project is critical for any
> good GUI.  I would not mind seeing it worked once the major features
> are completed.
>
> I did want to start on push work myself.  But if a student comes
> along and does it for me, then awesome. :)

Push implies writing a highly efficient packing mechanism (actually could
work without the efficiency in form of delya packing, but I'm not sure we want 
that).

> > operation) as main-task, but Merge also looks interesting. What is more,
> > I perceive Merge and Fetch operations as most wanted, because being on
> > project critical-path, blocking implementation of other tasks. Merge
> > seems also to be used very often itself.
>
> Yes.  Lots of interesting git features (cherry-pick, revert, rebase)
> are heavily based upon diff/apply and merge.  diff/apply can be
> implemented in terms of merge in many cases, and we do that a lot
> in C Git. So we do really need a working merge implementation.

A subproject of that is making graphical merge resolution inside Eclipse. That
might not be such a big project though depending on how much support
there is in eclipse. I haven't looked into it really.

> > For Merge: I'll need to implement 3-way merge algorithm. Wiki says that
> > one can look at Eclipse, if it is already implemented there. IMHO it's
> > better to implement it at jgit level, independently of egit, basing on
> > original git implementation, don't you think (I've seen that there are
> > some efforts to implement git plugin for NetMeans, basing on jgit
> > library)? I wonder how laborious this task is, i.e. as for Fetch/Push:
> > does translation of this algorithm written in C to Java requires
> > significant changes or improvements to existing jgit code?
>
> Yea, I wrote that statement about using merge code from Eclipse,
> but lately I have been thinking that is a bad idea.  Another person
> has stated to talk about using jgit to build a Git NetBeans plugin,
> and in another case there may be some idea of compiling jgit to
> CLR and using it for a Git Mono plugin.
My mind wanders...

> There is currently little-to-no merge support in jgit.  All of it
> needs to be ported in, or obtained from a suitable Java library that
> we can embed and ship as part of the package.  C Git for example
> embeds and ships libxdiff for most of its diff/merge code.
>
> > What do you think about importance of each task and what is your feeling
> > about time needed to implement each?
>
> I think merge is a huge task, especially if you have to do your own
> file level merge implementation by porting in say libxdiff to Java.
> Focusing on merge may take the better part (or all!) of a GSoC
> summer, but if you finished early then I would suggest doing the
> much needed UI additions like you discussed above.
>
> > I've got some experiences with JSch (bad experience...) and Trilead SSH
> > (much more postive) - BSD-licensed SSH implementations for Java.
>
> Interesting that JSch wasn't a good experience.  Eclipse has just made it
> an official part of the core platform, even without the IDE and its CVS
> plugin.  I was sort of hoping that given it is now fully integrated into
> Eclipse, and its key management is just part of the workspace, that we
> could take advantage of that in egit.
The current fetch uses Ganymedes for SSH. It seesm to work well and has
lots of features for key management. There is an SHA-1 implementation there
also that should be a bit faster than Sun's default.

>
> > To this time I was working on many smaller projects, and one bigger (>1
> > year, BSc and grids related), mostly in small teams, sometimes
> > distributed. It seems to be great experience if I can join git
> > community:)
>
> You can join any time.  GSoC is just a good execuse.  :)
>
> > PS BTW: I've wonder what "Checkpoint project" in Eclipse Team menu
> > stands for?
>
> That's my fault.  A long time ago egit stored the workspace back
> out as trees, rather than using the standard .git/index file.
> It was slow and cost a lot of time, so I tried to put that activity
> onto a low priority background job.  The job caused more problems,
> so Robin got rid of it.  But the menu option to allow the user to
> force a checkpoint is still there.
I left it and almost forgot about it. We should probably drop it completely.

(sorry for not cleaning up my response fully here. bedtime)

-- robin

Re: [SoC] egit: pre-proposal, problem recognition

[Marek Zawirski]

Hi, coming back to git-dev. Was nice to get some valuable comments from 
Shawn and Robin :]

Robin Rosenberg <robin.rosenberg.lists@dewire.com> wrote:
>> Fetch and clone are in progress.  Robin Rosenburg (the current egit
>> maintainer) has it just about finished and will probably contribute
>> it soon.  But I think its going to cause merge conflicts with my own
>> revwalk work.  So that's why fetch isn't on the GSoC 2008 ideas list.
> Yes, fetch (and clone) over git and git+ssh works on my machine. I did not go 
> into the http based version which is a different story altoghether so it is 
> still open.
> 
>> I did want to start on push work myself.  But if a student comes
>> along and does it for me, then awesome. :)
> 
> Push implies writing a highly efficient packing mechanism (actually could
> work without the efficiency in form of delya packing, but I'm not sure we want 
> that).

I've tried to look at this problem closer.
I've seen that egit supports packed objects reading already. What makes
me not-sleeping is builtin-pack-objects.c code - mentioned packs
creation support. Hacking git story.... IMHO porting it into Java is a 
really big task , also because of its dependencies - especially 
delta/diff related code. It looks for me that it could be as difficult 
as merge support, or even harder :/

So... by not efficient packing you probably mean only compressing 
objects (+possibly arranging some order), but not finding deltas? I 
believe that the main reason why Robin said that he don't know whether 
it is wanted, is that when we are creating and sending such packs, 
they'll be stored on remote side in such inefficient way. Until repo 
admin make some cleanup (packing). I suspect that it's not enough reason 
for creating git-receive-pack --repack option (configuration/protocol 
negotiated, in SSH case, simply forced by client)?
BTW it shows that git receiver is very trusty (efficiency on top?) for
repo users: not only about what they transfer, but also about the way
they transfer and format it.

Anyway, I though that even push code without highly efficient
packing mechanism may be valuable for a meantime - at least as a base 
for further enhancements. I imagine that it maybe easy to abstract packs 
creation algorithm, that could be enhanced later. While rest of a commit 
operation would be already implemented.
In a meantime user could get warning in Commit dialog, that produced 
pack is not so efficient (yeah, not so pretty...).

>> Yes.  Lots of interesting git features (cherry-pick, revert, rebase)
>> are heavily based upon diff/apply and merge.  diff/apply can be
>> implemented in terms of merge in many cases, and we do that a lot
>> in C Git. So we do really need a working merge implementation.

.. so I now see that merge and packing operations look like most needed 
and most hard to do;)

> A subproject of that is making graphical merge resolution inside Eclipse. That
> might not be such a big project though depending on how much support
> there is in eclipse. I haven't looked into it really.

I've been looking around. There is general Eclipse Synchronization API 
as part of Team API, and related Synchronize View to provide user 
similar look&feel across different synchronization schemes.
Subclipse and CVS for Eclipse implements this API. I'm not very 
experienced CVS user, but saw that there are 2 Synchronization Views for 
CVS: for synchronization with repository (head?) and merging/conflicts 
resolutions. The second one is easier, and is what you probably mean. So 
there is pretty nice support from Eclipse. However, without stable merge 
API (not existing yet) is hard to do anything around that.

At the beginning, there could be just Compare Editor (CompareUI) for 
conflicting file edition without whole synchronization perspective and 
view  - not a big project. It is nice Eclipse editor, that "only" needs 
tree of DiffNode objects as input: result tree of three-way or two-way 
merge algorithm. Maybe some ad-hoc version could be provided without 
merge API (not sure), basing on git-merge result file with marked 
conflicts for creation of DiffNode tree (showing just two-way merge).


>> Yea, I wrote that statement about using merge code from Eclipse,
>> but lately I have been thinking that is a bad idea.  Another person
>> has stated to talk about using jgit to build a Git NetBeans plugin,
>> and in another case there may be some idea of compiling jgit to
>> CLR and using it for a Git Mono plugin.

Wow;) However, it seems that Microsoft is abandonning J# language? Don't 
know what is current status, I've been coding in C# only.

>> There is currently little-to-no merge support in jgit.  All of it
>> needs to be ported in, or obtained from a suitable Java library that
>> we can embed and ship as part of the package.  C Git for example
>> embeds and ships libxdiff for most of its diff/merge code.

That makes things harder. On the other hand, some people also faced this 
problem before. SVNKit ("jgit for svn") provides some 3-way merge and 
Differencer class from Eclipse org.eclipse.compare, possibly used by 
CVS. However, both of these solutions are product specific, so the code 
(if used; license issues) need to be copied & adapted for jgit. 
Especially SVNKit seems to depend on their own classes.

>> I think merge is a huge task, especially if you have to do your own
>> file level merge implementation by porting in say libxdiff to Java.
>> Focusing on merge may take the better part (or all!) of a GSoC
>> summer, but if you finished early then I would suggest doing the
>> much needed UI additions like you discussed above.

Well, words "better part (or all!)" doesn't sound very enthiusiatic for 
me... (explanation below)

>> Interesting that JSch wasn't a good experience.  Eclipse has just made it
>> an official part of the core platform, even without the IDE and its CVS
>> plugin.  I was sort of hoping that given it is now fully integrated into
>> Eclipse, and its key management is just part of the workspace, that we
>> could take advantage of that in egit.
> The current fetch uses Ganymedes for SSH. It seesm to work well and has
> lots of features for key management. There is an SHA-1 implementation there
> also that should be a bit faster than Sun's default.

Trilead for SSH is successor of Ganymed for SSH, as WWW tells. Yes, it 
was surprising for me, that JSch became official Eclipse core part. 
AFAIR especially annoying things were 1) you had to do polling for 
checking results of some operations(!) 2) no javadoc. I thought that in 
21st century people used to blocking operations, or non-blocking with 
Listener pattern and so on;> Or maybe something changed in library or 
I'm stupid and couldn't read library usage examples correctly.
Both libraries uses same OpenSSH key formats possibly, so maybe key 
management in workspace is possible even if case of "mixing" them.

>> You can join any time.  GSoC is just a good execuse.  :)

Especially good if I don't have much of free time except holidays for 
GSoC :)

To sum up a little. If you don't mind much, I would prefer doing some 
jgit task that will not fill my whole project, as I would like to do 
some Eclipse-related stuff also. Full merge implementation (with 
diff&merge algorithm adaptation) and full push operation (with efficient 
packing)  seems to be tasks that may take a really long time, as I've 
learned from your comments.

My reasons for interesting in tasks that are related to Eclipse are:
- it's nice to play with Eclipse, learn it more
- don't want to dig into real git internals for whole summer, if 
possible; I believe that coming into details for git-newbie may take 
much time
- There is friendly IBM Eclipse Support Center team in Poznan. As they 
are very enthusiastic about supporting Eclipse-related products, they've 
proposed to support GSOC Eclipse-related projects for free:) This means 
that I can work on GSoC in their office and get advices, ask for their 
knowledge (they are pros, Eclipse Foundation contributors) anytime. I 
think it's reasonable to benefit from such proposal.

Reasons for taking tasks related more to git:
- I always appreciate some network/distributed stuff ;)
- possibly interesting algorithms or Java profiling
- merge&packing is much NEEDED

So I would balance them if possible. For example, providing push 
implementation for GIT and SSH with some simple (extensible) packing 
algorithm at first and/or providing HTTP-based (commit walkers as you 
name them) fetch. Then moving to Eclipse/UI related stuff. What do you 
think?

I also see that project is very dynamic, so in 2 months some things may 
change, even some basic merge may support may come? If you agree, I can 
make in my application/proposal longer tasks list with priorities, and 
annotation that specific tasks for implementation will be chosen from 
top, depending on what is current dependencies status.

Maybe it would be easier to chat on IRC for us. What are your nicks if 
you are there (mine: zawir)?

Again, RFC & thanks for support :)

-- 
Marek Zawirski [zawir]
marek.zawirski@gmail.com