Adds support for disabling installation from pre-built wheels i.e. the
package must be built from source locally.
We will still always use pre-built wheels for metadata during
resolution.
Available via `--no-binary` and `--no-binary-package <name>` flags in
`pip install` and `pip sync`. There is no flag for `pip compile` since
no installation happens there.
```
--no-binary
Don't install pre-built wheels.
When enabled, all installed packages will be installed from a source distribution.
The resolver will still use pre-built wheels for metadata.
--no-binary-package <NO_BINARY_PACKAGE>
Don't install pre-built wheels for a specific package.
When enabled, the specified packages will be installed from a source distribution.
The resolver will still use pre-built wheels for metadata.
```
When packages are already installed, the `--no-binary` flag will have no
affect without the `--reinstall` flag. In the future, I'd like to change
this by tracking if a local distribution is from a pre-built wheel or a
locally-built wheel. However, this is significantly more complex and
different than `pip`'s behavior so deferring for now.
For reference, `pip`'s flag works as follows:
```
--no-binary <format_control>
Do not use binary packages. Can be supplied multiple times, and each time adds to the
existing value. Accepts either ":all:" to disable all binary packages, ":none:" to empty the
set (notice the colons), or one or more package names with commas between them (no colons).
Note that some packages are tricky to compile and may fail to install when this option is
used on them.
```
Note we are not matching the exact `pip` interface here because it seems
complicated to use. I think we may want to consider adjusting our
interface for this behavior since we're not entirely compatible anyway
e.g. I think `--force-build` and `--force-build-package` are clearer
names. We could also consider matching the `pip` interface or only
allowing `--no-binary <package>` for compatibility. We can of course do
whatever we want in our _own_ install interfaces later.
Additionally, we may want to further consider the semantics of
`--no-binary`. For example, if I run `pip install pydantic --no-binary`
I expect _just_ Pydantic to be installed without binaries but by default
we will build all of Pydantic's dependencies too.
This work was prompted by #895, as it is much easier to measure
performance gains from building source distributions if we have a flag
to ensure we actually build source distributions. Additionally, this is
a flag I have used frequently in production to debug packages that ship
Cythonized wheels.
This is https://github.com/astral-sh/puffin/pull/947 again but this time
merging into main instead of downstack, sorry for the noise.
---
Windows has a default stack size of 1MB, which makes puffin often fail
with stack overflows. The PR reduces stack size by three changes:
* Boxing `File` in `Dist`, reducing the size from 496 to 240.
* Boxing the largest futures.
* Boxing `CachePolicy`
## Method
Debugging happened on linux using
https://github.com/astral-sh/puffin/pull/941 to limit the stack size to
1MB. Used ran the command below.
```
RUSTFLAGS=-Zprint-type-sizes cargo +nightly build -p puffin-cli -j 1 > type-sizes.txt && top-type-sizes -w -s -h 10 < type-sizes.txt > sizes.txt
```
The main drawback is top-type-sizes not saying what the `__awaitee` is,
so it requires manually looking up with a future with matching size.
When the `brotli` features on `reqwest` is active, a lot of brotli types
show up. Toggling this feature however seems to have no effect. I assume
they are false positives since the `brotli` crate has elaborate control
about allocation. The sizes are therefore shown with the feature off.
## Results
The largest future goes from 12208B to 6416B, the largest type
(`PrioritizedDistribution`, see also #948) from 17448B to 9264B. Full
diff: https://gist.github.com/konstin/62635c0d12110a616a1b2bfcde21304f
For the second commit, i iteratively boxed the largest file until the
tests passed, then with an 800KB stack limit looked through the
backtrace of a failing test and added some more boxing.
Quick benchmarking showed no difference:
```console
$ hyperfine --warmup 2 "target/profiling/main-dev resolve meine_stadt_transparent" "target/profiling/puffin-dev resolve meine_stadt_transparent"
Benchmark 1: target/profiling/main-dev resolve meine_stadt_transparent
Time (mean ± σ): 49.2 ms ± 3.0 ms [User: 39.8 ms, System: 24.0 ms]
Range (min … max): 46.6 ms … 63.0 ms 55 runs
Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs. It might help to use the '--warmup' or '--prepare' options.
Benchmark 2: target/profiling/puffin-dev resolve meine_stadt_transparent
Time (mean ± σ): 47.4 ms ± 3.2 ms [User: 41.3 ms, System: 20.6 ms]
Range (min … max): 44.6 ms … 60.5 ms 62 runs
Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs. It might help to use the '--warmup' or '--prepare' options.
Summary
target/profiling/puffin-dev resolve meine_stadt_transparent ran
1.04 ± 0.09 times faster than target/profiling/main-dev resolve meine_stadt_transparent
```
## Summary
I got confused by why `VerbatimUrl` was on `Path`. Since it's directly
computed from it, I think we should just compute it as-needed. I think
it's also possibly-buggy because the URL is the URL of the _directory_,
not the artifact itself, which differs from other distributions.
## Summary
It turns out that storing an absolute URL for every file caused a
significant performance regression. This PR attempts to address the
regression with two changes.
The first is that we now store the raw string if the URL is an absolute
URL. If the URL is relative, we store the base URL alongside the raw
relative string. As such, we avoid serializing and deserializing URLs
until we need them (later on), except for the base URL.
The second is that we now use the internal `Url` crate methods for
serializing and deserializing. If you look inside `Url`, its standard
serializer and deserialization actually convert it to a string, then
parse the string. But the crate exposes some other methods for faster
serialization and deserialization (with fewer guarantees). I think this
is totally fine since the cache is entirely internal.
If we _just_ change the `Url` serialization (and no other code -- so
continue to store URLs for every file), then the regression goes down to
about 5%:
```shell
❯ python -m scripts.bench \
--puffin-path ./target/release/main \
--puffin-path ./target/release/relative --puffin-path ./target/release/puffin \
scripts/requirements/home-assistant.in --benchmark resolve-warm
Benchmark 1: ./target/release/main (resolve-warm)
Time (mean ± σ): 496.3 ms ± 4.3 ms [User: 452.4 ms, System: 175.5 ms]
Range (min … max): 487.3 ms … 502.4 ms 10 runs
Benchmark 2: ./target/release/relative (resolve-warm)
Time (mean ± σ): 284.8 ms ± 2.1 ms [User: 245.8 ms, System: 165.6 ms]
Range (min … max): 280.3 ms … 288.0 ms 10 runs
Benchmark 3: ./target/release/puffin (resolve-warm)
Time (mean ± σ): 300.4 ms ± 3.2 ms [User: 255.5 ms, System: 178.1 ms]
Range (min … max): 295.4 ms … 305.1 ms 10 runs
Summary
'./target/release/relative (resolve-warm)' ran
1.05 ± 0.01 times faster than './target/release/puffin (resolve-warm)'
1.74 ± 0.02 times faster than './target/release/main (resolve-warm)'
```
So I considered _just_ making that change. But 5% is kind of
borderline...
With both of these changes, the regression is down to 1-2%:
```
Benchmark 1: ./target/release/relative (resolve-warm)
Time (mean ± σ): 282.6 ms ± 7.4 ms [User: 244.6 ms, System: 181.3 ms]
Range (min … max): 275.1 ms … 318.5 ms 30 runs
Benchmark 2: ./target/release/puffin (resolve-warm)
Time (mean ± σ): 286.8 ms ± 2.2 ms [User: 247.0 ms, System: 169.1 ms]
Range (min … max): 282.3 ms … 290.7 ms 30 runs
Summary
'./target/release/relative (resolve-warm)' ran
1.01 ± 0.03 times faster than './target/release/puffin (resolve-warm)'
```
It's consistently ~2%-ish, but at this point it's unclear if that's due
to the URL change or something other change between now and then.
Closes#943.
Add directory `--find-links` support for local paths to pip-compile.
It seems that pip joins all sources and then picks the best package. We
explicitly give find links packages precedence if the same exists on an
index and locally by prefilling the `VersionMap`, otherwise they are
added as another index and the existing rules of precedence apply.
Internally, the feature is called _flat index_, which is more meaningful
than _find links_: We're not looking for links, we're picking up local
directories, and (TBD) support another index format that's just a flat
list of files instead of a nested index.
`RegistryBuiltDist` and `RegistrySourceDist` now use `WheelFilename` and
`SourceDistFilename` respectively. The `File` inside `RegistryBuiltDist`
and `RegistrySourceDist` gained the ability to represent both a url and
a path so that `--find-links` with a url and with a path works the same,
both being locked as `<package_name>@<version>` instead of
`<package_name> @ <url>`. (This is more of a detail, this PR in general
still work if we strip that and have directory find links represented as
`<package_name> @ file:///path/to/file.ext`)
`PrioritizedDistribution` and `FlatIndex` have been moved to locations
where we can use them in the upstack PR.
I added a `scripts/wheels` directory with stripped down wheels to use
for testing.
We're lacking tests for correct tag priority precedence with flat
indexes, i only confirmed this manually since it is not covered in the
pip-compile or pip-sync output.
Closes#876
Previously, the url on file could either be a relative or an absolute
url, depending on the index, and we would finalize it lazily. Now we
finalize the url when converting `pypi_types::File` to
`distribution_types::File`. This change is required to make the hashes
on `File` optional (https://github.com/astral-sh/puffin/pull/910), which
are currently the only unique field usable for caching.
## Summary
Now that `get_or_build_wheel` will often _also_ handle the unzip step,
we need to move our per-target locking (`OnceMap`) up a level.
Previously, it was only applied to the unzip step, to prevent us from
attempting to unzip into the same target concurrently; now, it's applied
at the `get_wheel` level, which includes both downloading and unzipping.
## Test Plan
It seems like none of our existing tests catch this -- perhaps because
they're too "simple"? You need to run into a situation in which you're
doing multiple source distribution builds concurrently (since they'll
all try to download `setuptools`):
```
rm -rf foo && virtualenv --clear .venv && cargo run -p puffin-cli -- pip-compile ./scripts/requirements/pydantic.in --verbose --cache-dir foo
```
## Summary
This PR adds support for `prepare_metadata_for_build_wheel`, which
allows us to determine source distribution metadata without building the
source distribution. This represents an optimization for the resolver,
as we can skip the expensive build phase for build backends that support
it.
For reference, `prepare_metadata_for_build_wheel` seems to be supported
by:
- `hatchling` (as of
[1.0.9](https://hatch.pypa.io/latest/history/hatchling/#hatchling-v1.9.0)).
- `flit`
- `setuptools`
In fact, it seems to work for every backend _except_ those using legacy
`setup.py`.
Closes#599.
Changes `File::size` from a `usize` to a `u64`.
The motivations are that with tensorflow wheels being 475 MB
(https://pypi.org/project/tensorflow/2.15.0.post1/#files), we're already
only one order of magnitude away and to avoid target dependent failures.
I've tried to investigate puffin's performance wrt to builds and
parallelism in general, but found the previous instrumentation to
granular. I've tried to add spans to every function that either needs
noticeable io or cpu resources without creating duplication. This also
fixes some wrong tracing usage on async functions
(https://docs.rs/tracing/latest/tracing/struct.Span.html#in-asynchronous-code)
and some spans that weren't actually entered.
## Summary
This PR adds support for relative URLs in the simple JSON responses. We
already support relative URLs for HTML responses, but the handling has
been consolidated between the two. Similar to index URLs, we now store
the base alongside the metadata, and use the base when resolving the
URL.
Closes#455.
## Test Plan
`cargo test` (to test HTML indexes). Separately, I also ran `cargo run
-p puffin-cli -- pip-compile requirements.in -n
--index-url=http://localhost:3141/packages/pypi/+simple` on the
`zb/relative` branch with `packse` running, and forced both HTML and
JSON by limiting the `accept` header.
This crate started off as generic caching utilities, but we started
adding a lot of Puffin-specific stuff (like the cache buckets
abstraction that knows about Git vs. direct URL vs. indexes and so on).
This PR moves the generic stuff into a new `cache-key` crate.
This is a pure refactor to follow-up #690, to separate the metadata that
we know upfront about distributions (like the version, for
registry-based distributions) vs. the metadata that requires building
(like the version, for URL-based distributions).
## Summary
This PR adds a `VerbatimUrl` struct to preserve verbatim URLs throughout
the resolution and installation pipeline. In short, alongside the parsed
`Url`, we also keep the URL as written by the user. This enables us to
display the URL exactly as written by the user, rather than the
serialized path that we use internally.
This will be especially useful once we start expanding environment
variables since, at that point, we'll be able to write the version of
the URL that includes the _unexpected_ environment variable to the
output file.
## Summary
At present, we have two separate phases within the installation pipeline
related to populating wheels into the cache. The first phase downloads
the distribution, and then builds any source distributions into wheels;
the second phase unzips all the built wheels into the cache.
This PR merges those two phases into one, such that we seamlessly
download, build, and unzip wheels in one pass. This is more efficient,
since we can start unzipping while we build. It also ensures that if the
install _fails_ partway through, we don't end up with a bunch of
downloaded wheels that we never had a chance to unzip. The code is also
much simpler.
The main downside is that the user-facing feedback isn't as granular,
since we only have one phase and one progress bar for what was
originally three distinct phases.
Closes https://github.com/astral-sh/puffin/issues/571.
## Test Plan
I ran the benchmark script on two separate requirements files, and saw a
7% and 31% speedup respectively:
```text
+ TARGET=./scripts/benchmarks/requirements.txt
+ hyperfine --runs 100 --warmup 10 --prepare 'virtualenv --clear .venv' './target/release/main pip-sync ./scripts/benchmarks/requirements.txt --no-cache' --prepare 'virtualenv --clear .venv' './target/release/puffin pip-sync ./scripts/benchmarks/requirements.txt --no-cache'
Benchmark 1: ./target/release/main pip-sync ./scripts/benchmarks/requirements.txt --no-cache
Time (mean ± σ): 269.4 ms ± 33.0 ms [User: 42.4 ms, System: 117.5 ms]
Range (min … max): 221.7 ms … 446.7 ms 100 runs
Benchmark 2: ./target/release/puffin pip-sync ./scripts/benchmarks/requirements.txt --no-cache
Time (mean ± σ): 250.6 ms ± 28.3 ms [User: 41.5 ms, System: 127.4 ms]
Range (min … max): 207.6 ms … 336.4 ms 100 runs
Summary
'./target/release/puffin pip-sync ./scripts/benchmarks/requirements.txt --no-cache' ran
1.07 ± 0.18 times faster than './target/release/main pip-sync ./scripts/benchmarks/requirements.txt --no-cache'
```
```text
+ TARGET=./scripts/benchmarks/requirements-large.txt
+ hyperfine --runs 100 --warmup 10 --prepare 'virtualenv --clear .venv' './target/release/main pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache' --prepare 'virtualenv --clear .venv' './target/release/puffin pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache'
Benchmark 1: ./target/release/main pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache
Time (mean ± σ): 5.053 s ± 0.354 s [User: 1.413 s, System: 6.710 s]
Range (min … max): 4.584 s … 6.333 s 100 runs
Benchmark 2: ./target/release/puffin pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache
Time (mean ± σ): 3.845 s ± 0.225 s [User: 1.364 s, System: 6.970 s]
Range (min … max): 3.482 s … 4.715 s 100 runs
Summary
'./target/release/puffin pip-sync ./scripts/benchmarks/requirements-large.txt --no-cache' ran
```
## Summary
When installing a local wheel, we need to avoid removing the zipped
wheel (since it lives outside of the cache), _and_ need to ensure that
we unzip the wheel into the cache (rather than replacing the zipped
wheel, which may even live outside of the project).
Closes https://github.com/astral-sh/puffin/issues/553.
This PR modifies the source distribution building to replace any
existing targets after building the new wheel. In some cases, the
existence of an existing target may be indicative of a bug, so we warn.
It's partially a workaround for some (but not all) of the errors in
https://github.com/astral-sh/puffin/issues/554.
## Summary
This is hard to reproduce, but if you run a long installation process
that errors part-way through, you can end up with zipped wheels in the
`Wheels` cache, which is intended to contain only unzipped wheels. This
PR avoids returning those entries from the registry, which will then
lead to errors downstream when we treat them as directories.
After this change, two wheel caches remain: `built-wheels-v0` and
`wheels-v0`, docs screenshots below. Each contains both the wheel
metadata, cache policy and zip or unzipped wheels under the same name.
The zipped/unzipped strategy is as follows: In `pip-compile`, when we
build a wheel, we store it zipped. When `pip-sync` or a source dist
build in `pip-compile` need to install the wheel, we unzip it, remove
the file and replace it with the unzipped wheel.
This removes `WheelCache` and `UrlIndex` in favor of `Cache` plus
`WheelCache`. The non-built wheel cache now considers index urls and the
url for url wheels.
I'm unsure if we need the `Unzipper` type, this could just be a
function.
I move `no_index` into `IndexUrls` and started using `IndexUrl` up to
the clap level.
I left a number of TODOs in the code, namely performing the actual
invalidation of unzipped wheels and making the `InstallPlan` understand
cache invalidation (i.e. uninstall wheels when their remote changed).
Remove built wheels alongside their metadata when their index source
dist or url source dist changed. For git source dists, we currently
don't clear the previous build but use a new directory (not sure what's
right here - are there any generic cache GC approaches out there? I've
seen that e.g. spotify keeps its cache at 10GB max, but i also haven't
seen any reusable, well tested approaches for this). Path distributions
are unchanged (#478).
I like the structure of metadata alongside the wheel for cache
invalidation, i'll try to do that for `wheels-v0`/`wheel-metadata-v0`
too. (The unzipped wheels afaik currently lack cache invalidation when
the remote changed.) This should give is roughly the same structure for
wheel and built wheels and a very similar pattern of invalidation.
This is mostly a mechanical refactor that moves 80% of our code to the
same cache abstraction.
It introduces cache `Cache`, which abstracts away the path of the cache
and the temp dir drop and is passed throughout the codebase. To get a
specific cache bucket, you need to requests your `CacheBucket` from
`Cache`. `CacheBucket` is the centralizes the names of all cache
buckets, moving them away from the string constants spread throughout
the crates.
Specifically for working with the `CachedClient`, there is a
`CacheEntry`. I'm not sure yet if that is a strict improvement over
`cache_dir: PathBuf, cache_file: String`, i may have to rotate that
later.
The interpreter cache moved into `interpreter-v0`.
We can use the `CacheBucket` page to document the cache structure in
each bucket:
## Summary
We need to pass in the distribution with the "precise" URL to avoid
refetching.
## Test Plan
Ran `cargo run -p puffin-cli -- pip-compile requirements.in --verbose`
with `flask @ git+https://github.com/pallets/flask.git` and verified
that we only checked out Flask once.
## Summary and motivation
For a given source dist, we store the metadata of each wheel built
through it in `built-wheel-metadata-v0/pypi/<source dist
filename>/metadata.json`. During resolution, we check the cache status
of the source dist. If it is fresh, we check `metadata.json` for a
matching wheel. If there is one we use that metadata, if there isn't, we
build one. If the source is stale, we build a wheel and override
`metadata.json` with that single wheel. This PR thereby ties the local
built wheel metadata cache to the freshness of the remote source dist.
This functionality is available through `SourceDistCachedBuilder`.
`puffin_installer::Builder`, `puffin_installer::Downloader` and
`Fetcher` are removed, instead there are now `FetchAndBuild` which calls
into the also new `SourceDistCachedBuilder`. `FetchAndBuild` is the new
main high-level abstraction: It spawns parallel fetching/building, for
wheel metadata it calls into the registry client, for wheel files it
fetches them, for source dists it calls `SourceDistCachedBuilder`. It
handles locks around builds, and newly added also inter-process file
locking for git operations.
Fetching and building source distributions now happens in parallel in
`pip-sync`, i.e. we don't have to wait for the largest wheel to be
downloaded to start building source distributions.
In a follow-up PR, I'll also clear built wheels when they've become
stale.
Another effect is that in a fully cached resolution, we need neither zip
reading nor email parsing.
Closes#473
## Source dist cache structure
Entries by supported sources:
* `<build wheel metadata cache>/pypi/foo-1.0.0.zip/metadata.json`
* `<build wheel metadata
cache>/<sha256(index-url)>/foo-1.0.0.zip/metadata.json`
* `<build wheel metadata
cache>/url/<sha256(url)>/foo-1.0.0.zip/metadata.json`
But the url filename does not need to be a valid source dist filename
(<https://github.com/search?q=path%3A**%2Frequirements.txt+master.zip&type=code>),
so it could also be the following and we have to take any string as
filename:
* `<build wheel metadata
cache>/url/<sha256(url)>/master.zip/metadata.json`
Example:
```text
# git source dist
pydantic-extra-types @ git+https://github.com/pydantic/pydantic-extra-types.git
# pypi source dist
django_allauth==0.51.0
# url source dist
werkzeug @ ff1904eb5e2853bf83db817a7dd53d/werkzeug-3.0.1.tar.gz
```
will be stored as
```text
built-wheel-metadata-v0
├── git
│ └── 5c56bc1c58c34c11
│ └── 843b753e9e8cb74e83cac55598719b39a4d5ef1f
│ └── metadata.json
├── pypi
│ └── django-allauth-0.51.0.tar.gz
│ └── metadata.json
└── url
└── 6781bd6440ae72c2
└── werkzeug-3.0.1.tar.gz
└── metadata.json
```
The inside of a `metadata.json`:
```json
{
"data": {
"django_allauth-0.51.0-py3-none-any.whl": {
"metadata-version": "2.1",
"name": "django-allauth",
"version": "0.51.0",
...
}
}
}
```
Charlie Marsh
Zanie Blue
konsti
bojanserafimov