Python's sys.meta_path finders are the primary objects whose job it
is to find a module at import time. When import is called, Python
iterates objects in this list and calls o.find_spec(...) to find
a ModuleSpec (or None if the module couldn't be found by that
finder). If no meta path finder can find a module, import fails.
One of the default meta path finders is PathFinder. Its job is to
import modules from the filesystem and is probably the most important
importer. This finder looks at sys.path and sys.path_hooks to do
its job.
The ModuleSpec returned by MetaPathImporter.find_spec() has a
loader attribute, which defines the concrete module loader to use.
sys.path_hooks is a hook point for teaching PathFinder to
instantiate custom loader types.
Previously, we injected a custom sys.path_hook that told PathFinder
to wrap the default loaders with a loader that creates a module object
that is lazy.
This approach worked. But its main limitation was that it only applied
to the PathFinder meta path importer. There are other meta path
importers that are registered. And in the case of PyOxidizer loading
modules from memory, PathFinder doesn't come into play since
PyOxidizer's own meta path importer was handling all imports.
This commit changes our approach to lazy module loading by proxying
all meta path importers. Specifically, we overload the find_spec()
method to swap in a wrapped loader on the ModuleSpec before it
is returned. The end result of this is all meta path importers should
be lazy.
As much as I would have loved to utilize .class manipulation to
achieve this, some meta path importers are implemented in C/Rust
in such a way that they cannot be monkeypatched. This is why we
use getattribute to define a proxy.
Also, this change could theoretically open us up to regressions in
meta path importers whose loader is creating module objects which
can't be monkeypatched. But I'm not aware of any of these in the
wild. So I think we'll be safe.
According to hyperfine, this change yields a decent startup time win of
5-6ms:
Benchmark #1: ~/.pyenv/versions/3.6.10/bin/python ./hg version Time (mean ± σ): 86.8 ms ± 0.5 ms [User: 78.0 ms, System: 8.7 ms] Range (min … max): 86.0 ms … 89.1 ms 50 runs Time (mean ± σ): 81.1 ms ± 2.7 ms [User: 74.5 ms, System: 6.5 ms] Range (min … max): 77.8 ms … 90.5 ms 50 runs Benchmark #2: ~/.pyenv/versions/3.7.6/bin/python ./hg version Time (mean ± σ): 78.9 ms ± 0.6 ms [User: 70.2 ms, System: 8.7 ms] Range (min … max): 78.1 ms … 81.2 ms 50 runs Time (mean ± σ): 73.4 ms ± 0.6 ms [User: 65.3 ms, System: 8.0 ms] Range (min … max): 72.4 ms … 75.7 ms 50 runs Benchmark #3: ~/.pyenv/versions/3.8.1/bin/python ./hg version Time (mean ± σ): 78.1 ms ± 0.6 ms [User: 70.2 ms, System: 7.9 ms] Range (min … max): 77.4 ms … 80.9 ms 50 runs Time (mean ± σ): 72.1 ms ± 0.4 ms [User: 64.4 ms, System: 7.6 ms] Range (min … max): 71.4 ms … 74.1 ms 50 runs