...

I haven't created pull request yet. (Can I create without issue, as trivial patch?)

Trivial, no-issue PRs are meant for things like typo fixes that need no discussion or record.

Moving imports in violation of the PEP 8 rule, "Imports are always put at the top of the file, just after any module comments and docstrings, and before module globals and constants", is not trivial. Doing so voluntarily for speed, as opposed to doing so necessarily to avoid circular import errors, is controversial.

-- Terry Jan Reedy

Make sense. I'll create issues for each module if it seems really worth enough. Thanks,

On Tue, Sep 5, 2017 at 11:13 AM, Jelle Zijlstra <jelle.zijlstra@gmail.com> wrote:

2017-09-05 6:02 GMT-07:00 INADA Naoki <songofacandy@gmail.com>:

...

With this profile, I tried optimize `python -c 'import asyncio'`, logging and http.client.

https://gist.github.com/methane/1ab97181e74a33592314c7619bf34233#file-0-opti...

This patch moves a few imports inside functions. I wonder whether that kind of change actually helps with real applications—doesn't any real application end up importing the socket module anyway at some point?

I don't know if this particular change is worthwhile, but one place where startup slowness is particularly noticed is with commands like 'foo.py --help' or 'foo.py --generate-completions' (the latter called implicitly by hitting <tab> in some shell), which typically do lots of imports that end up not being used. -n -- Nathaniel J. Smith -- https://vorpus.org

On Wed, Sep 6, 2017 at 2:30 PM, INADA Naoki <songofacandy@gmail.com> wrote:

...

...

This patch moves a few imports inside functions. I wonder whether that kind of change actually helps with real applications—doesn't any real application end up importing the socket module anyway at some point?

I don't know if this particular change is worthwhile, but one place where startup slowness is particularly noticed is with commands like 'foo.py --help' or 'foo.py --generate-completions' (the latter called implicitly by hitting <tab> in some shell), which typically do lots of imports that end up not being used.

Yes. And There are more worse scenario.

1. Jinja2 supports asyncio. So it imports asyncio. 2. asyncio imports concurrent.futures, for compatibility with Future class. 3. concurrent.futures package does `from concurrent.futures.process import ProcessPoolExecutor` 4. concurrent.futures.process package imports multiprocessing.

So when I use Jinja2 but not asyncio or multiprocessing, I need to import large dependency tree. I want to make `import asyncio` dependency tree smaller.

FYI, current version of Jinja2 has very large regex which took more than 100ms when import time. It is fixed in master branch. So if you try to see Jinja2, please use master branch.

How significant is application startup time to something that uses Jinja2? Are there short-lived programs that use it? Python startup time matters enormously to command-line tools like Mercurial, but far less to something that's designed to start up and then keep running (eg a web app, which is where Jinja is most used). ChrisA

On Wednesday, September 6, 2017, INADA Naoki <songofacandy@gmail.com> wrote:

...

How significant is application startup time to something that uses Jinja2? Are there short-lived programs that use it? Python startup time matters enormously to command-line tools like Mercurial, but far less to something that's designed to start up and then keep running (eg a web app, which is where Jinja is most used).

Since Jinja2 is very popular template engine, it is used by CLI tools like ansible.

SaltStack uses Jinja2. It really is a good idea to regularly restart the minion processes. Celery can also cycle through worker processes, IIRC.

Additionally, faster startup time (and smaller memory footprint) is good for even Web applications. For example, CGI is still comfortable tool sometimes. Another example is GAE/Python.

Short-lived processes are sometimes preferable from a security standpoint. Python is currently less viable for CGI use than other scripting languages due to startup time. Resource leaks (e.g. memory, file handles, database references; valgrind) do not last w/ short-lived CGI processes. If there's ASLR, that's also harder. Scale up operations with e.g. IaaS platforms like Kubernetes and PaaS platforms like AppScale all incur Python startup time on a regular basis.

Anyway, I think researching import tree of popular library is good startline about optimizing startup time. For example, modules like ast and tokenize are imported often than I thought.

Jinja2 is one of libraries I often use. I'm checking other libraries like requests.

Thanks,

INADA Naoki <songofacandy@gmail.com <javascript:;>> _______________________________________________ Python-Dev mailing list Python-Dev@python.org <javascript:;> https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/ wes.turner%40gmail.com

On Sep 6, 2017, at 00:42, INADA Naoki <songofacandy@gmail.com> wrote:

Additionally, faster startup time (and smaller memory footprint) is good for even Web applications. For example, CGI is still comfortable tool sometimes. Another example is GAE/Python.

Anyway, I think researching import tree of popular library is good startline about optimizing startup time. For example, modules like ast and tokenize are imported often than I thought.

Improving start up time may indeed help long running processes but start up costs will generally be amortized across the lifetime of the process, so it isn’t as noticeable. However, startup time *is* a real issue for command line tools. I’m not sure however whether burying imports inside functions (as a kind of poor man’s lazy import) is ultimately going to be satisfying. First, it’s not natural, it generally violates coding standards (e.g. PEP 8), and can make linters complain. Second, I think you’ll end up chasing imports all over the stdlib and third party modules in any sufficiently complicated application. Third, I’m not sure that the gains you’ll get won’t just be overwhelmed by lots of other things going on, such as pkg_resources entry point processing, pth file processing, site.py effects, command line processing libraries such as click, and implicitly added distribution exception hooks (e.g. Ubuntu’s apport). Many of these can’t be blamed on Python itself, but all can contribute significantly to Python’s apparent start up time. It’s definitely worth investigating the details of Python import, and a few of us at the core sprint have looked at those numbers and thrown around ideas for improvement, but we’ll need to look at the effects up and down the stack to improve the start up performance for the average Python application. Cheers, -Barry

I should mention that I have a prototype design for improving importlib's lazy loading to be easier to turn on and use. See https://notebooks.azure.com/Brett/libraries/di2Btqj7zSI/html/Lazy%20importin... for my current notes. Part of it includes an explicit lazy_import() function which would negate needing to hide imports in functions to delay their importation. On Wed, 6 Sep 2017 at 20:50 INADA Naoki <songofacandy@gmail.com> wrote:

...

I’m not sure however whether burying imports inside functions (as a kind of poor man’s lazy import) is ultimately going to be satisfying. First, it’s not natural, it generally violates coding standards (e.g. PEP 8), and can make linters complain.

Of course, I tried to move imports only when (1) it's only used one or two of many functions in the module, (2) it's relatively heavy, (3) it's rerely imported from other modules.

...

Second, I think you’ll end up chasing imports all over the stdlib and third party modules in any sufficiently complicated application.

Agree. I won't use much time to optimization by moving import from top to inner function in stdlib.

I think my import-profiler patch can be polished and committed in Python to help library maintainers to know import time easily. (Maybe, `python -X import-profile`)

...

Third, I’m not sure that the gains you’ll get won’t just be overwhelmed by lots of other things going on, such as pkg_resources entry point processing, pth file processing, site.py effects, command line processing libraries such as click, and implicitly added distribution exception hooks (e.g. Ubuntu’s apport).

Yes. I noticed some of them while profiling imports. For example, old-style namespace package imports types module for types.Module. Types module imports functools, and functools imports collections. So some efforts in CPython (Avoid importing collections and functools from site) is not worth enough when there are at least one old-style namespace package is installed.

...

Many of these can’t be blamed on Python itself, but all can contribute

significantly to Python’s apparent start up time. It’s definitely worth investigating the details of Python import, and a few of us at the core sprint have looked at those numbers and thrown around ideas for improvement, but we’ll need to look at the effects up and down the stack to improve the start up performance for the average Python application.

...

Yes. I totally agree with you. That's why I use import-profile.patch for some 3rd party libraries.

Currently, I have these ideas to optimize application startup time.

* Faster, or lazily compiling regular expression. (pkg_resources imports pyparsing, which has lot regex) * More usable lazy import. (which can be solved "PEP 549: Instance Properties (aka: module properties)") * Optimize enum creation. * Faster namedtuple (There is pull request already) * Faster ABC * Breaking large import tree in stdlib. (PEP 549 may help this too)

Regards,

INADA Naoki <songofacandy@gmail.com> _______________________________________________ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/brett%40python.org