This reverts commit 2f49dab586
.
12 KiB
Maintainers' Guide
Branch organization
The essence of this branching scheme is that you create "release branches" of independently releasable units of work. These can then be released by their maintainers when ready.
Master branch
Master is what's released on the network. Deployment instructions are in the next section, but tagged releases should always come from this branch.
Feature branches
Anyone can create feature branches. For those with commit access to urbit/urbit, you're welcome to create them in this repo; otherwise, fork the repo and create them there.
Usually, new development should start from master, but if your work depends on work in another feature branch or release branch, start from there.
If, after starting your work, you need changes that are in master, merge it into your branch. If you need changes that are in a release branch or feature branch, merge it into your branch, but understand that your work now depends on that release branch, which means it won't be released until that one is released.
Release branches
Release branches are code that is ready to release. All release branch names
should start with release/
.
All code must be reviewed before being pushed to a release branch. Thus, feature branches should be PR'd against a release branch, not master.
Create new release branches as needed. You don't need a new one for every PR, since many changes are relatively small and can be merged together with little risk. However, once you merge two branches, they're now coupled and will only be released together -- unless one of the underlying commits is separately put on a release branch.
Here's a worked example. The rule is to make however many branches are useful, and no more. This example is not prescriptive, the developers making the changes may add, remove, or rename branches in this flow at will.
Suppose you (plural, the dev community at large) complete some work in a
userspace app, and you put it in release/next-userspace
. Separately, you make
a small JS change. If you PR it to release/next-userspace
, then it will only
be released at the same time as the app changes. Maybe this is fine, or maybe
you want this change to go out quickly, and the change in
release/next-userspace
is relatively risky, so you don't want to push it out
on Friday afternoon. In this case, put the change in another release branch,
say release/next-js
. Now either can be released independently.
Suppose you do further work that you want to PR to release/next-userspace
, but
it depends on your fixes in release/next-js
. Simply merge release/next-js
into either your feature branch or release/next-userspace
and PR your finished
work to release/next-userspace
. Now there is a one-way coupling:
release/next-userspace
contains release/next-js
, so releasing it will
implicitly release release/next-js
. However, you can still release
release/next-js
independently.
This scheme extends to other branches, like release/next-kernel
or
release/os1.1
or release/ford-fusion
. Some branches may be long-lived and
represent simply the "next" release of something, while others will have a
definite lifetime that corresponds to development of a particular feature or
numbered release.
Since they are "done", release branches should be considered "public", in the sense that others may depend on them at will. Thus, never rebase a release branch.
When cutting a new release, you can filter branches with git branch --list 'release/*'
or by typing "release/" in the branch filter on Github. This will
give you the list of branches which have passed review and may be merged to
master and released. When choosing which branches to release, make sure you
understand the risks of releasing them immediately. If merging these produces
nontrivial conflicts, consider asking the developers on those branches to merge
between themselves. In many cases a developer can do this directly, but if it's
sufficiently nontrivial, this may be a reviewed PR of one release branch into
another.
Non-OTAable release branches
In some cases, work is completed which cannot be OTA'd as written. For example, the code may lack state adapters, or it may not properly handle outstanding subscriptions. It could also be code which is planned to be released only upon a breach (network-wide or rolling).
In this case, the code may be PR'd to a na-release/
branch. All rules are the
same as for release branches, except that the code does not need to apply
cleanly to an existing ship. If you later write state adapter or otherwise make
it OTAable, then you may PR it to a release branch.
Other cases
Outside contributors can generally target their PRs against master unless specifically instructed. Maintainers should retarget those branches as appropriate.
If a commit is not something that goes into a release (eg changes to README or CI), it may be committed straight to master.
If a hotfix is urgent, it may be PR'd straight to master. This should only be done if you reasonably expect that it will be released soon and before anything else is released.
If a series of commits that you want to release is on a release branch, but you really don't want to release the whole branch, you must cherry-pick them onto another release branch. Cherry-picking isn't ideal because those commits will be duplicated in the history, but it won't have any serious side effects.
Hotfixes
Here lies an informal guide for making hotfix releases and deploying them to the network.
Take this PR, as an example. This constituted a great hotfix. It's a single commit, targeting a problem that existed on the network at the time. Here's how it should be released and deployed OTA.
If the thing is acceptable to merge, merge it to master
Unless it's very trivial, it should probably have a single "credible looking" review from somebody else on it.
You should avoid merging the PR in GitHub directly. Instead, use the
sh/merge-with-custom-msg
script -- it will produce a merge commit with
message along the lines of:
Merge branch FOO (#PR_NUM)
* FOO:
bar: ...
baz: ...
Signed-off-by: SIGNER <signer@example.com>
We do this as it's nice to have the commit log information in the merge commit,
which GitHub's "Merge PR" button doesn't do (at least by default).
sh/merge-with-custom-msg
performs some useful last-minute urbit-specific
checks, as well.
You might want to alias sh/merge-with-custom-msg
locally, to make it easier
to use. My .git/config contains the following, for example:
[alias]
mu = !sh/merge-with-custom-msg
so that I can type e.g. git mu origin/foo 1337
.
Prepare a release commit
If you're making a Vere release, just play it safe and update all the pills.
For an Urbit OS release, after all the merge commits, make a release with the commit message "release: urbit-os-v1.0.xx". This commit should have up-to-date artifacts from pkg/interface and a new solid pill. If neither the pill nor the JS need to be updated (e.g if the pill was already updated in the previous merge commit), consider making the release commit with --allow-empty.
If anything in pkg/interface
has changed, ensure it has been built and
deployed properly. You'll want to do this before making a pill, since you want
the pill to have the new files/hash. For most things, it is sufficient to run
npm install; npm run build:prod
in pkg/interface
.
However, if you've made a change to Landscape's JS, then you will need to build
a "glob" and upload it to bootstrap.urbit.org. To do this, run npm install; npm run build:prod
in pkg/interface
, and add the resulting
pkg/arvo/app/landscape/index.[hash].js
to a fakezod at that path (or just create a
new fakezod with urbit -F zod -B bin/solid.pill -A pkg/arvo
). Run
:glob|make
, and this will output a file in fakezod/.urb/put/glob-0vXXX.glob
.
Upload this file to bootstrap.urbit.org, and modify +hash
at the top of
pkg/arvo/app/glob.hoon
to match the hash in the filename of the .glob
file.
Amend pkg/arvo/app/landscape/index.html
to import the hashed JS bundle, instead
of the unversioned index.js. Do not commit the produced index.js
and
make sure it doesn't end up in your pills (they should be less than 10MB each).
Tag the resulting commit
What you should do here depends on the type of release being made.
First, for Urbit OS releases:
If it's a very trivial hotfix that you know isn't going to break
anything, tag it as urbit-os-vx.y.z
. Here 'x' refers to the product version
(e.g. OS1, OS2..), 'y' to the continuity era in that version, and 'z' to an
OTA patch counter. So for a hotfix version, you'll just want to increment 'z'.
Use an annotated tag, i.e.
git tag -a urbit-os-vx.y.z
The tag format should look something like this:
urbit-os-vx.y.z
This release will be pushed to the network as an over-the-air update.
Release notes:
[..]
Contributions:
[..]
You can get the "contributions" section by the shortlog between the last release and this release:
git shortlog LAST_RELEASE..
I originally tried to curate this list somewhat, but now just paste it verbatim. If it's too noisy, yell at your colleagues to improve their commit messages.
Try to include a high-level summary of the changes in the "release notes" section. You should be able to do this by simply looking at the git log and skimming the commit descriptions (or perhaps copying some of them in verbatim). If the commit descriptions are too poor to easily do this, then again, yell at your fellow contributors to make them better in the future.
If it's not a trivial hotfix, you should probably make any number of release
candidate tags (e.g. urbit-os-vx.y.z.rc1
, urbit-os-vx.y.z.rc2
, ..), test
them, and after you confirm one of them is good, tag the release as
urbit-os-vx.y.z
.
For Vere releases:
Tag the release as urbit-vx.y.z
. The tag format should look something like
this:
urbit-vx.y.z
Note that this Vere release will by default boot fresh ships using an Urbit OS
va.b.c pill.
Release binaries:
(linux64)
https://bootstrap.urbit.org/urbit-vx.y.z-linux64.tgz
(macOS)
https://bootstrap.urbit.org/urbit-vx.y.z-darwin.tgz
Release notes:
[..]
Contributions:
[..]
Ensure the Vere release is marked as the 'latest' release and upload the two
.tgz
files to the release as darwin.tgz
and linux64.tgz
;
this allows us to programmatically retrieve the latest release at
urbit.org/install/mac/latest/ and
urbit.org/install/linux64/latest,
respectively.
The same schpeel re: release candidates applies here.
Note that the release notes indicate which version of Urbit OS the Vere release will use by default when booting fresh ships. Do not include implicit Urbit OS changes in Vere releases; this used to be done, historically, but shouldn't be any longer. If there are Urbit OS and Vere changes to be released, make two separate releases.
Deploy the update
(Note: the following steps are automated by some other Tlon-internal
tooling. Just ask ~nidsut-tomdun
for details.)
For Urbit OS updates, this means copying the files into ~zod's %home desk. The changes should be merged into /~zod/kids and then propagated through other galaxies and stars to the rest of the network.
For consistency, I create a release tarball and then rsync the files in.
$ wget https://github.com/urbit/urbit/archive/urbit-os-vx.y.z.tar.gz
$ tar xzf urbit-os-vx.y.z.tar.gz
$ herb zod -p hood -d "+hood/mount /=home="
$ rsync -zr --delete urbit-urbit-os-vx.y.z/pkg/arvo/ zod/home
$ herb zod -p hood -d "+hood/commit %home"
$ herb zod -p hood -d "+hood/merge %kids our %home"
For Vere updates, this means simply shutting down each desired ship, installing the new binary, and restarting the pier with it.
Continuous deployment
A subset of release branches are deployed continuously to the network. Thus far
this only includes release/next-userspace
, which deploys livenet-compatible
changes to select QA ships. Any push to master will automatically
merge master into release/next-userspace
to keep the streams at parity.
Announce the update
Post an announcement to urbit-dev. The tag annotation, basically, is fine here -- I usually add the %base hash (for Urbit OS releases) and the release binary URLs (for Vere releases). Check the urbit-dev archives for examples of these announcements.