Since we're discussing it, this was the only consideration I had in
moving to PR's:
* Strongly suggest that a PR is applied as one commit (rebased by the
contributor or developer applying the change). Avoiding the word "must"
because there are edge-cases where multiple commits would be better
* Developers must rebase a change (avoid the merge commit) so the merge
is always a fast-forward merge.
The lineage just gets so hairy for history if we start getting a bunch
of branches/merges. This is what we've observed in PR's (at least, I
have always seen this happen) -- if we are defining policy, it might be
good to also codify this :)
Julian Hyde wrote:
In principle it is hard to compute the delta of a pull request, but in practice
it is easy. A well-formed pull request is a branch that is a small number of
commits away from the master branch at the time, and the pull requests that we
tend to accept are well-formed.
Since we don’t rewrite the master branch, you can easily apply the pull request
using “git rebase”. Because git knows where where the pull request’s branch
meets the master branch, it can do a better job than “patch” could.
Julian
On Feb 8, 2017, at 12:21 PM, Alan Gates<[email protected]> wrote:
I agree that PRs are easier to manage than attaching patches to JIRA. And now
days most contributors seem to prefer them as well.
One question I have is about traceability and findability. It is very nice for
people to be able to come to JIRA and figure out if others have had the same
problem they have, and if so if and where it's fixed, and exactly which commits
they need to pick up if they want the fix. Can all this be achieved with just
PRs?
If the answer is that PRs can't achieve that, I'd still vote for moving to
them. But I would also suggest continuing to open JIRAs that point to the PRs.
Alan.
On Feb 8, 2017, at 11:33 AM, Julian Hyde<[email protected]> wrote:
Our current policy is that we accept patches attached to JIRA case and pull requests
to https://github.com/apache/calcite<https://github.com/apache/calcite>. I
would like to propose that we no longer support patches.
Why? I argue that it makes the process easier for the committer. The pull
request implicitly does “git add” and “git remove”, whereas when applying a
patch you have to remember to apply these. The pull request comes in a branch,
so if I modify the code as I am reviewing it, I can easily save and restore my
state. Also, a pull request is “valid” as a contribution, from an IP
standpoint, even when not accompanied by a JIRA case.
Recently I went through 5 rounds of patches for a particular feature. I couldn’t
tell what had changed between one iteration of the patch and the next (you can’t
“diff" patches - you need to apply the patches to separate git branches and
diff the branches - yuck!). And I went through 3 test cycles and 24 hours before I
managed to “git add” all of the files. Yes, I did “git status” and I missed the 2
new files among all of the “.orig” and “.rej” files in my sandbox.
In summary. I propose that we accept contributions only as pull requests to
https://github.com/apache/calcite<https://github.com/apache/calcite>. If they
are non-trivial they should be accompanied by a JIRA case. Committers can propose
changes any way they like, as long as they commit the changes themselves, but if they
want to make it easier for others to review, they should use either a personal git
branch or a pull request.
Julian
