We currently have a fairly intricate way of running our PostgreSQL and MSSQL integration tests (not the API tests). By splitting them out, we can simplify this a lot. Most prominently, we can rely on Cabal to be our argument parser instead of writing our own.
We can also simplify how they're run in CI. They are currently (weirdly) run alongside the Python integration tests. This breaks them out into their own jobs for better visibility, and to avoid conflating the two.
The changes are as follows:
- The "unit" tests that rely on a running PostgreSQL database are extracted out to a new test directory so they can be run separately.
- Most of the `Main` module comes with them.
- We now refer to these as "integration" tests instead.
- Likewise for the "unit" tests that rely on a running MS SQL Server database. These are a little simpler and we can use `hspec-discover`, with a `SpecHook` to extract the connection string from an environment variable.
- Henceforth, these are the MS SQL Server integration tests.
- New CI jobs have been added for each of these.
- There wasn't actually a job for the MS SQL Server integration tests. It's pretty amazing they still run well.
- The "haskell-tests" CI job, which used to run the PostgreSQL integration tests, has been removed.
- The makefiles and contributing guide have been updated to run these.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/6912
GitOrigin-RevId: 67bbe2941bba31793f63d04a9a693779d4463ee1
With the vague idea that we can eventually avoid publishing ports at all, at least in CI, while still having the flexibility to publish ports locally.
If we can get there, it should help with the issues we're seeing on CI, where ports are already allocated from previous runs and containers aren't properly cleaned up.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/6836
GitOrigin-RevId: 1d636c73ae889e45e80ad52042c56daa1b9d5838
This regenerates _cabal.project.freeze_ using the new `--normalize` flag in _scripts/cabal-freeze-update.sh_. This flag simply skips the `cabal update` step.
I also added a section to preserve the liberal GHC version, as we currently run GHC v9.2.4.20220919 on CI, and most developers just run v9.2.4 locally. This can be removed once we upgrade to v9.2.5.
Finally, I have added a GitHub Action which ensures that the checked-in cabal.project.freeze file does not change upon re-normalization. As we don't run `cabal update`, this should be stable.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/6815
GitOrigin-RevId: 8cd3b5f3bbe5c61d8d3f61dfcc93b9d7c278419f
This upgrades the version of Ormolu required by the HGE repository to v0.5.0.1, and reformats all code accordingly.
Ormolu v0.5 reformats code that uses infix operators. This is mostly useful, adding newlines and indentation to make it clear which operators are applied first, but in some cases, it's unpleasant. To make this easier on the eyes, I had to do the following:
* Add a few fixity declarations (search for `infix`)
* Add parentheses to make precedence clear, allowing Ormolu to keep everything on one line
* Rename `relevantEq` to `(==~)` in #6651 and set it to `infix 4`
* Add a few _.ormolu_ files (thanks to @hallettj for helping me get started), mostly for Autodocodec operators that don't have explicit fixity declarations
In general, I think these changes are quite reasonable. They mostly affect indentation.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/6675
GitOrigin-RevId: cd47d87f1d089fb0bc9dcbbe7798dbceedcd7d83
This installs the ODBC Driver 18 for SQL Server in all our shipped Docker images, and update our tests and documentation accordingly.
This version supports arm64, and therefore can run natively (or via Docker) on macOS on aarch64.
`msodbcsql17` is still installed in production-targeted Docker images so that users do not _have_ to migrate to the new driver.
Nix expressions are packaged for the new driver, as it is not yet available in nixpkgs.
In this version, [the default encryption setting was changed from "no" to "yes"](https://techcommunity.microsoft.com/t5/sql-server-blog/odbc-driver-18-0-for-sql-server-released/ba-p/3169228). In addition, "mandatory" and "optional" were added as synonyms for "yes" and "no" respectively.
I have therefore modified all connection strings in tests to specify `Encrypt=optional` (and changed some from `Encrypt=no`). I chose "optional" rather than "no" because I feel it's more honest; these connection strings will work with or without an encrypted connection.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/6241
GitOrigin-RevId: 959f88dd1f271ef06a3616bc46b358f364f6cdfd
This upgrades CI and anyone using Nix to HLint v3.4.1.
If you're not using Nix, this doesn't actually _do_ anything on your
local machine; it's just a suggestion.
It also applies a bunch of simple HLint refactors, using
`make lint-hs-fix`.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/6324
GitOrigin-RevId: de8267e4909d6dcd3f83543188517f3aaeebc5f3
Now that we use the mssql-tools image for running `sqlcmd`, we no longer need this (or indeed use it very much).
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/6152
GitOrigin-RevId: be8852a93889553c3ffd2ec00c0dda81b36c3d05
Improvements to the Nix configuration so that macOS is supported.
Microsoft SQL Server is still not supported (yet; I have something in mind there), but the rest works. You can still use Homebrew to install the SQL Server drivers.
I had to make the following changes:
* I updated nixpkgs, because it's been a while.
* I made `ODBCINSTINI` optional so that it's not loaded on macOS, as it depends on `msodbcsql17`, which is broken on macOS.
* I upgraded OpenSSL.
* I set `DYLD_LIBRARY_PATH` on macOS so GHC finds OpenSSL; otherwise, it uses the wrong version of `libcrypto`, and fails with a fun error:
> WARNING: ghc is loading libcrypto in an unsafe way
* I patched GHC to fix compilation on macOS (copied from https://github.com/NixOS/nixpkgs/pull/149942).
To test this out, you can run `nix develop` (or install [direnv][] and add `use flake` to _.envrc.local_), and then try building HGE in the shell provided.
Fair warning: GHC needs to be built, and takes _aaages_ the first time. If this becomes useful to others, we can set up a shared cache.
[direnv]: https://direnv.net/
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5989
GitOrigin-RevId: 7130738d47709c37778b31c134061758ce23e959
* Add `source=` directives.
* Quote variables.
* Don't export variables that can fail to assign; instead, export afterwards.
* Write HGE stderr to the log file.
* Fix warnings around the PIDs.
* Disable a couple of false positives.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/6097
GitOrigin-RevId: ed1c696b8735cd5d63ba30b3105040a0a9eca63c
This makes a few changes to the test scripts and makefiles in order to make things simpler for the average Apple user.
First of all, we change the `wait_for_mysql` function to use "localhost", not "127.0.0.1", as this fixed an issue on my system when attempting to connect to the MySQL server.
Secondly, we split the SQL Server test image into two:
* The first is the server itself, which now automatically uses `azure-sql-edge` as the image if you are on an aarch64 chip and using the `make` commands.
* The second is the initialization script. Because `sqlcmd` is not available in the `azure-sql-edge` image on aarch64, we use a separate container based on `mssql-tools` to initialize the server.
The README has been updated.
Tested on both macOS/aarch64 (with other changes) and Linux/x86_64.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5986
GitOrigin-RevId: b16e079861dcbcc66773295c47d715e443b67eea
`spawn-dc-sqlite-agent` was incorrectly named, and the `start-dc-sqlite-agent` target was missing. In addition, we usually only make the `start-xyz` target public in `make help`, and so only that needs a comment.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5804
GitOrigin-RevId: 8064ce148d004ef73dbd8570ef30435423ede0a1
This means that if `remote_schemas/nodejs/package.json` changes, the
dependencies will be automatically reinstalled.
It also moves `package-lock.json` to the correct location (in the
directory in which we run `npm install`), and updates it.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5481
GitOrigin-RevId: f3fb431afd19de150f39ec2e4cb6572b896c870f
Making it easier to inject different ones later.
I also included a change to _.prettierignore_ so Visual Studio Code doesn't keep trying to reformat the JavaScript or YAML files in `server/tests-py`, as it can cause diffs to balloon for no obvious benefit.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5456
GitOrigin-RevId: bc6d548708160a328e1e61a00e19be8e124da025
It's about time.
To do this I had to check a few more boxes.
* I copied the flags from `graphql-engine.cabal` to the libraries in `server/lib`.
* I moved `Cacheable` instances of schema parser types beside the typeclass declaration.
* I removed imports of `Hasura.Prelude` from the tests, and rewrote them accordingly.
* I copied the `TestMonad` parse monad into `server/src-test/Hasura/GraphQL/Schema/RemoteTest.hs`, which was using it. I think this could be done with the real thing, but I tried replacing it with constraints and it messed with my head somewhat.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5311
GitOrigin-RevId: ebebcc50a16f2d517b7f730fe72410827ca3e86c
I'm trying to shore up the Python integration tests to make them more reliable. In doing so, I noticed this.
---
This ensures that the data directories are always in a mutable file system, and not going through overlayfs.
In my testing, this sped up PostgreSQL database initialization a lot. In particular, `CREATE DATABASE` commands went from > 5s to < 0.1s.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5248
GitOrigin-RevId: 5104aea32c92559c323803020d727bebfdafb8e6
I'm trying to shore up the Python integration tests to make them more reliable. In doing so, I noticed this.
---
Rather than hard-coding hostnames and ports, we can (and already do) inject these into the HGE process using environment variables.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5255
GitOrigin-RevId: 6bb593999ece42cedef6619f31f9d9b2e39f30ef
Docker Compose is now a plugin for Docker, bundled by default in Docker Desktop and many Linux distribution packages. The standalone `docker-compose` binary has been deprecated since Docker Compose v2.
Using the new version directly allows us to write development scripts that do not require `docker-compose` to be installed.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5185
GitOrigin-RevId: c8542b8b2405d1aa32288991688c6fde4af96383
We currently use `hpack` to generate the Cabal files from _package.yaml_
files for the two small libraries in _server/lib_. While this is more
convenient, we also check the Cabal files into the repository to avoid
needing an extra step upon pulling changes.
In order to ensure that the Cabal files do not get out of sync with the
hpack files, this introduces a few improvements:
1. Makefile targets to automatically generate the Cabal files without
needing to know the correct incantation. These targets are a
dependency of all build targets, so you can simply run
`make build-all` and it will work.
2. An extra comment at the top of all generated Cabal files that
explains how to regenerate it.
3. A `lint-hpack` Makefile target that verifies that the Cabal files
are up-to-date.
4. A CI job that runs `make lint-hpack`, to stop inconsistencies
getting merged into trunk.
Most of these changes are ported from #4794.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5217
GitOrigin-RevId: d3dfbe19ec00528368d357b6d0215a7ba4062f68
* Fix `make start-sqlserver` so it can find `sqlcmd`.
* Start the DC reference agent when testing all backends.
* Start backends in parallel so it doesn't take so long to spin them up. This is achieved by splitting all `start-*` tasks into `spawn-*` and `wait-for-*`.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5176
GitOrigin-RevId: ea483a755daa3fa6a8701f9d3dd0f393bc56510f
There's no reason to pin it to v3.7. This wasn't causing the Docker
Compose issues; it's a dependency version issue.
We actually don't need to install `docker-compose` at all. It's now a
Docker plugin (and is bundled in Docker Desktop) since Docker Compose
v2. You can run it with `docker compose`.
This also fixes a bug in _dev.sh_ due to a moved `make` target.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5197
GitOrigin-RevId: 638b2eb39c99e7a990e7a6c66388831438aed45f
Rather than a homebrewed approach, we can use `make` to figure out when
it's necessary to regenerate our venv.
This Makefile will regenerate _requirements.txt_ from
_requirements-top-level.txt_ when the latter is changed.
It will also regenerate the venv when _requirements.txt_ is changed
(i.e. changes are pulled, or it's regenerated as described above).
`make` uses file/directory timestamps to figure out what to rebuild.
This is probably more reliable than expecting people to update a version
number whenever they change a file.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5152
GitOrigin-RevId: 24b27d49bf6c4ba1d57ac38ea38ae278216c6d66
This helps us use the same versions locally as in CI. If you're using the Nix setup, it guarantees it.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5111
GitOrigin-RevId: 6e00cd7a78593df1e60fac37cc1195aba60e488f
I intend to use the Make targets around starting and stopping backend databases in order to wrap the Python integration tests in targets.
To that end, I have:
1. Factored out test infrastructure targets into _test-infrastructure.mk_.
2. Actually run the functions from _util.sh_; previously we were just running a bash script with no explicit commands. _util.sh_ must be `source`d and then you invoke the relevant function.
3. Run `docker-compose` through Nix if possible.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5136
GitOrigin-RevId: d7a01958045a82199c969ec3f88387d7d76d254b
I tried re-freezing _server/tests-py/requirements-top-level.txt_
recently, and discovered that it caused the tests to fail.
This pins a couple of dependencies so that we can safely re-freeze.
Specifically:
- `cryptography` is pinned at v3.*
- `graphene` is pinned at v2.*
- `PyJWT` is pinned at v2.3.*
- `websocket-client` is pinned at v0.56.0 (this was done in
_requirements.txt_ already, but that file is supposed to be
regenerated)
Upgrading `SQLAlchemy` required changing PostgreSQL URLs to use
"postgresql://" as the URL scheme, not "postgres://".
Updating `ruamel.yaml` caused a few tests to fail as we are passing
`ruamel.yaml.scalarstring.LiteralScalarString` values as header values.
This is fixed by explicitly converting header values to strings.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5120
GitOrigin-RevId: 9c12a3013c3d1f23dddbe781037663838b23f6f5
When pytest rewrites assertions to make them more useful, it also
truncates long assertion messages in the middle, often obscuring the
actual error.
Disabling this allows us to see the full message, which should hopefully
allow us to see the actual error.
In order to support this, we need to make sure that custom assertion
messages include the actual assertion information, as this will no
longer be rendered by pytest.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/4835
GitOrigin-RevId: de6839a3b40d0edc7cb96e46625eebca0aaf0c95
This improves the experience of testing against BigQuery in a few ways:
1. There is an explicit instruction to create a dataset, which was not
documented.
2. The verification script uses the environment variables, just like the
tests.
3. The verification script passes the correct content type headers.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/4706
GitOrigin-RevId: 15b5bd28b1762c130c0ec09d6babe8c08e05ab15
(Work here originally done by awjchen, rebased and fixed up for merge by
jberryman)
This is part of a merge train towards GHC 9.2 compatibility. The main
issue is the use of the new abstract `KeyMap` in 2.0. See:
https://hackage.haskell.org/package/aeson-2.0.3.0/changelog
Alex's original work is here:
#4305
BEHAVIOR CHANGE NOTE: This change causes a different arbitrary ordering
of serialized Json, for example during metadata export. CLI users care
about this in particular, and so we need to call it out as a _behavior
change_ as we did in v2.5.0. The good news though is that after this
change ordering should be more stable (alphabetical key order).
See: https://hasurahq.slack.com/archives/C01M20G1YRW/p1654012632634389
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/4611
Co-authored-by: awjchen <13142944+awjchen@users.noreply.github.com>
GitOrigin-RevId: 700265162c782739b2bb88300ee3cda3819b2e87
…p CI (#3975)
Trying out a new convention here for jobs that generate multiple
artifacts at different points in time, and where downstream jobs depend
on just a particular subset of artifacts and don't need to wait for the
entire job. We accomplish this with separate `Input` steps, which we
unblock manually via calls to the buildkite API once particular sets of
artifacts are complete.
We could accomplish the same thing by having multiple build server jobs
focused on particular targets ( this is in fact how it used to work)
but that is quite wasteful of CI resources; in order to build any of
the binaries quickly, we need significant number of cores; but if we
only build particular targets, we don't take advantage I've all those
cores relative to the parallelism a full cabal build gives us. That
would likely be more difficult to maintain as well.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/4256
Co-authored-by: Vishnu Bharathi <4211715+scriptnull@users.noreply.github.com>
GitOrigin-RevId: 5f5603da8a85f9a491877d51a7c08764c7199602
UPDATE: After testing in CI it turns out that the compile time Improvement is better than expected: even though we always have to recompile the OSS lib (due to Version.hs), downstream packages like Pro and multi-tenant can still benefit from some caching and avoid full recompilation. In the best case this takes us from 22 minutes to 13 minutes total.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/4104
GitOrigin-RevId: 76cbfc157064b33856e30f4c2b2ab2366f9c6089
## Description
This PR updates the notes-extracting script, to allow for a stable sort across runs:
- files are now treated in alphabetical order of the full file name:
- within a file, notes are treated in order
- that order is reflected in the generated index file
- within a note, references are sorted by file, then by line number
Additionally it fixes(?) a note whose format was unexpected.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/3352
GitOrigin-RevId: 2b2b57ec0aa2657d75a88e4951e6b19bb2d665e3
_(This PR is on top of #3352.)_
## Description
This PR overhauls our documentation CI steps to push all generated server documentation to the `gh-pages` branch of the OSS repo. The goal of this PR is to arrive in the situation where `https://hasura.github.io/graphql-engine/server/` is automatically populated to contain the following:
- all the markdown files from `server/documentation`, copied verbatim, no transformation applied
- all the notes, collected from the code by the `extract-notes.sh` script, in `server/notes`
- the generated haddock documentation for each major release or branch in `server/haddock`.
To do so, this PR does the following:
- it includes the script to extract notes from #3352,
- it rewrites the documentation checking CI step, to generate the notes and publish the resulting "server/documentation" folder,
- it includes a new CI step to deploy the documentation to the `gh-pages` branch
Of note:
- we will generate a different haddock folder for each main branch and release; in practice, that means the _main_, _stable_, _alpha_, _beta_ branches, and every build tagged with a version number
- the step that builds the haddock documentation checks that ALL projects in the repo build, including pro, but the deploy only deploys the graphql-engine documentation, as it pushes it to a publicly-accessible place
## Required work
**DO NOT MERGE THIS PR IT IS NOT READY**. Some work needs to go into this PR before it is ready.
First of all: the `gh-pages` branch of the OSS repo does NOT yet contain the documentation scaffolding that this new process assumes. At the bare minimum, it should be a orphan branch that contains a top-level README.md file, and a _server_ folder. An example of the bare minimum required can be previewed [on my fork](https://nicuveo.github.io/graphql-engine/server/).
The content of the `server/documentation` folder needs to be adjusted to reflect this; at the very least, a `README.md` file needs to be added to do the indexing (again, see the placeholder [on my fork](https://nicuveo.github.io/graphql-engine/server/) for an example).
This way of publishing documentation must be validated against [proposed changes to the documentation](https://github.com/hasura/graphql-engine-mono/pull/3294). @marionschleifer what do you think?
~~The buildkite code in this branch is currently untested, and I am not sure how to test it.~~
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/3380
GitOrigin-RevId: b24f6759c64ae29886c1f1b481b172febc512032
There's two major parts to this change:
1. cut down on environment variables needed to run the tests:
- lux parameters that don't change are now in code
- most remaining parameters have reasonable defaults
- the only variable that is still required to be set is HASURA_MULTITENANT_SPEC_DB_URL
- env.sh is no longer needed
2. find a better work-around for the problems running graphql-engine-multitenant
via cabal run (https://github.com/haskell/cabal/issues/7914), by adding a shell
script that implements a more correct version of cabal run.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/3461
GitOrigin-RevId: 0939a79cc45cd3c1c103719552b12099678850dd
So far we've only used `hlint` to lint the OSS code. This moves some things around to lint the Pro code as well.
Note that the CI action only runs `hlint` on Haskell files that are _changed_ by a PR, relative to its merge-base (usually `main`).
As a reminder, you can use the `ignore-server-hlint-checks` label to prevent this from blocking a merge.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/3383
GitOrigin-RevId: d5779c63d780f526a1d58ae4107f0d5262a23ec1
This PR upgrades some of the pinned dependencies do not build with python 3.10 - cffi, ruamel, py. Further, it upgrades other packages where the effort is minimal.
For the reviewers: Please review it commit by commit.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/3367
GitOrigin-RevId: c5401fe289d3185a79c4d382297f86fbde139825
The motivation here is:
- save some CI resources
- make it cleaner/more sensible to do library caching
Since OSS is a dependency of pro, and at the same time the various
executables can be built in parallel, doing this all in the same job
probably is more sensible.
Future improvements: there are separate pro and oss pipelines and one or
the other of those will need to wait unnecessarily while everything
finishes. In practice, currently this isn't a big deal, but we could use
the BuildKite blocking job API to get more fine-grained downstream
triggering.
GitOrigin-RevId: aa917996cbfc678becaaafca490d6edc7de89b1f
### Description
This PR fixes two small issues with the `dev.sh` script:
- when running the script without a specific test flag (`./scripts/dev.sh test`), we would start some of the containers more than once; this PR fixes this by checking whether the container is already running before starting it
- we were still doing some bigquery setup in the main code, despite that setup being already done in `parse-pytest-backend`, which would result in biquery setup errors even when testing non-bigquery backends; this PR just deletes the redundant line.
PR-URL: https://github.com/hasura/graphql-engine-mono/pull/3017
GitOrigin-RevId: 86a53b1fe974aae733a02bdf045389a0d7436cb7
### Description
The inherited roles integration tests were behind a flag, and its corresponding fixture, presumably to avoid enabling the option globally. However, #2288 introduced a new test using inherited roles that was not gated behind the flag, which fails when run with `dev.sh`. However, that test works on CI... because inherited roles are globally enabled there.
Consequently, this PR:
- globally enables inherited roles in dev.sh
- removes the flag and the associated fixture
https://github.com/hasura/graphql-engine-mono/pull/2358
Co-authored-by: Vishnu Bharathi <4211715+scriptnull@users.noreply.github.com>
GitOrigin-RevId: ebfa6754873324bed15b2cc5e37ec2d8008e8f8d
### Description
When running the tests with `dev.sh`, we always build the engine, before attempting to run the tests. This builds the engine twice, due to the flag change. This PR changes this to only do this first build if we're planning to actually run the integration tests.
https://github.com/hasura/graphql-engine-mono/pull/1964
GitOrigin-RevId: 32e8a6568f3665ad99adaf42421a711241974638
This essentially restores the original code from c425b554b8
(https://github.com/hasura/graphql-engine/pull/4013). Prior to this
commit we would slurp messages as fast as possible from the database
(one thing c425b55 fixed).
Another thing broken as a consequence of the same logic was the
removeEventFromLockedEvents logic which unlocks in-flight events
(breaking at-least-once delivery)
Some archeology, post-c425b55:
- cc8e2ccc erroneously attempted to refactor using `bracket`, resulting
in the same slurp-all-events behavior (since we don't ever wait for
processEvent to complete)
- at some point event processing within a batch is made serial, this
reported as a bug. See: https://github.com/hasura/graphql-engine/issues/5189
- in 0ef52292b5 (which I approved...) an `async` is added, again
causing the same issue...
GitOrigin-RevId: d8cbaab385267a4c3f1f173e268a385265980fb1
This claws back ~7min from integration tests (run serially, as with `dev.sh test --integration`
Further improvements would do well to focus on optimizing metadata operations, as `setup` dominates
GitOrigin-RevId: 76637d6fa953c2404627c4391447a05bf09355fa
