Use `setup.py` directly rather than `pip install` #203

mdboom · 2015-01-30T03:27:16Z

This results in building the project in the same directory every time, giving ccache a better chance of doing its thing. An alternate solution to #201.

pip install /some/local/path has been used in asv to install the benchmarked project on the assumption that pip knows better about the details of how to install a Python package beyond just the usual setup.py install. However, I wonder if that's really the case...

Note that I tried to use the --build flag for pip install to force it to build in a specific directory. This seems to work for remote packages, but it's ignored when given a local path...

Attn: @pv

pv · 2015-01-30T14:57:52Z

I think pip refuses to uninstall packages installed by setup.py install --- at least for the case when setup.py is not using setuptools.

Perhaps just running python setup.py build; pip install . will work (or just build_ext instead of build)?
At least for Numpy, this appears to skip the build completely in the second step, so ccache has the opportunity to work with non-random paths.

pv · 2015-01-30T14:59:58Z

The second question is then what to do with building dependencies --- if the environment in asv is not torn down every time, using ccache for those is probably not so important?

mdboom · 2015-01-30T15:16:33Z

I think pip refuses to uninstall packages installed by setup.py install --- at least for the case when setup.py is not using setuptools.

Perhaps just running python setup.py build; pip install . will work (or just build_ext instead of build)?

At least experimentally on a dummy package that uses only vanilla distutils and no setuptools, this appears to be true. python setup.py install makes something that isn't pip uninstall-able. python setup.py build followed by pip install ./project works, though. That seems like a good solution, then.

As for the dependencies, you are correct that they are still being built with random paths, therefore ccache doesn't work. But they are cached in the env directory tree and don't need to be built each time, so it's less of an issue. If it were an issue, we could pass --build ./some-static-tmp-directory to pip when building the dependencies, since --build does work when installing from downloaded packages, just not when installing from the local filesystem.

pv · 2015-02-01T02:08:49Z

This works fine for me. Does it break the parallel build feature however?

mdboom · 2015-02-01T02:58:55Z

I does break the parallel build feature. For a couple of other corner case reasons, I've been meaning to fix the parallel build feature so that each parallel thread would use its own git clone anyway, so this might just be onus to do that...

mdboom · 2015-02-02T23:10:20Z

This now makes a separate clone of the project repository for each environment (these are --shared clones, so their .git directories don't take up much space). That way parallel building can work without clobbering another environment's build. This also includes @pv's wheel caching support, so the last N builds are retained and can be reinstalled next time much faster.

The combination of parallel building, ccache working and cached wheels makes things much faster overall. Too bad there isn't a tool to measure that ;)

project. This gives ccache a better chance of doing its thing.

Conflicts: asv/benchmarks.py asv/commands/find.py asv/commands/profiling.py asv/commands/run.py asv/environment.py asv/plugins/conda.py asv/plugins/virtualenv.py

pv · 2015-02-04T20:45:57Z

Seems to break asv dev, cf pv/asv@1addad0
asv dev also seems to unnecessarily clone the repo.

pv · 2015-02-04T20:48:41Z

asv/wheel_cache.py

+        if self._wheel_cache_size == 0:
+            return None
+
+        if not os.path.isdir(self._path):


This check should be moved --- the directory is first created by the call to ~~_get_wheel~~ _create_wheel_cache_path below.
My logic here with _get_wheel_cache_path creating the directory was a bit funky...
pv/asv@6762b7dd29f

pv · 2015-02-04T21:16:00Z

Worksforme apart from those two

mdboom · 2015-02-04T22:09:20Z

Thanks. I think I've addressed those two issues.

pv · 2015-02-05T00:50:52Z

The isdir check probably needs to be added to _get_wheel pv/asv@6762b7d ?

mdboom · 2015-02-05T01:22:22Z

Indeed. I think it's fixed now.

pv · 2015-02-05T01:25:36Z

asv/wheel_cache.py

+    def _get_wheel(self, commit_hash):
+        cache_path = self._get_wheel_cache_path(commit_hash)
+
+        if not os.isdir(cache_path):


os.path.isdir

pv · 2015-02-07T01:41:21Z

asv/console.py

+        orig_level = self._logger.level
+        self._logger.setLevel(level)
+        yield
+        self._logger.setLevel(orig_level)


This doesn't set the level back if an exception is raised (eg build failure), may be related to gh-209

pv/asv@044b8d6f9c

Thanks. I've cherry-picked 044b8d6 over to this branch.

I've been using this PR for the Scipy benchmarks, and didn't meet other problems than what is fixed above.

…manager

Use `setup.py` directly rather than `pip install`

mdboom mentioned this pull request Jan 30, 2015

Add ccache plugin #202

Closed

mdboom mentioned this pull request Feb 2, 2015

Cache build results in Python wheels for faster installation #206

Closed

mdboom force-pushed the install-same-place branch from f931397 to ad4b91d Compare February 2, 2015 23:07

mdboom and others added 9 commits February 3, 2015 14:36

Use setup.py directly rather than pip install for benchmarked

0a7ee16

project. This gives ccache a better chance of doing its thing.

Build with setup.py, then install with pip

902720d

Pass cwd to subprocess rather than manually using chdir

c889a73

Use separate checkouts for each environment

3317ab7

Cache build results in Python wheels for faster installation

e53ed7e

Conflicts: asv/benchmarks.py asv/commands/find.py asv/commands/profiling.py asv/commands/run.py asv/environment.py asv/plugins/conda.py asv/plugins/virtualenv.py

Resolve the wheel cache stuff with the environment-local-building stuff

a1f6b62

Cleanups

82a015b

Add docs.

bee5aaf

Upgrade virtualenv

201042a

mdboom force-pushed the install-same-place branch from 00427ec to 201042a Compare February 3, 2015 19:37

More test fixes

56e2c30

pv reviewed Feb 4, 2015
View reviewed changes

mdboom added 2 commits February 4, 2015 17:08

Fix asv dev

6182b49

Remove unnecessary checkc

3e94fe9

pv reviewed Feb 5, 2015
View reviewed changes

Fix logic about missing wheel directory.

cf95250

mdboom force-pushed the install-same-place branch from 46a9ffa to cf95250 Compare February 5, 2015 13:43

pv mentioned this pull request Feb 7, 2015

Failed build messes up printing #209

Closed

pv reviewed Feb 7, 2015
View reviewed changes

pv mentioned this pull request Feb 7, 2015

Replace benchmark suite with airspeed velocity scipy/scipy#4501

Merged

mdboom and others added 2 commits February 9, 2015 10:17

Test with wheel_cache_size

09fe1ce

Fix bug in restoring logging level if exception is raised in context …

ba1fbb8

…manager

mdboom mentioned this pull request Feb 9, 2015

Add mercurial support #193

Closed

mdboom added a commit that referenced this pull request Feb 9, 2015

Merge pull request #203 from mdboom/install-same-place

77a2944

Use `setup.py` directly rather than `pip install`

mdboom merged commit 77a2944 into airspeed-velocity:master Feb 9, 2015

mdboom mentioned this pull request Feb 9, 2015

pip confuses ccache #201

Closed

pv mentioned this pull request Jul 24, 2017

--build-dir is not respected when installing from an sdist without a req pypa/pip#4242

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use `setup.py` directly rather than `pip install` #203

Use `setup.py` directly rather than `pip install` #203

mdboom commented Jan 30, 2015

pv commented Jan 30, 2015

pv commented Jan 30, 2015

mdboom commented Jan 30, 2015

pv commented Feb 1, 2015

mdboom commented Feb 1, 2015

mdboom commented Feb 2, 2015

pv commented Feb 4, 2015

pv Feb 4, 2015

pv commented Feb 4, 2015

mdboom commented Feb 4, 2015

pv commented Feb 5, 2015

mdboom commented Feb 5, 2015

pv Feb 5, 2015

pv Feb 7, 2015

pv Feb 7, 2015

mdboom Feb 9, 2015

pv Feb 9, 2015 via email

Use setup.py directly rather than pip install #203

Use setup.py directly rather than pip install #203

Conversation

mdboom commented Jan 30, 2015

pv commented Jan 30, 2015

pv commented Jan 30, 2015

mdboom commented Jan 30, 2015

pv commented Feb 1, 2015

mdboom commented Feb 1, 2015

mdboom commented Feb 2, 2015

pv commented Feb 4, 2015

pv Feb 4, 2015

Choose a reason for hiding this comment

pv commented Feb 4, 2015

mdboom commented Feb 4, 2015

pv commented Feb 5, 2015

mdboom commented Feb 5, 2015

pv Feb 5, 2015

Choose a reason for hiding this comment

pv Feb 7, 2015

Choose a reason for hiding this comment

pv Feb 7, 2015

Choose a reason for hiding this comment

mdboom Feb 9, 2015

Choose a reason for hiding this comment

pv Feb 9, 2015 via email

Choose a reason for hiding this comment

Use `setup.py` directly rather than `pip install` #203

Use `setup.py` directly rather than `pip install` #203