I used GitHub to check the number of pull requests we've merged to a few
of our projects over the last year in two month-chunks [1]. Each list
begins today and works backwards:
brooklyn-server: 86, 79, 85, 68, 95, 78.
brooklyn-client: 7, 4, 3, 8, 2, 4.
brooklyn-library: 11, 12, 9, 7, 17, 8.
total: 104, 95, 97, 83, 114, 90.
It's my opinion that these numbers show a consistency in the rate at
which we merge pull requests. Since a few committers have recently
joined I expect the rate to increase over the next few months. I would
like to know other figures like the average length of time between
opening a pull request and a committer merging it and the average length
of time before the first review comment. I'm sure there are many other
interesting metrics that could inform this discussion.
Maybe there's just one or two cases that we have managed badly.
Identifying those pull requests and understanding the cause of the delay
to each one would be valuable.
Regarding the complexity of Brooklyn and speaking as someone that has
used it in anger (sometimes literally) I would posit that many of us
have been caught out several times by the corner cases and unexpected
interactions between components that complexity so often implies. This
leads to cautious reviews and slow turnaround times. Quick-scan reviews
solve the speed at which we get code into the product. We need to think
harder about how we reduce its overall complexity.
Sam
1. I used GitHub's interface to search for "is:pr
merged:2017-03-04..2017-05-03", etc.
On 03/05/2017 09:41, Aled Sage wrote:
Hi Alex,
I agree with you that we have a problem with the reviewing of some of
our PRs - it's a bad situation for all concerned when these PRs stay
open for as long as they do.
I agree with your "eye-ball test" for a certain class of PR. I think
where we probably disagree is where the line is for "low-risk" PRs.
There are some examples where they "eye-ball test" would certainly
have helped.
I deliberately expanded the scope of this discussion, because there's
not one single solution - I'm not "turning this discussion in to
guidelines about PR". I'm adding to the discussion.
*Contributing Factors*
Breaking this down into the contributing factors, I believe those are:
1. Some PRs don't get enough attention from reviewers.
2. Not as much time as we'd like is spent by the community reviewing PRs.
3. The bar for being merged is too high (for at least some PRs).
4. Some PRs are very hard to review.
There are clearly examples of (1) where it's a no-brainer we should
have given it more attention, commented and merged. Your proposal for
the "eye-ball test" will help for some.
However, there are also examples of (1) that are caused by (4) -
commonly due to the complexity of the code being changed (e.g.
config-inheritance), sometimes made worse by it changing many things
(so it's more daunting to review).
Given (2) and (3), it suggests we should spread that time across more
PRs (i.e. some PRs that are getting a very thorough review could get
less, and folk try to use that "saved" time on some other PRs). I'm
not convinced that would actually happen in practice though!
*Solutions*
1. Add the "eye-ball test" to our reviewer guidelines (as described by
Alex) - and adjust our perception of "low-risk" over time, as we see
how it works.
2. Guidelines for PRs - what will make the reviewers' job considerably
easier, so we can get things merged faster? For example, small and
focused PRs, with good test coverage, linking to a jira issue where
appropriate.
3. Invest more time simplifying Brooklyn (see below).
*Complexity of Brooklyn*
I've heard from quite a few people that certain areas of Brooklyn are
far too hard to understand. Some people avoid reviewing PRs that touch
it, because it's so hard to understand the implications - they focus
their time on PRs that they feel more confident to review.
This is a symptom of an overly complex project. It would be great to
find more time to simplify things - e.g. to delete things from
Brooklyn, to make things more consistent, to refactor or even rewrite
some sections, and to add more javadoc.
*Accepted Limitations to Timely Review*
PRs that make far reaching changes to low-level details of Brooklyn
will always require a thorough review. Clearly we should try to find
the time for that promptly, but should always view them as high-risk.
*YOML*
If you insist on generalising YOML here, rather than a separate email
thread specifically about it, then: we should have commented very
quickly and discussed it promptly on the mailing list - at the level
of whether we want it (ignoring much of its technical details). If it
was pretty much anyone but Alex, then we should have commented saying:
"Very interesting, but this is a huge amount of code to add and
maintain in Brooklyn. Can you instead create a new github project
for this library, so that it can be worked on and maintained
separately? We'd then be interested to see how it can be used within
Brooklyn. Can you close this PR and let us know when/where you
create that library."
Like I said, that's for pretty much anyone but Alex. The difference is
that Alex wrote the first version of our yaml/camp parsing and knows
it better than anyone else. That original code definitely deserves a
re-write: it's become increasingly complicated as the supported yaml
has evolved. Alex has investigated different approaches and has come
up with a way that could greatly improve that code, and be used in
other places as well. Doing that in Brooklyn is simpler for him,
because it can evolve in tandem to satisfy requirements of Brooklyn.
I therefore suggest we discuss YOML separately, rather than generalising.
Aled
On 03/05/2017 02:13, Alex Heneveld wrote:
Aled,
> *Light-weight Review*
> I agree with you - where PRs look sensible, low-risk and unit
tested we should take more risk and
> merge them sooner (even if there's not been time for a thorough
review by the community).
I'm saying something a little different: we should _try_ for a
thorough review of *all* PRs. Which I think is uncontroversial.
> What should we do with a PR when we aren't able to review things in
as much depth as we'd like?
This is the question I'm asking, to ensure we handle PR's in a good
time frame. To summarise, I'm suggesting we make more of an effort,
and we fall back to an "eyeball test" a certain period of time (7
days max, less if it's simple?), triage the review to look at:
* clearly helpful & not obviously wrong
* low-risk / doesn't break compatibility
* good test coverage (and passing)
* likely to be maintained
If these can't be confirmed, the reviewer should say what they have
doubts about, maybe suggest what the contributor could do to help, or
appeal to other committers more familiar with an area. In any case
get a discussion going.
If these do seem confirmed, I still suggest we don't merge
immediately in the absence of a thorough review, but ping specific
committers likely to be interested. If no thorough review after a
few more days, _then_ merge.
I'm not suggesting any heavyweight process, but just enough to put
healthy forces on us as reviewers.
This is not a theoretical question, nor is it restricted to the YOML
PR. We're pretty good with most of our PRs and reviews but there are
plenty of examples where we've dropped the ball. Look at [1] which
is tiny and tests-only and took nine days to get a review. Or [2]
which yes combines a few related-but-different things but is by no
means a hard thing to review. It would take far more time to split
that up into 3 branches, test those locally, then babysit each of
those PR's than it would take for a reviewer to just get on with a
review. It's been sitting there for 2 months and doesn't even have a
comment.
This is not a good state of affairs. Turning this discussion in to
guidelines about PR's misses the point. If there's any change to our
docs/process made as a result of this discussion I'd like to see the
eyeball test added to a review process discussion.
Finally re YOML, there is an ML thread started when the issue was
raised. There was chatter beforehand but it wasn't an easy thing to
discuss until there was prototype code. The point is for 7 months
there have been no comments in any of these places, even after I've
run a public session explaining it and private sessions and the PR
itself says how it can be tested and how it is insulated from the
rest of the code (Thomas I think you missed that point). As there is
an ML thread and an open issue, either of which would be a fine place
to comment, but no one is -- the suggestion of a new separate ML
thread to solve the problem is bizarre. I say this is _exactly_ the
situation when we need guidelines for how we handle PR's that are not
being reviewed in a timely way.
Best
Alex
[1] https://github.com/apache/brooklyn-server/pull/600
[2] https://github.com/apache/brooklyn-server/pull/575
On 02/05/2017 19:21, Aled Sage wrote:
Hi Alex,
Interesting question. A few initial thoughts:
*YOML*
YOML (PR #363) is an exceptional case - we should not use that as an
example when discussing this meta-question. The PR is 12,000 lines
(including comments/notes), and was not discussed on the mailing
list before it was submitted. I suggest we have a separate email
thread specifically about merging that PR, as there are certainly
very useful things we'd get from YOML.
*Small PRs*
We should strongly encourage small focused PRs on a single thing,
wherever possible. That will make review faster, easier and lower
risk. For such PRs, we should strive for review+merge within days (7
days being an upper bound in normal circumstances, hopefully).
We can add some brief guidelines to this effect at
http://brooklyn.apache.org/developers/how-to-contribute.html
*Changing low-level Brooklyn*
PRs that change low-level things in Brooklyn (e.g. changes to
config-inheritance etc) deserve thorough review. They are high-risk
as the unforeseen consequences of the changes can be very subtle,
and break downstream blueprints that rely on old ways of doing things.
*Light-weight Review*
I agree with you - where PRs look sensible, low-risk and unit tested
we should take more risk and merge them sooner (even if there's not
been time for a thorough review by the community).
Aled
On 02/05/2017 15:50, Duncan Johnston Watt wrote:
Hi Alex
This is probably covered already but I guess there needs to be an
impact
assessment (by submitter?) before something is waved through by
default.
Best
Duncan
On 2 May 2017 at 06:52, Alex Heneveld
<[email protected]>
wrote:
Hi Brooklyners-
As many of you know, my YOML PR #363 [1] has been open for a
while. This
sets up a foundation for giving better documentation and feedback and
hugely simplifying how we do our parsing. However it's a very big
PR. I'm
eager to have people spend some time using it and ideally
extending it --
but here I wanted to raise a meta-question:
*W**hat should we do with a PR when we aren't able to review
things in as
much depth as we'd like?*
One option -- call it (A) -- is to say if we can't review things
thoroughly in a reasonable timeframe, we do a lighter review and
if the PR
looks promising and safe we merge it.
The other option -- call it (B) -- is to leave PRs open for as
long as it
takes for us to do the complete review.
I think most people have been approaching this with a mindset of
(B), and
while that's great for code quality and shared code understanding,
if we
can't deliver on that quickly, it's frankly anti-social. The
contributor
has to deal with merge conflicts (and the rudeness of his or her
contribution being ignored), and Brooklyn loses velocity. My PR is an
extreme example but many have been affected by slow reviews, and I
think
the expectation that reviews have to be so thorough is part of the
problem: it even discourages reviewers, as if you're not an
expert in an
area you probably don't feel qualified to review.
We have good test coverage so product risk of (A) is small, and we
have
great coders so I've no worry about us being able to solve
problems that
(A) might introduce. We should be encouraging reviewers to look
at any
area, and we need to solve the problem of slow reviews.
*I propose that the**standard we apply is that we quickly either
merge PRs
or identify what the contributor needs to resolve.
*I'm all for thorough reviews and shared understanding, but if we
can't do
this quickly I suggest we are better to sacrifice those things
rather than
block contributions, stifle innovation, and discourage reviews by
insisting
on a standards that we struggle to sustain.
As a general rule of thumb, maybe something like:
(1) After 7 days of no activity on a PR we go with an "eyeball test";
unless the following statement is untrue we say:
/I haven't done as much review as I'd like, but the code is clearly
helpful, not risky or obviously wrong or breaking compatibility,
it has
good test coverage, and we can reasonably expect the contributor or
committers to maintain it. Leaving open a bit longer in case
someone else
wants to review more but if nothing further in the next few days,
let's
merge.
/(If there are committers who are likely to be specifically
interested,
call them out as CC.)
(2) After 3 more days, if no activity, merge it.
And we encourage _anyone_ to review anything. If the above
response is
the baseline, everyone in our community is qualified to do it or
better and
we'll be grateful!
Best
Alex
[1] https://github.com/apache/brooklyn-server/pull/363
