feat: platforms-aware package arch aliases #100

jjmaestro · 2024-09-20T17:36:14Z

Note

Stacked on top of #97

feat: platforms-aware package arch aliases

Package repos can now be used directly without the /<ARCH> bit.

The "root-level" BUILD in the package repo now has platforms-aware aliases that use select to pick the right arch for the :data, :control, etc., targets.

Now, you can do:

load("@rules_oci//oci:defs.bzl", "oci_image")

oci_image(
    name = "foo",
    base = "@distroless_base_nossl_debian12",
    tars = [
        "@debian12//coreutils",
        "@debian12//bash",
    ],
)

when before you would have needed something like:

load("@rules_oci//oci:defs.bzl", "oci_image")

CPU_ARCH = [
    ("arm64", "arm64"),
    ("x86_64", "amd64"),
]

PACKAGES = [
    "@debian12//coreutils",
    "@debian12//bash",
]

PACKAGES_ARCH = select({
    "@platforms//cpu:%s" % cpu: [
        "%s/%s" % (package, arch)
        for package in PACKAGES
    ]
    for cpu, arch in CPU_ARCH
})

oci_image(
    name = "foo",
    base = "@distroless_base_nossl_debian12",
    tars = PACKAGES_ARCH,
)

To force a specific architecture the "arch packages" can still be used just like before with @repo//package/arch or via constraints and platforms (--platforms=...).

chore: remove `select()` from the examples

This is a separate diff so that we can test the previous diff with the changes plus the examples with select in place and make sure everything works as-is with the new package aliases.

* separate the script into a template file so it's easier to shellcheck and syntax highlight in editors. * shellcheck the script and remove all SC2086 warnings ("Double quote to prevent globbing and word splitting") * improve the buildozer help messages in the copy.sh script: * reduce duplication of buildozer command * add a more clear autofix bazel run command that can be easily copy-pasted * change some of the variable names in the copy.sh template for longer, easier to understand names (repo_name >> name, lock_label >> label) * move repo_name and workspace_relative_path into variables to reduce line length and improve readability

While working with some flaky mirrors and trying to figure out why they were failing I found the _fetch_package_index code a bit hard to follow so here's my attempt at streamlining it a bit: * Change the general flow of the for-loop so that we can directly set the reasons for failure in failed_attempts * Remove both integrity as an argument and as a return value since neither is ever used. * return the content of the Packages index instead of the path, since we already have the repository_context. This way we don't need rctx anywhere else. * Reword failure messages adding more context and debug information * Shorter lines and templated strings, trying to make the code easier to read and follow.

* remove the `state` "intermediary `struct`" in `package_resolution.bzl` since it wasn't used / needed. * refactor and move _set_dict from `util.bzl` to a `_package_set` method * use `dict .get()` with default values instead of the "nested `if`s" * renamed `package()` to `package_get` and make it return all package versions when the version is not specified so we can remove `_package_versions` * reordeder `(name, version, arch)` args to match the order of the index keys `(arch, name, version)`

Add testing for package_index mocking the external / side effects (downloads, decompression, etc).

The package resolution debugging that e.g. checks the package dependencies should all be within package_resolution.bzl _resolve_all() and not "leak out" returning the information that's only needed for debugging / logging. Also: * reduce the verbosity of the optional dependencies warning by just printing one message per root package instead of one per package. * break up the long lines to build the error messages and remove the "# buildifier: disable=print".

* move version constraint parsing from package_resolution to its own _parse_version_and_constraint method in version.bzl * refactor _version_relop into a compare method in version.bzl plus a VERSION_OPERATORS dict so that (1) we use the operator strings everywhere and (2) we can use the keys to validate the operators.

Previously we had: ```starlark pkgindex = package_index.new(rctx, sources = sources, archs = manifest["archs"]) pkgresolution = package_resolution.new(index = pkgindex) ``` And none of the code of package_resolution was used anywhere but in resolve.bzl and after initializing the `pkgindex`. Also, it makes sense since we are building the index from the manifest and once we have the index we use it to resolve the packages and populate the lock.

Cleanup resolve.bzl and package_index.bzl by moving all of the manifest functionality to a separate manifest.bzl file where we now do all of the work to generate the lock: manifest parsing, validation and the package index and resolution. IMHO this is how it should be because the lock is the "frozen state" of the manifest. * _parse() parses the YAML * _from_dict validates the manifest dict and does the rest of the changes that we need to produce a manifest struct * add extra validation for e.g. duplicated architectures * _lock is the only method that's exposed to the outside and it encapsulates all of the other parts, calling _from_dict and all of the package index and resolution, to produce the lock file. * move get_dupes to util.bzl * refactor the "source" struct into the new manifest where we can now centralize a lot of the structure and logic spread across multiple parts of the code. * remove yq_toolchain_prefix since it's always "yq" and, looking at GH code search, this seems to be a copy-paste leftover from rules_js (or the other way around)... the code is always the same and it never receives a string different from "yq".

* move all of the "package logic" to pkg.bzl * The v2 lockfile format: * doesn't need the fast_package_lookup dict because it's already using a dict to store the packages. * has the dependencies sorted so the lockfile now has stable serialization and the diffs of the lock are actually usable and useful to compare with the changes to the manifest. * removes the package and dependency key from the lockfile, now it's done via an external function (make_deb_import_key in deb_import.bzl) * Remove add_package_dependency from the lockfile API. Now, the package dependencies are passed as an argument to add_package. This way, the lockfile functionality is fully contained in lockfile.bzl and e.g. we can remove the "consistency checks" that were only needed because users could forget to add the dependency as a package to the lockfile. * Ensure backwards-compatibility by internally converting lock v1 to v2. Also, when a lock is set and it's in v1 format, there's a reminder that encourages the users to run @repo//:lock to update the lockfile format.

By separating the migration from the previous commit we get to 1. in the previous commit, run all tests with the new code while locks are still v1 2. update the locks n this commit to V2 so we can then re-run all tests in the final state.

…t.bzl Refactor the package repo templates into their own methods and massively cleanup the `for`-loop in `_deb_package_index_impl`. IMHO overall now there's a much better and clear separation of concerns between the "index repo" (`apt/private/index.bzl`) and the "package repos" (`apt/private/deb_import.bzl`).

@alexconrey

Fixes issue GoogleContainerTools#56 Follow-up and credit to @alexconrey (PR GoogleContainerTools#55), @ericlchen1 (PR GoogleContainerTools#64) and @benmccown (PR GoogleContainerTools#67) for their work on similar PRs that I've reviewed and drawn some inspiration to create "one 💍 PR to merge them all" 😅 Problem: Debian has two types of repos: "canonical" and "flat". Each has a different sources.list syntax: "canonical": ``` deb uri distribution [component1] [component2] [...] ``` (see https://wiki.debian.org/DebianRepository/Format#Overview) flat: ``` deb uri directory/ ``` (see https://wiki.debian.org/DebianRepository/Format#Flat_Repository_Format) A flat repository does not use the dists hierarchy of directories, and instead places meta index and indices directly into the archive root (or some part below it) Thus, the URL logic in _fetch_package_index() is incorrect for these repos and it always fails to fetch the Package index. Solution: Just use the Debian sources.list convention in the 'sources' section of the manifest to add canonical and flat repos. Depending on whether the channel has one directory that ends in '/' or a (dist, component, ...) structure the _fetch_package_index and other internal logic will know whether the source is a canonical or a flat repo. For example: ``` version: 1 sources: # canonical repo - channel: bullseye main contrib url: https://snapshot-cloudflare.debian.org/archive/debian/20240210T223313Z # flat repos, note the trailing '/' and the lack of distribution or components - channel: bullseye-cran40/ url: https://cloud.r-project.org/bin/linux/debian - channel: ubuntu2404/x86_64/ url: https://developer.download.nvidia.com/compute/cuda/repos archs: - amd64 packages: - bash - r-mathlib - nvidia-container-toolkit-base ``` Disregarding the "mixing" of Ubuntu and Debian repos for the purpose of the example, this manifest shows that you can mix canonical and flat repos and you can mix multiarch and single-arch flat repos and canonical repos. You will still have the same problems as before with packages that only exist for one architecture and/or repos that only support one architecture. In those cases, simply separate the repos and packages into their own manifests. NOTE: The NVIDIA CUDA repos don't follow Debian specs and have issues with the package filenames. This is addressed in a separate commit.

Although the Debian repo spec for 'Filename' (see https://wiki.debian.org/DebianRepository/Format#Filename) clearly says that 'Filename' should be relative to the base directory of the repo and should be in canonical form (i.e. without '.' or '..') there are cases where this is not honored. In those cases we try to work around this by assuming 'Filename' is relative to the sources.list directory/ so we combine them and normalize the new 'Filename' path. Note that, so far, only the NVIDIA CUDA repos needed this workaround so maybe this heuristic will break for other repos that don't conform to the Debian repo spec.

Package repos can now be used directly without the `/<ARCH>` bit. The "root-level" BUILD in the package repo now has platforms-aware aliases that use 'select' to pick the right arch for ':data', ':control', etc. Now, you can do: ```starlark load("@rules_oci//oci:defs.bzl", "oci_image") oci_image( name = "foo", base = "@distroless_base_nossl_debian12", tars = [ "@debian12//coreutils", "@debian12//bash", ], ) ``` when before you would have needed something like: ```starlark load("@rules_oci//oci:defs.bzl", "oci_image") CPU_ARCH = [ ("arm64", "arm64"), ("x86_64", "amd64"), ] PACKAGES = [ "@debian12//coreutils", "@debian12//bash", ] PACKAGES_ARCH = select({ "@platforms//cpu:%s" % cpu: [ "%s/%s" % (package, arch) for package in PACKAGES ] for cpu, arch in CPU_ARCH }) oci_image( name = "foo", base = "@distroless_base_nossl_debian12", tars = PACKAGES_ARCH, ) ``` To force a specific architecture the "arch packages" can still be used just like before with '@repo//package/arch' or via constraints and platforms ('--platforms=...').

This is a separate diff so that we can test the previous diff with the changes plus the examples with `select` in place and make sure everything works as-is with the new package aliases.

jjmaestro force-pushed the feat-platforms-aware-package-arch-aliases branch from aeab122 to 038afcc Compare September 22, 2024 18:43

jjmaestro force-pushed the feat-platforms-aware-package-arch-aliases branch 3 times, most recently from 53213e5 to edab37a Compare October 29, 2024 00:40

jjmaestro marked this pull request as ready for review October 29, 2024 01:19

jjmaestro force-pushed the feat-platforms-aware-package-arch-aliases branch from edab37a to 1a6532a Compare October 29, 2024 08:24

jjmaestro added 10 commits November 14, 2024 00:50

tests: package_index_test

22121d0

Add testing for package_index mocking the external / side effects (downloads, decompression, etc).

chore: migrate repo locks to v2

f47c059

By separating the migration from the previous commit we get to 1. in the previous commit, run all tests with the new code while locks are still v1 2. update the locks n this commit to V2 so we can then re-run all tests in the final state.

jjmaestro force-pushed the feat-platforms-aware-package-arch-aliases branch from 1a6532a to 581fc0b Compare November 14, 2024 01:07

jjmaestro added 6 commits November 14, 2024 11:49

chore: remove select() from the examples

13f8bab

This is a separate diff so that we can test the previous diff with the changes plus the examples with `select` in place and make sure everything works as-is with the new package aliases.

test: example with "all" arch only

4949412

jjmaestro force-pushed the feat-platforms-aware-package-arch-aliases branch from 581fc0b to 4949412 Compare November 14, 2024 11:54

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: platforms-aware package arch aliases #100

feat: platforms-aware package arch aliases #100

jjmaestro commented Sep 20, 2024

feat: platforms-aware package arch aliases #100

Are you sure you want to change the base?

feat: platforms-aware package arch aliases #100

Conversation

jjmaestro commented Sep 20, 2024

feat: platforms-aware package arch aliases

chore: remove select() from the examples

chore: remove `select()` from the examples