MIME-Version: 1.0
Date: Tue, 15 Sep 2015 00:10:37 -0400
Message-ID: <CADm_WcY8Vy+k+5BaBS+jV6D6tmSXrok8rAxoPxxKOzUhyPWgMg@mail.gmail.com>
From: Jeff Garzik <jgarzik@gmail.com>
To: Bitcoin development mailing list <bitcoin-dev@lists.linuxfoundation.org>
Content-Type: text/plain; charset=UTF-8
Subject: [bitcoin-dev] libconsensus and bitcoin development process
Precedence: list

[collating a private mail and a github issue comment, moving it to a
better forum]

On libconsensus
---------------
In general there exists the reasonable goal to move consensus state
and code to a specific, separate lib.

To someone not closely reviewing the seemingly endless stream of
libconsensus refactoring PRs, the 10,000 foot view is that there is a
rather random stream of refactors that proceed in fits and starts
without apparent plan or end other than a one sentence "isolate
consensus state and code" summary.

I am hoping that
* There is some plan
* We will not see a five year stream of random consensus code movement
patches causing lots of downstream developer headaches.

I read every code change in every pull request that comes into
github/bitcoin/bitcoin with three exceptions:
* consensus code movement changes - too big, too chaotic, too
frequent, too unfocused, laziness guarantees others will inevitably
ACK it without me.
* some non-code changes (docs)
* ignore 80% of the Qt changes

As with any sort of refactoring, they are easy to prove correct, easy
to reason, and therefore quick and easy to ACK and merge.

Refactors however have a very real negative impact.
bitcoin/bitcoin.git is not only the source tree in the universe.
Software engineers at home, at startups, and at major companies are
maintaining branches of their own.

It is very very easy to fall into a trap where a project is merging
lots of cosmetic changes and not seeing the downstream ripple effects.
Several people complained to me at the conference about all the code
movement changes breaking their own work, causing them to stay on
older versions of bitcoin due to the effort required to rebase to each
new release version - and I share those complaints.

Complex code changes with longer development cycles than simple code
movement patches keep breaking.  It is very frustrating, and causes
folks to get trapped between a rock and a hard place:
- Trying to push non-trivial changes upstream is difficult, for normal
and reasonable reasons (big important changes need review etc.).
- Maintaining non-trivial changes out of tree is also painful, for the
aforementioned reasons.

Reasonable work languishes in constant-rebase hell, and incentivizes
against keeping up with the latest tree.


Aside from the refactor, libconsensus appears to be engineering in the
dark.  Where is any sort of plan?  I have low standards - a photo of a
whiteboard or youtube clip will do.

The general goal is good.   But we must not stray into unfocused
engineering for a non-existent future library user.

The higher priority must be given to having a source code base that
maximizes the collective developers' ability to maintain The Router --
the core bitcoin full node P2P engine.

I recommend time-based bursts of code movement changes.  See below;
for example, just submit & merge code movement changes on the first
week of every 2nd month.  Code movement changes are easy to create
from scratch once a concrete goal is known.  The coding part is
trivial and takes no time.

As we saw in the Linux kernel - battle lessons hard learned - code
movement and refactors have often unseen negative impact on downstream
developers working on more complicated changes that have more positive
impact to our developers and users.


On Bitcoin development release cycles & process
------------------------------------------------------------------

As I've outlined in the past, the Linux kernel maintenance phases
address some of these problems.  The merge window into git master
opens for 1 week, a very chaotic week full of merging (and rebasing),
and then the merge window closes.  Several weeks follow as the "dust
settles" -- testing, bug fixing, moving in parallel OOB with
not-yet-ready development.  Release candidates follow, then the
release, then the cycle repeats.

IMO a merge window approach fixes some of the issues with refactoring,
as well as introduces some useful -developer discipline- into the
development process.  Bitcoin Core still needs rapid iteration --
another failing of the current project -- and so something of a more
rapid pace is needed:
- 1st week of each month, merge changes.  Lots of rebasing during this week.
- remaining days of the month, test, bug fix
- release at end of month

If changes are not ready for merging, then so be it, they wait until
next month's release.  Some releases have major features, some
releases are completely boring and offer little of note.  That is the
nature of time-based development iteration.  It's like dollar cost
averaging, a bit.


And frankly, I would like to close all github pull requests that are
not ready to merge That Week.  I'm as guilty of this as any, but that
stuff just languishes.  Excluding a certain category of obvious-crap,
pull requests tend to default to a state of either (a) rapid merging,
(b) months-long issues/projects, (c) limbo.

Under a more time-based approach, a better pull request process would be to
* Only open pull requests if it's a bug fix, or the merge window is
open and the change is ready to be merged in the developer's opinion.
* Developers CC bitcoin-dev list to discuss Bitcoin Core-bound projects
* Developers maintain and publish projects via their own git trees
* Pull requests should be closed if unmerged after 7 days, unless it
is an important bug fix etc.

The problem with projects like libconsensus is that they can get
unfocused and open ended.  Code movement changes in particular are
cheap to generate.  It is low developer cost for the developer to
iterate all the way to the end state, see what that looks like, and
see if people like it.  That end state is not something you would
merge all in one go.  I would likely stash that tree, and then start
again, seek the most optimal and least disruptive set of refactors,
and generate and merge those into bitcoin/bitcoin.git in a time-based,
paced manner.  Announce the pace ahead of time - "cosmetic stuff that
breaks your patches will be merged 1st week of every second month"

To underscore, the higher priority must be given to having a source
code base and disciplined development process that maximizes the
collective developers' ability to maintain The Router that maintains
most of our network.

Modularity, refactoring, cleaning up grotty code generates a deep
seated happiness in many engineers.  Field experience however shows
refactoring is a never ending process which sometimes gets in the way
of More Important Work.