I recently wrote up the Stackage data flow. The primary intent was to assist the rest of the Stackage curation team see how all the pieces fit together. However, it may also be of general interest to the rest of the community. In particular, some of the components used are not widely known and may be beneficial for completely separate projects (such as all-cabal-metadata).
Please check out the above linked copy of the file for the most up-to-date content. For convenience, I'm copying in the current content as of publication time below.
The Stackage project is really built on top of a number of different subcomponents. This page covers how they fit together. The Stackage data flow diagram gives a good bird's-eye view:
There are three inputs into the data flow:
Hackage is the upstream repository of all available open source Haskell packages that are part of our ecosystem. Hackage provides both cabal file metadata (via the 00-index.tar file) and tarballs of the individual packages.
build-constraints.yaml is the primary Stackage input file. This is where package maintainers can add packages to the Stackage package set. This also defines upper bounds, skipped tests, and a few other pieces of metadata.
stackage-content is a Github repository containing static file content served from stackage.org
For various reasons, we leverage Travis CI for running some processes. In particular:
all-cabal-files clones all cabal files from Hackage's 00-index.tar file into a Git repository without any modification
all-cabal-metadata uses all-cabal-metadata-tool to query extra metadata from Hackage about packages and put them into YAML files. As we'll see later, this avoids the need to make a lot of costly calls to Hackage APIs
Travis does not currently provide a means of running jobs on a regular basis. Therefore, we have a simple cron job on the Stackage build server that triggers each of the above builds every 30 minutes.
The heart of running Stackage builds is the stackage-curator tool. We run this on a daily basis on the Stackage build server for Stackage Nightly, and on a weekly basis for LTS Haskell. The build process is highly automated and leverages Docker quite a bit.
stackage-curator needs to know about the most recent versions of all packages, their tarball contents, and some metadata, all of which it gets from the Travis-generated sources mentioned in the previous section. In addition, it needs to know about build constraints, which can come from one of two places:
- When doing an LTS Haskell minor version bump (e.g., building lts-5.13), it grabs the previous version (e.g., lts-5.12) and converts the previous package set into constraints. For example, if lts-5.12 contains the package foo-5.6.7, this will be converted into the constraint
foo >= 5.6.7 && < 5.7.
- When doing a Stackage Nightly build or LTS Haskell major version bump (e.g., building lts-6.0), it grabs the latest version of the build-constraints.yaml file.
By combining these constraints with the current package data, stackage-curator can generate a build plan and check it. (As an aside, this build plan generation and checking also occurs every time you make a pull request to the stackage repo.) If there are version bounds problems, one of the Stackage curators will open up a Github issue and will add upper bounds, temporarily block a package, or some other corrective action.
Once a valid build plan is found, stackage-curator will build all packages, build docs, and run test suites. Assuming that all succeeds, it generates some artifacts:
- Uploads the build plan as a YAML file to either stackage-nightly or lts-haskell
- Uploads the generated Haddock docs and a package index (containing all used .cabal files) to haddock.stackage.org.
On the Stackage build server, we run the stackage-server-cron executable regularly, which generates:
- A SQLite database containing information on snapshots, the packages they contain, Hackage metadata about packages, and a bit more. This database is uploaded to S3.
- A Hoogle database for each snapshot, which is also uploaded to S3
The software running stackage.org is a relatively simple Yesod web application. It pulls data from the stackage-content repo, the SQLite database, the Hoogle databases, and the build plans for Stackage Nightly and LTS Haskell. It doesn't generate anything important of its own except for a user interface.
Stack takes advantage of many of the pieces listed above as well:
- It by default uses the all-cabal-hashes repo for getting package metadata, and downloads package contents from the hackage.fpcomplete.com mirror (using the hashes in the repo for verification)
- There are some metadata files in stackage-content which contain information on, for example, where to download GHC tarballs from to make
- Stack downloads the raw build plans for Stackage Nightly and LTS Haskell from the Github repo and uses them when deciding which packages to build for a given stack.yaml file