Skip to content

Latest commit

 

History

History
460 lines (344 loc) · 21.1 KB

CURATORS.md

File metadata and controls

460 lines (344 loc) · 21.1 KB
Raw

This is a collection of instructions covering the processes that the Stackage curators - the team who maintain the Stackage project itself - should be doing on a regular basis. Curation activities are mostly automated, and do not take up a significant amount of time. The following is the current list of curators, in alphabetical order:

  • Adam Bergmark (@bergmark)
  • Alexey Zabelin (@alexeyzab)
  • Chris Dornan (@cdornan)
  • Dan Burton (@danburton)
  • Jens Petersen (@juhp)
  • Joe Kachmar (@jkachmar)
  • Michael Snoyman (@snoyberg)
  • Mihai Maruseac (@mihaimaruseac)

Workflow overview

This section sketches out at a high level how the entire Stackage build/curation process works:

  • build-constraints.yaml specifies packages to be included in Stackage
  • curator combines build-constraints.yaml with the current state of Hackage to create a build plan for a Stackage Nightly
  • curator can check that build plan to ensure all version bounds are consistent
    • The Travis job performs these two steps to provide immediate feedback on pull requests
  • Docker Hub builds a Docker image for running builds
  • The stackage-build server (described below) is able to run automated builds using the build.sh script
  • When a new Nightly build is completed, it is uploaded to the nightly repo
  • Once a week, we run an LTS minor bump. Instead of using build-constraints.yaml, that job takes the previous LTS release, turns it into constraints, and then bumps the version numbers to the latest on Hackage, in accordance with the version bounds in the build plan. This plans are uploaded to the LTS repo
  • Cutting a new LTS major release is essentially just a Stackage Nightly that gets rebuilt and uploaded as an LTS

Pull requests

The typical story on pull requests is: If Travis accepts it and the author only added packages under his/her own name, merge it. If the build later fails (see Adding Debian packages), then block the package until it's fixed.

If benchmarks, haddocks, or test suites fails at this point we typically also block the package until these issues are fixed. This in order to add packages with a clean slate.

Optionally we can check if packdeps says the package is up to date. Visit http://packdeps.haskellers.com/feed?needle=

Builds may fail because of unrelated bounds changes. If this happens, first add any version bounds to get master into a passing state (see "Fixing bounds issues"), then re-run the travis build.

A common issue is that authors submit newly uploaded packages, it can take up to an hour before this has synced across the stack infrastructure. You can usually compare the versions of the package in https://github.com/commercialhaskell/all-cabal-metadata/tree/master/packages/ to what's on hackage to see if this is the case. Wait an hour and re-run the pull request.

Tests also commonly fail due to missing test files, and sometimes due to doctest limitations. You can point the maintainer to https://github.com/bergmark/blog/blob/master/2016/package-faq.md

Fixing bounds issues

The most common activity you'll deal with in Stackage curation is a version bound issue, usually a restrictive upper bound. You fix this by opening an issue on the Stackage repo about the problem, and modifying the build-constraints.yaml file to work around it in one of the ways below. Be sure to refer to the issue for workarounds added to that file.

Temporary upper bounds

Most common technique, just prevent a new version of a library from being included immediately. This also applies to when only benchmarks and tests are affected.

  • Copy the curator output and create a new issue, see e.g commercialhaskell#2108

  • Add a new entry under the "stackage upper bounds" section of build-constraints.yaml. For the above example it would be

    "Stackage upper bounds":
        # https://github.com/commercialhaskell/stackage/issues/2108
        - pipes < 4.3.0
  • Commit (message e.g. "Upper bound for #2108")
  • Optionally: Verify with ./check locally
  • Push
  • Verify that everything works on the build server (you can restart the build or wait for it to to run again)

Sometimes releases for different packages are tightly coupled. Then it can make sense to combine them into one issue, as in commercialhaskell#2143.

If a dependency that is not explicitly in stackage is causing test or benchmark failures you can skip or expect them to fail (see "Skipping tests and benchmarks" and "Expecting test/benchmark/haddock failures"). Bonus points for reporting this upstream to that packages' maintainer.

Lifting upper bounds

You can try this when you notice that a package has been updated. You can also periodically try to lift bounds (I think it's good to do this at the start of your week /@bergmark)

If not all packages have been updated check if any of them are missing from the original issue and if so add a new comment mentioning them. A new package may appear if its dependencies were part of this issue but have been updated since the last time we checked. We want to give these new packages ample time to be upgraded.

If curator is happy commit the change ("Remove upper bounds and close #X"). After doing this the next nightly build may fail because some packages didn't have an upper bound in place, but compilation failed. In this case revert the previous commit so any disabled packages are enabled again, re-open the issue, and add a new comment with the failing packages. This is to give all maintainers enough time to upgrade for this case as well.

Amending upper bounds

With the pipes example above there was later a new release of pipes-safe that required the newer version of pipes. You can add that package to the same upper bounds section, (e.g. https://github.com/commercialhaskell/stackage/commit/6429b1eb14db3f2a0779813ef2927085fa4ad673) as we want to lift them simultaneously.

Skipping tests and benchmarks

Sometimes tests and benchmark dependencies are forgotten or not cared for. To disable compilation for them add them to skipped-tests or skipped-benchmarks. If a package is added to these sections they won't be compiled, and their dependencies won't be taken into account.

There are sub sections under these headers that is used to group types of failures together, and also to document what type of failures exist.

Expecting test/benchmark/haddock failures

The difference from the skipped sections is that items listed here are compiled and their dependencies are taken into account (but they are allowed to fail to build). These sections also have subsections with groups and descriptions.

One big category of test suites in this section are those requiring running services. We don't want to run those, but we do want to check dependencies and compile them.

If there are no version bounds that would fix the issue or if you can't figure it out, file it (e.g. commercialhaskell#2133) to ask the maintainer for help.

Waiting for new releases

Sometimes there is a failure reported on a (now possibly closed) issue on an external tracker. If an issue gets resolved but there is no hackage release yet we'd like to get notified when it's uploaded.

Add the package with its current version to the tell-me-when-its-released section. This will cause the build to stop when the new version is out.

Excluding packages

In an extreme case of a non-responsive maintainer, you can remove the package entirely from Stackage. We try to avoid that whenever possible.

This typically happens when we move to a new major GHC release or when there are only a few packages waiting for updates on an upper bounds issue.

Comment out the offending packages from the "packages" section and add a comment saying why it was disabled:

        # - swagger # bounds: aeson 1.0

If this causes reverse dependencies to be disabled we should notify the maintainers of those packages.

Updating the content of the Docker image used for building

Adding Debian packages for required system tools or libraries

Additional (non-Haskell) system libraries or tools should be added to stackage/debian-bootstrap.sh. After you've committed those changes, merging them into the nightly branch should trigger a DockerHub build. Simply run:

    $ git checkout nightly
    $ git merge master
    $ git push

Use Ubuntu Package content search to determine which package provides particular dev files (it defaults to xenial which is the version used to build Nightly).

Note that we generally don't install/run services needed for testsuites in the docker images - packages with tests requiring some system service can be added to expected-test-failures. It's good to inform the maintainer of any disabled tests (commenting in the PR is sufficient).

If a new package fails to build because of missing system libraries we often ask the maintainer to help figure out what to install.

Upgrading GHC version

The Dockerfile contains information on which GHC versions should be used. You can modify it and push it to Github to trigger a DockerHub build. The nightly branch is used for nightlies. For LTSes, we use the ltsX branch, where X is the major version number (e.g., lts3 for lts-3.*).

Note that when starting a new LTS major release, you'll need to modify Docker Hub to create a new Docker tag for the relevant branch name.

Update GHCVER in Dockerfile. (This env var automatically gets passed to debian-bootstrap.sh.)

Ensure that the global-hints.yaml file is updated with information on the latest GHC release by cloning that repo and running ./update-global-hints.yaml ghc-X.Y.Z.

Also required to build an LTS minor bump with a ghc version change: On the build server, modify /var/stackage/stackage/automated/wrk/lts-$THIS_LTS_MAJOR_VER/constraints.yaml and update the ghc-version. Then run NOPLAN=1 /var/stackage/stackage/automated/build lts-$THIS_LTS_MINOR_BUMP to build the LTS.

Getting the new image to the build server

Once a new Docker image is available, you'll need to pull it onto the stackage-build server (see below). Instead of pulling an unbounded number of images, I typically just delete all of the old images and let the new ones get downloaded:

docker rm $(docker ps -a -q)
docker rmi $(docker images -q)

but docker pull commercialhaskell/stackage:nightly can also be run instead just to update the nightly image say.

For a new GHC version you should also delete the cache .stack-work snapshot install directories on the stackage-build server to force all packages to be rebuilt clear up some space. See: issue#746. Eg:

# for example
SNAP_SERIES=nightly # or lts16 
OLD_GHCVER=8.10.1
rm -r work/$SNAP_SERIES/unpack-dir/.stack-work/install/x86_64-linux/*/$OLD_GHCVER/

This should also be done when moving the Nightly docker image to a new version of Ubuntu.

If you're impatient and would like to build the Docker image on the build server instead of waiting for Docker Hub, you can run the following command (replacing BRANCH=nightly if the image for a different branch is desired):

BRANCH=nightly
DIR=$(mktemp -d)
(cd $DIR \
  && git clone https://github.com/commercialhaskell/stackage \
  && cd stackage \
  && git checkout $BRANCH \
  && docker build --tag commercialhaskell/stackage:$BRANCH .)
rm -rf $DIR

Note that we do a clean clone of the stackage repo instead of using the existing checkout because of how docker build works: it will send the entire local directory contents as context to the Docker daemon, which in the case of the build tree is a lot of content. (We can discuss the wisdom—or lack thereof—of Docker's approach separately.)

stackage-build server

You'll need to get your SSH public key added to the machine. ~/.ssh/config info:

Host stackage-build
    User curators
    Hostname build.stackage.org

Running the build script

We currently run the builds manually so make it easy to see when there are bounds issues that need to be corrected. Automated this would be even better, we're just not there yet.

# Run a nightly build
/var/stackage/stackage/automated/run-nightly.sh

# Run an LTS minor bump
/var/stackage/stackage/automated/build.sh lts-15.1

# Run an LTS major bump
/var/stackage/stackage/automated/build.sh lts-16.0

Recommended: run these from inside a tmux session. If you get version bound problems on nightly or LTS major, you need to fix build-constraints.yaml (see info above).

Building LTS minor releases

Before running the build, please make sure that the Dockerfile in automated/dockerfiles/lts-X.Y is up to date, where X is the major version that you're building and Y is the latest minor version of X for which a Dockerfile exists.

  • If any changes need to be made, (eg, new GHC version), copy automated/lts-X.Y/Dockerfile to automated/lts-X.Z/Dockerfile, where Z is the minor version you're building, and include the new changes.
  • If you are building the first release of a new LTS major version, create a new lts-X.Z/Dockerfile based on the previous LTS's, and adjust the variables at the top to match the requirements of the snapshot. Ensure that STACK_VERSION is the latest release of Stack, and BOOTSTRAP_COMMIT is the commit ID of this repo containing the version of the bootstrap-commit.sh used to build the snapshot. Also ensure the FROM image's Ubuntu version matches that used in the root Dockerfile used to build this snapshot.

First run build.sh to regenerate updated ltsXX/work/constraints.yaml and ltsXX/work/snapshot-incomplete.yaml files.

For an LTS minor bump, you'll typically want to:

  • Add constraints to package range: fields under the source: field in that constraints.yaml.
  • Add new packages to the constraints.yaml file
  • Test, benchmark, haddock failures can also be added to package fields in the constraints.yaml if necessary, though it should be avoided if possible for LTS.

Then run NOPLAN=1 build.sh to build the generate an updated snapshot.

This replaces CONSTRAINTS=...' /var/stackage/stackage/automated/build.sh lts-x.y for the old curator-1.

If a build fails for bounds reasons, see all of the advice above. If the code itself doesn't build, or tests fail, open up an issue and then either put in a version bound to avoid that version or something else. It's difficult to give universal advice on how to solve things, since each situation is unique. Let's develop this advice over time. For now: if you're not sure, ask for guidance.

NOPLAN=1 If you wish to rerun a build without recalculating a build plan, you can set the environment variable NOPLAN=1. This is useful for such cases as an intermittent test failure, out of memory condition, or manually tweaking the plan file. (When using NOPLAN=1, if one needs to revert one package, say due to a build or test regression, one can edit current-plan.yaml and updated the SHA256 hash of the .cabal file, to avoid having to rebuild everything again.)

Sadly no longer true currently: Note LTS builds inherit the current Hackage data (stack updated for Nightly) to avoid excess extra rebuilding.

Timing

A looping script on the build server keeps trying to build nightly with sleep 30m interleaved. It only publishes the nightly once per day. This way new package versions or build failures can be caught early and hopefully the nightlies will be timely.

LTS minor bumps typically are run on Sundays.

Diskspace errors (and website sync debugging)

  • You can detect the problem by running df. If you see that / is out of space, we have a problem.

  • If you see that /var/stackage/ is out of space, you can:

    • run ./etc/diskspace/clean-old-stack-libs.sh [nightly|lts-XX]
    • If that is insufficient then remove all the old builds under the previous ghc/Cabal version:
      • rm -r /var/stackage/stackage/automated/work/[nightly|lts-XX]/unpack-dir/unpacked/*/.stack-work/dist/x86_64-linux/Cabal-X.Y.0.0/

    optionally (not recommended?):

    • rm -r /var/stackage/stackage/automated/work/lts*/unpack-dir/unpacked/
    • rm -r /var/stackage/stackage/automated/work/nightly/unpack-dir/unpacked/

Wiping the cache

Sometimes the cache can get corrupted which might manifest as can't load .so/.DLL. You can wipe the nightly cache and rebuild everything by doing rm -rf /var/stackage/stackage/automated/nightly. Replace nightly with lts7 to wipe the LTS 7 cache.

Force a single package rebuild

You can force a single package to rebuild by deleting its "previous result" file, e.g.:

$ rm /var/stackage/stackage/automated/nightly/work/builds/nightly/prevres/Build/cryptohash-0.11.9

Restarting docker

If docker hangs with e.g. docker: Error response from daemon: connection error: desc = "transport: dial unix /var/run/docker/containerd/docker-containerd .sock: connect: connection refused". you can restart it with sudo systemctl restart docker.service.

Local curator setup

We do not run the full stackage build locally as that might take too much time. However, some steps on the other hand are much faster to do yourself, e.g. verifying constraints without building anything.

To get started, install curator via Git:

$ git clone git@github.com:commercialhaskell/curator.git
$ cd curator && stack install curator

It is a good idea to upgrade curator at the start of your week. Then, clone the stackage repo, get the latest packages and run dependency resolution:

$ git clone git@github.com:commercialhaskell/stackage.git
$ cd stackage
$ ./check

This can be used to make sure all version bounds are in place, including for test suites and benchmarks, to check whether bounds can be lifted, and to get tell-me-when-its-released notifications.

curator does not build anything, so you wont see any compilation errors for builds, tests and benchmarks.

Adding new curators

  1. Add public ssh key to ~/.ssh/authorized_keys on build server
  2. Add to commercialhaskell/stackage project.

Dealing with a new GHC release

As mentioned in the GHC upgrade note, the major impact of a new GHC release is on the packages that are causing upper bounds to be put in place. In order to minimise out-of-date breakage and allow maintainers to have a solid chance of getting their packages into the newest LTS, we try to do the following:

Make an early announcement (in the form of a blog post, typically) of the new GHC release on the nightly build and the planned deadline for the new LTS release. Make it clear, that in the time coming up to this, we hope package maintainers will upgrade their packages to allow for the new GHC release.

We prefer to prune packages causing upper bounds constraints after the LTS release to allow the maximum amount of packages to get into the newest LTS.

You will almost always need to update the Win32 package version listed in the build-constraints.yaml file.

After the first LTS release, the package pruning process may begin in the nightly build in order to move forward with getting the latest versions of packages compatible with the new GHC release.

New LTS major bump

Every 3-6 months, we make a new major release of LTS. The procedure we follow for this is:

  1. Write a blog post on stackage.org announcing the intent to cut a major release. Give an estimated date two weeks in the future from the publication date of the post.
  2. Spread the blog post on social media and mailing lists as much as possible.
  3. Expect maintainers to send significant requests for added packages and relaxed upper bounds. There will likely be some hard decisions to be made regarding relaxing a bound versus keeping more packages. All of these changes occur on master and affect nightly.
  4. Once the estimated date hits, push a new lts-XX branch and trigger Docker Hub to build a Docker image for the new release.
  5. Run the build procedure for the new LTS release.
  6. After the LTS build completes, more aggressively prune upper bounds from build-constraints.yaml.
  7. Once both (5) and (6) are done, publish a new blog post on stackage.org announcing the new LTS and Nightly, with links to the change pages on stackage.org. Include a reminder that requests for packages to be added to LTS may be made on commercialhaskell/lts-haskell.