depot_tools/README.codereview

The git-cl README describes the git-cl command set.  This document
describes how code review and git work together in general, intended
for people familiar with git but unfamiliar with the code review
process supported by Rietveld.

== Concepts and terms
A Rietveld review is for discussion of a single change or patch.  You
upload a proposed change, the reviewer comments on your change, and
then you can upload a revised version of your change.  Rietveld stores
the history of uploaded patches as well as the comments, and can
compute diffs in between these patches.  The history of a patch is
very much like a small branch in git, but since Rietveld is
VCS-agnostic the concepts don't map perfectly.  The identifier for a
single review+patches+comments in Rietveld is called an "issue".

Rietveld provides a basic uploader that understands git.  This program
is used by git-cl, and is included in the git-cl repo as upload.py.

== Basic interaction with git
The fundamental problem you encounter when you try to mix git and code
review is that with git it's nice to commit code locally, while during
a code review you're often requested to change something about your
code.  There are a few different ways you can handle this workflow
with git:

1) Rewriting a single commit.  Say the origin commit is O, and you
   commit your initial work in a commit A, making your history like
   O--A.  After review comments, you commit --amend, effectively
   erasing A and making a new commit A', so history is now O--A'.
   (Equivalently, you can use git reset --soft or git rebase -i.)

2) Writing follow-up commits.  Initial work is again in A, and after
   review comments, you write a new commit B so your history looks
   like O--A--B.  When you upload the revised patch, you upload the
   diff of O..B, not A..B; you always upload the full diff of what
   you're proposing to change.

The Rietveld patch uploader just takes arguments to "git diff", so
either of the above workflows work fine.  If all you want to do is
upload a patch, you can use the upload.py provided by Rietveld with
arguments like this:

  upload.py --server server.com <args to "git diff">

The first time you upload, it creates a new issue; for follow-ups on
the same issue, you need to provide the issue number:

  upload.py --server server.com --issue 1234 <args to "git diff">

== git-cl to the rescue
git-cl simplifies the above in the following ways:

1) "git cl config" puts a persistent --server setting in your .git/config.

2) The first time you upload an issue, the issue number is associated with
   the current *branch*.  If you upload again, it will upload on the same
   issue.  (Note that this association is tied to a branch, not a commit,
   which means you need a separate branch per review.)

3) If your branch is "tracking" (in the "git checkout --track" sense)
   another one (like origin/master), calls to "git cl upload" will
   diff against that branch by default.  (You can still pass arguments
   to "git diff" on the command line, if necessary.)

In the common case, this means that calling simply "git cl upload"
will always upload the correct diff to the correct place.

== Patch series
The above is all you need to know for working on a single patch.

Things get much more complicated when you have a series of commits
that you want to get reviewed.  Say your history looks like
O--A--B--C.  If you want to upload that as a single review, everything
works just as above.

But what if you upload each of A, B, and C as separate reviews?
What if you then need to change A?

1) One option is rewriting history: write a new commit A', then use
   git rebase -i to insert that diff in as O--A--A'--B--C as well as
   squash it.  This is sometimes not possible if B and C have touched
   some lines affected by A'.
 
2) Another option, and the one espoused by software like topgit, is for
   you to have separate branches for A, B, and C, and after writing A'
   you merge it into each of those branches.  (topgit automates this
   merging process.)  This is also what is recommended by git-cl, which
   likes having different branch identifiers to hang the issue number
   off of.  Your history ends up looking like:

       O---A---B---C
            \   \   \
             A'--B'--C'

   Which is ugly, but it accurately tracks the real history of your work, can
   be thrown away at the end by committing A+A' as a single "squash" commit.

In practice, this comes up pretty rarely.  Suggestions for better workflows
are welcome.