site.py imports relatively large `sysconfig` module.

[methane]

BPO	29585
Nosy	@malemburg, @gpshead, @vstinner, @tiran, @ned-deily, @merwok, @methane
PRs	bpo-29585: optimize site.py startup time #136 bpo-29585: Fix Windows build failure #2476 bpo-29585: Define PYTHONFRAMEWORK in PC/pyconfig.h #2477 Revert "bpo-29585: optimize site.py startup time" #2478 bpo-29585: Fix sysconfig.get_config_var("PYTHONFRAMEWORK") #2483 bpo-28095: Re-enable temporarily disabled part of test_startup_import… #2927 bpo-29585: fix test fail on macOS Framework build #2928

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2017-07-28.12:35:27.441>
created_at = <Date 2017-02-17.09:57:03.001>
labels = ['3.7', 'library', 'performance']
title = 'site.py imports relatively large `sysconfig` module.'
updated_at = <Date 2017-07-28.12:35:27.441>
user = 'https://github.com/methane'

bugs.python.org fields:

activity = <Date 2017-07-28.12:35:27.441>
actor = 'methane'
assignee = 'none'
closed = True
closed_date = <Date 2017-07-28.12:35:27.441>
closer = 'methane'
components = ['Library (Lib)']
creation = <Date 2017-02-17.09:57:03.001>
creator = 'methane'
dependencies = []
files = []
hgrepos = []
issue_num = 29585
keywords = []
message_count = 22.0
messages = ['287981', '287983', '287984', '287985', '287988', '287990', '287997', '287999', '288000', '288001', '288012', '288020', '288057', '297192', '297194', '297197', '297258', '299368', '299371', '299379', '299380', '299383']
nosy_count = 7.0
nosy_names = ['lemburg', 'gregory.p.smith', 'vstinner', 'christian.heimes', 'ned.deily', 'eric.araujo', 'methane']
pr_nums = ['136', '2476', '2477', '2478', '2483', '2927', '2928']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue29585'
versions = ['Python 3.7']

[methane]

site.py uses sysconfig (and sysconfigdata, _osx_support) module for user-site package.

But sysconfig module is not so lightweight, and very rarely used.
Actually speaking, only tests and distutils uses sysconfig in stdlibs.

And it takes about 7% of startup time, only for searching user-site path.

I tried to port minimal subset of sysconfig into site.py (GH-136).
But 'PYTHONFRAMEWORK' is only in sysconfigdata. So I couldn't get rid sysconfig dependency completely.

How can I do to solve this?

a) Drop "osx_framework_user" (~/Library/Python/3.7/) support completely.
b) Add "sys._osx_framework" attribute
c) Create minimal sysconfigdata only for site.py
d) anything else?

[vstinner]

Instead of using slow sysconfig and loading the big _sysconfig_data dictionary in memory, would it be possible to extract the minimum set of sysconfig needed by the site module and put it in a builtin module? In site.py, I only found 4 variables:

    from sysconfig import get_config_var
    USER_BASE = get_config_var('userbase')

    from sysconfig import get_path
            USER_SITE = get_path('purelib', 'osx_framework_user')
    USER_SITE = get_path('purelib', '%s_user' % os.name)

            from sysconfig import get_config_var
            framework = get_config_var("PYTHONFRAMEWORK")

Because of the site module, the _sysconfig_data module dictionary is always loaded in memory even for for a dummy print("Hello World!").

I suggest to start building a _site builtin module: subset of site.py which would avoid sysconfig and reimplement things in C for best performances.

speed.python.org:

• python_startup: 14 ms
• python_startup_nosite: 8 ms

Importing site takes 6 ms: 42% of 14 ms...

I'm interested to know if it would be possible to reduce these 6 ms by rewriting some parts of site.py in C.

[vstinner]

Serhiy collected interesting numbers, copy/paste of this message:
http://bugs.python.org/issue28637#msg280380

On my computer:

Importing empty module: 160 us
Creating empty class: 30 us
Creating empty function: 0.16 us
Creating empty Enum/IntEnum: 125/150 us
Creating Enum/IntEnum member: 25/27 us
Creating empty namedtuple: 600 us
Creating namedtuple member: 50 us
Importing the itertools module: 40 us
Importing the io module: 900 us
Importing the os module: 1600 us
Importing the functools module: 2100 us
Importing the re module (with all sre submodules): 3300 us
Python startup time: 43000 us

[tiran]

What's your platform, Inada? Are you running macOS? I optimized site.py for Linux and BSD users a couple of years ago.

[methane]

Christian: I'm using macOS on office and Linux on home.

sysconfig is imported even on Linux
https://github.com/python/cpython/blob/master/Lib/site.py#L247-L248
https://github.com/python/cpython/blob/master/Lib/site.py#L263-L271

[malemburg]

I don't think rewriting party of site.py in C is a good idea. It's a rather maintenance intense module.

However, optimizing access is certainly something that's possible, e.g. by placing the few variables that are actually needed by site.py into a bootstrap module for sysconfig, which only contains the few variables needed by interpreter startup.

Alternatively, sysconfig data could be made available via a C lookup function; with the complete dictionary only being created on demand. get_config_var() already is such a lookup API which could be used as front-end.

[vstinner]

Marc-Andre Lemburg added the comment:

I don't think rewriting party of site.py in C is a good idea. It's a rather maintenance intense module.

However, optimizing access is certainly something that's possible, e.g. by placing the few variables that are actually needed by site.py into a bootstrap module for sysconfig, which only contains the few variables needed by interpreter startup.

Right, I don't propose to rewrite the 598 lines of site.py in C, but
only rewrite the parts which have a huge impact on the startup time.
It seems like the minimum part would be to write a _site module which
provide the 4 variables currently read from sysconfig.

I'm proposing to add a new private module because I don't want to
pollute site which already contains too many things.

I looked at site.py history: I don't see *major* changes last 2 years.
Only small enhancements, updates and fixes.

Alternatively, sysconfig data could be made available via a C lookup function; with the complete dictionary only being created on demand. get_config_var() already is such a lookup API which could be used as front-end.

I don't think that it's worth it to reimplement partially sysconfig in
C. This module is huge, complex, and platform dependant.

Well, I'm not sure about what is the best approach, but I'm sure that
we can do something to optimize site.py. 6 ms is a lot!

I never liked site.py. It seems like a huge workaround. I also dislike
having a different behaviour if site is imported or not.

That's why I asked Steve Dower to removing the code to create the
cpXXX alias for the mbcs codec from site.py to encodings/init.py:
see commit f5aba58. I'm happy that
this code was removed from site.py!

[tiran]

Instead of _site, would it make sense to include the four vars in sys, perhaps as named structure like sys.flags?

[methane]

FYI, here is profile of site:
https://gist.github.com/methane/1f1fe4385dad84f03eb429359f0f917b

[methane]

no-site
python_startup_no_site: Median +- std dev: 9.13 ms +- 0.02 ms

default:
python_startup: Median +- std dev: 15.6 ms +- 0.0 ms

GH-136 + skip abs_paths().
python_startup: Median +- std dev: 14.2 ms +- 0.0 ms

profile of GH-136 + skip abs_paths():
https://gist.github.com/methane/26fc0a2382207655a6819a92f867620c

Most of time is consumed by importlib.

[malemburg]

On 17.02.2017 13:06, STINNER Victor wrote:

> Alternatively, sysconfig data could be made available via a C lookup function; with the complete dictionary only being created on demand. get_config_var() already is such a lookup API which could be used as front-end.

I don't think that it's worth it to reimplement partially sysconfig in
C. This module is huge, complex, and platform dependant.

Sorry, I was just referring to the data part of sysconfig,
not sysconfig itself.

Having a lookup function much like we have for unicodedata
makes things much more manageable, since you don't need to
generate a dictionary in memory for all the values in the
config data. Creating that dictionary takes a while (in terms
of ms).

[methane]

I create bpo-29592 for abs_paths(). Let's focus on sysconfig in this issue.

PR 136 ports really needed part of sysconfig into site.py already.
'PYTHONFRAMEWORK' on macOS is the only variable we need import from sysconfig.

Adding site.cfg like pyvenv.cfg make sense?