369-getting-lazy-with-python-imports-and-pep-690.vtt

WEBVTT

00:00:00.000 --> 00:00:04.000
- Hello YouTube, hello Barry, Erman, Carl.


00:00:04.000 --> 00:00:04.960
Welcome to the show.


00:00:04.960 --> 00:00:07.400
- Hey, hello.


00:00:07.400 --> 00:00:10.520
- It's wonderful to have you all here.


00:00:10.520 --> 00:00:14.800
I'm very excited about the work that you're all doing


00:00:14.800 --> 00:00:17.160
around Python performance.


00:00:17.160 --> 00:00:19.520
We're gonna focus on imports and this pep


00:00:19.520 --> 00:00:21.760
that you three proposed today,


00:00:21.760 --> 00:00:24.520
but it's really just the tip of the iceberg


00:00:24.520 --> 00:00:26.840
in terms of a bunch of cool stuff that's going on.


00:00:26.840 --> 00:00:28.720
So I'm very excited to dive into that


00:00:28.720 --> 00:00:29.840
with the three of you.


00:00:29.840 --> 00:00:31.300
Now, before we get to it though,


00:00:31.300 --> 00:00:33.700
let's just do a quick round of introductions.


00:00:33.700 --> 00:00:35.300
Barry, you've been on the show before


00:00:35.300 --> 00:00:40.300
talking about 1994 Python stuff and other things.


00:00:40.300 --> 00:00:44.380
So maybe just a quick introduction for yourself.


00:00:44.380 --> 00:00:46.860
- Yeah, Barry Warsaw.


00:00:46.860 --> 00:00:50.900
So I'm still here, still hanging around, I guess.


00:00:50.900 --> 00:00:56.700
You know, I should mention, you know,


00:00:56.700 --> 00:00:59.120
as we get into this PEP though,


00:00:59.120 --> 00:01:00.720
I'm really just the sponsor.


00:01:00.720 --> 00:01:04.920
I gotta hand a lot of thanks to Hermon and Carl


00:01:04.920 --> 00:01:06.360
for doing all the work on it.


00:01:06.360 --> 00:01:13.120
Most peps these days require a core developer to sponsor.


00:01:13.120 --> 00:01:17.280
So I'm just super interested in the topic.


00:01:17.280 --> 00:01:19.560
I think it's a super clever approach.


00:01:19.560 --> 00:01:24.320
And so I think it'll help specific needs


00:01:24.320 --> 00:01:25.160
that I have at work.


00:01:25.160 --> 00:01:27.760
And so I was really eager to...


00:01:27.760 --> 00:01:31.640
eager to sponsor it.


00:01:31.640 --> 00:01:34.640
- Yeah, fantastic.


00:01:34.640 --> 00:01:36.440
All right, very cool.


00:01:36.440 --> 00:01:37.600
Hermann, how about you?


00:01:37.600 --> 00:01:45.040
- Well, I'm just working at Meta with Carl


00:01:45.040 --> 00:01:50.040
and I wrote the initial approach for Lazy Imports


00:01:50.040 --> 00:01:52.020
and well, here I am.


00:01:52.020 --> 00:01:54.320
- Right on.


00:01:56.080 --> 00:01:58.340
- Carl, quick introduction.


00:01:58.340 --> 00:02:04.440
- I've been around the Python community for a while.


00:02:04.440 --> 00:02:08.080
I think the first maybe semi-notable thing that I did


00:02:08.080 --> 00:02:12.760
was write the first version of pip uninstall back in 2009.


00:02:12.760 --> 00:02:16.040
That led to being a maintainer of pip and virtualen


00:02:16.040 --> 00:02:19.840
for a while, and I worked on the Django core team


00:02:19.840 --> 00:02:21.080
for a while.


00:02:21.080 --> 00:02:24.320
And so yeah, I've been doing Python things for a long time


00:02:24.320 --> 00:02:25.840
and I've been working at Meta since--


00:02:25.840 --> 00:02:27.840
- Those are a lot of big projects there.


00:02:27.840 --> 00:02:30.600
- Yeah, they're projects that I was using


00:02:30.600 --> 00:02:32.560
and so I was interested in working on them


00:02:32.560 --> 00:02:34.500
and same goes for Python itself.


00:02:34.500 --> 00:02:36.760
- Yeah, absolutely.


00:02:36.760 --> 00:02:37.880
Sorry, I kind of cut you off there.


00:02:37.880 --> 00:02:40.640
You said working at Meta since 2016, is that what you said?


00:02:40.640 --> 00:02:43.520
- Yeah, I've been working at Meta since 2016.


00:02:43.520 --> 00:02:47.360
Working mostly on how Meta uses Python,


00:02:47.360 --> 00:02:48.680
how Instagram uses Python.


00:02:48.680 --> 00:02:50.800
- Yeah.


00:02:50.800 --> 00:02:53.400
There've been some really cool looks inside what's going on,


00:02:53.400 --> 00:02:57.440
especially at Instagram there with some of the typing talks


00:02:57.440 --> 00:03:02.440
that Lucas gave, as well as sort of suspending


00:03:02.440 --> 00:03:05.160
the garbage collector for other various things.


00:03:05.160 --> 00:03:07.880
And a lot-- or was that you guys?


00:03:07.880 --> 00:03:10.480
I think it was, right?


00:03:10.480 --> 00:03:11.480
Yeah, we had a whole--


00:03:11.480 --> 00:03:13.240
A lot of stuff coming out of there.


00:03:13.240 --> 00:03:13.920
A whole saga.


00:03:13.920 --> 00:03:15.380
We turned off the garbage collector,


00:03:15.380 --> 00:03:16.960
turned back on the garbage collector.


00:03:16.960 --> 00:03:19.840
There's been multiple blog posts along the way


00:03:19.840 --> 00:03:24.240
explaining why we've done each of those silly sounding things.


00:03:24.240 --> 00:03:25.920
>> Sure, I mean, they seem insane.


00:03:25.920 --> 00:03:28.360
Well, a lot of the projects that,


00:03:28.360 --> 00:03:29.880
a lot of the stuff that's coming out here,


00:03:29.880 --> 00:03:32.080
it has to do with the Cinder project, right?


00:03:32.080 --> 00:03:35.200
>> Yeah.


00:03:35.200 --> 00:03:37.160
>> And you may be, I don't know who's best


00:03:37.160 --> 00:03:38.560
to give the introduction for Cinder,


00:03:38.560 --> 00:03:43.560
but Cinder's a really cool project about, you know,


00:03:43.560 --> 00:03:46.280
taking a whole bunch of optimizations


00:03:46.280 --> 00:03:48.160
and specializations you all have done


00:03:48.160 --> 00:03:50.920
and sort of sharing that back with the community


00:03:50.920 --> 00:03:51.920
a little bit, right?


00:03:51.920 --> 00:03:57.960
- Yeah, so we started Cinder in I think 2017 or 2018.


00:03:57.960 --> 00:03:59.680
There was actually two projects


00:03:59.680 --> 00:04:01.080
kind of started simultaneously.


00:04:01.080 --> 00:04:04.280
We realized around that time that the trajectory


00:04:04.280 --> 00:04:09.040
for Instagram's kind of server footprint


00:04:09.040 --> 00:04:11.600
was not really sustainable just in terms of how much


00:04:11.600 --> 00:04:14.640
server CPU time we were spending running Python code.


00:04:14.640 --> 00:04:16.440
And so we kicked off two projects.


00:04:16.440 --> 00:04:19.380
One was called, it's now called Sky Bison.


00:04:19.380 --> 00:04:23.220
It was like a ground up rewritten Python interpreter,


00:04:23.220 --> 00:04:27.140
you know, using all the modern dynamic language VM ideas,


00:04:27.140 --> 00:04:28.580
like a moving garbage collector


00:04:28.580 --> 00:04:30.180
and all these different things.


00:04:30.180 --> 00:04:32.740
We weren't the first people to try that.


00:04:32.740 --> 00:04:36.940
And we also weren't the first people to fail.


00:04:36.940 --> 00:04:39.720
That project was wound down last year,


00:04:39.720 --> 00:04:41.380
just weren't able to get the performance,


00:04:41.380 --> 00:04:43.740
particularly trying to emulate compatibility


00:04:43.740 --> 00:04:46.020
with the C API and all the C extensions,


00:04:46.020 --> 00:04:50.180
which is the same reason many prior rewrite efforts


00:04:50.180 --> 00:04:51.220
didn't go very far.


00:04:51.220 --> 00:04:53.860
So at the same time as we'd kicked off SkyBison,


00:04:53.860 --> 00:04:55.740
we had sort of kicked off what we thought


00:04:55.740 --> 00:04:58.320
was a short-term project of just like,


00:04:58.320 --> 00:05:00.060
what can we squeeze out of CPython?


00:05:00.060 --> 00:05:02.220
Where can we get a little more performance out of it?


00:05:02.220 --> 00:05:04.000
And that turned into Cinder,


00:05:04.000 --> 00:05:06.300
and then ended up kind of becoming our primary approach


00:05:06.300 --> 00:05:07.640
to Python performance.


00:05:07.640 --> 00:05:13.660
- Isn't it always the story that the C interop stuff


00:05:13.660 --> 00:05:16.900
is the big sticking point here.


00:05:16.900 --> 00:05:17.740
- Yeah.


00:05:17.740 --> 00:05:19.100
- I mean, Barry, you must have seen a whole bunch


00:05:19.100 --> 00:05:21.980
of examples from a core dev perspective.


00:05:21.980 --> 00:05:23.140
All right.


00:05:23.140 --> 00:05:23.980
- Yeah.


00:05:23.980 --> 00:05:24.800
- We could just change the guild,


00:05:24.800 --> 00:05:28.580
but the C API, we could just, but you know.


00:05:28.580 --> 00:05:29.620
- Yeah, yeah.


00:05:29.620 --> 00:05:33.820
I mean, it's both, you know, Python's, you know,


00:05:33.820 --> 00:05:35.780
advantage, you know, pros and cons, right?


00:05:35.780 --> 00:05:39.500
Like the approachability and the usability


00:05:39.500 --> 00:05:44.260
the C API has led directly to the incredible ecosystem


00:05:44.260 --> 00:05:47.120
of extension modules, right?


00:05:47.120 --> 00:05:50.980
But those also are also the hindrance for moving,


00:05:50.980 --> 00:05:52.860
you know, ahead in a revolutionary way


00:05:52.860 --> 00:05:55.900
with the interpreter, I think, you know.


00:05:55.900 --> 00:05:57.900
- Right, you kind of got to live within the box,


00:05:57.900 --> 00:06:01.020
with the walls that are put up by those constraints.


00:06:01.020 --> 00:06:02.340
But I do think it's super important.


00:06:02.340 --> 00:06:03.940
You know, a lot of people would maybe just think,


00:06:03.940 --> 00:06:04.820
well, we'll just get rid of it.


00:06:04.820 --> 00:06:06.380
Like, let's just try to move beyond it.


00:06:06.380 --> 00:06:09.800
but when you hear people say that Python is slow


00:06:09.800 --> 00:06:11.380
or it has these other problems,


00:06:11.380 --> 00:06:13.200
so often what you'll see as well,


00:06:13.200 --> 00:06:15.820
and what I did was I did a for loop in Python


00:06:15.820 --> 00:06:18.160
and did some math and that was slow.


00:06:18.160 --> 00:06:19.860
It's like, well, but if a thing is slow,


00:06:19.860 --> 00:06:22.180
so often that gets rewritten in C


00:06:22.180 --> 00:06:23.940
and then all of a sudden it's faster than,


00:06:23.940 --> 00:06:26.180
I don't know what Java or whatever else it is


00:06:26.180 --> 00:06:28.020
they're trying, you know, Node,


00:06:28.020 --> 00:06:29.120
but they're trying to compare it to.


00:06:29.120 --> 00:06:32.320
And so there's this kind of crazy switch


00:06:32.320 --> 00:06:35.800
that gets flipped for like really high performance.


00:06:35.800 --> 00:06:39.920
and then maybe acceptable most of the time performance.


00:06:39.920 --> 00:06:43.920
And it's this interop with C that is the thing,


00:06:43.920 --> 00:06:45.080
that's the escape hatch.


00:06:45.080 --> 00:06:51.200
- Yeah, and LinkedIn, we don't have the same kind


00:06:51.200 --> 00:06:55.080
of workloads that they have at Meta,


00:06:55.080 --> 00:06:58.480
but we've done analysis.


00:06:58.480 --> 00:07:00.480
One of the reasons why I was particularly interested


00:07:00.480 --> 00:07:02.800
in this project is because we write a lot


00:07:02.800 --> 00:07:04.840
of our CLIs in Python.


00:07:04.840 --> 00:07:09.840
And so, we've had proponents of other languages say,


00:07:09.840 --> 00:07:11.920
complain, "Hey, Python's really slow.


00:07:11.920 --> 00:07:13.600
"I write a CLI in Python."


00:07:13.600 --> 00:07:16.080
And then it takes a long time to start up.


00:07:16.080 --> 00:07:18.520
But if you actually do the analysis,


00:07:18.520 --> 00:07:22.520
what you find is that people are not writing,


00:07:22.520 --> 00:07:25.880
a lot of the internal libraries that those CLIs import


00:07:25.880 --> 00:07:27.620
do things like they go hit the network


00:07:27.620 --> 00:07:29.280
and they go try to do the service


00:07:29.280 --> 00:07:31.560
and create really expensive resources


00:07:31.560 --> 00:07:33.080
at module scope time, right?


00:07:33.080 --> 00:07:35.440
So like those are the types of analysis


00:07:35.440 --> 00:07:36.800
that you really need to do to say,


00:07:36.800 --> 00:07:39.600
is it really Python or is it the way we use Python?


00:07:39.600 --> 00:07:42.340
You know, for us, it's a little of both for sure,


00:07:42.340 --> 00:07:46.200
but often it's sort of the way Python is being used


00:07:46.200 --> 00:07:50.080
in a non-idiomatic or not in the highest performance way.


00:07:50.080 --> 00:07:50.920
Right?


00:07:50.920 --> 00:07:52.960
So a little bit of rewrite of some internal code


00:07:52.960 --> 00:07:54.080
can get you a long way.


00:07:54.080 --> 00:07:56.400
- It sure can.


00:07:56.400 --> 00:07:58.840
I also want to just give a quick shout out


00:07:58.840 --> 00:08:02.240
back on episode 347, I talked with Dino Vila


00:08:02.240 --> 00:08:03.900
like all about the Cinder project.


00:08:03.900 --> 00:08:04.840
So if you want to check that out,


00:08:04.840 --> 00:08:07.280
they can definitely go find that.


00:08:07.280 --> 00:08:11.280
But there's a lot of cool things that were taken out of Cinder


00:08:11.280 --> 00:08:12.440
and are being proposed.


00:08:12.440 --> 00:08:17.200
And I guess that sort of brings us to our main topic for today


00:08:17.200 --> 00:08:21.200
is imports and the PEP 690 that you all have proposed,


00:08:21.200 --> 00:08:23.280
which I am also excited for.


00:08:23.280 --> 00:08:26.640
I think imports are one of those things that


00:08:26.640 --> 00:08:29.560
are a little bit mysterious to people


00:08:29.560 --> 00:08:33.240
because they have conceptions coming from other languages,


00:08:33.240 --> 00:08:35.560
especially compiled languages that are very different


00:08:35.560 --> 00:08:37.020
than what the reality is.


00:08:37.020 --> 00:08:41.240
So I do want to talk a whole bunch


00:08:41.240 --> 00:08:42.560
about what you're offering here,


00:08:42.560 --> 00:08:46.720
what you're proposing here in this lazy imports pep.


00:08:46.720 --> 00:08:49.120
But before we do, let's maybe just set the stage


00:08:49.120 --> 00:08:54.080
with what happens when I write import requests


00:08:54.080 --> 00:08:56.040
or I write import FastAPI,


00:08:56.040 --> 00:08:59.800
or even something built in like import collections?


00:08:59.800 --> 00:09:02.720
Who wants to take this one?


00:09:02.720 --> 00:09:03.880
Tell people what really happens


00:09:03.880 --> 00:09:06.440
when the import statement runs, is there.


00:09:06.440 --> 00:09:11.240
- I don't know.


00:09:11.240 --> 00:09:14.420
I guess I'll say,


00:09:14.420 --> 00:09:19.160
there was a time where the import system


00:09:19.160 --> 00:09:21.200
for Python was written in C


00:09:21.200 --> 00:09:23.560
and it was even more mysterious, I think at that point.


00:09:23.560 --> 00:09:26.560
And I don't remember when it was,


00:09:26.560 --> 00:09:31.140
but Brett Cannon rewrote the import system into Python.


00:09:31.140 --> 00:09:32.900
And then about that time,


00:09:32.900 --> 00:09:36.260
I sort of went through sort of line by line


00:09:36.260 --> 00:09:38.760
and tried to understand, I mean, because right,


00:09:38.760 --> 00:09:42.620
like imports are probably one of the oldest features


00:09:42.620 --> 00:09:43.460
of Python, right?


00:09:43.460 --> 00:09:44.340
They've been around so long


00:09:44.340 --> 00:09:47.220
and there's so many corner cases, right?


00:09:47.220 --> 00:09:51.700
It's a very complex system and has lots of features,


00:09:51.700 --> 00:09:56.980
lots of hooks, lots of places where you can hook in your own behavior and really


00:09:56.980 --> 00:09:59.260
understanding how all of that fits together.


00:09:59.260 --> 00:10:04.140
I think I probably took a couple of months to sort of walk through the entire


00:10:04.140 --> 00:10:05.580
system line by line and document it.


00:10:05.580 --> 00:10:11.500
Fortunately now, if you go to the language reference guide, I think it goes into all


00:10:11.500 --> 00:10:14.180
the gory detail about how imports work.


00:10:14.180 --> 00:10:19.660
Everything from namespace packages to, you know, concrete packages and all the file


00:10:19.660 --> 00:10:23.900
system hooks and meta path hooks and blah, blah, blah, blah.


00:10:23.900 --> 00:10:24.400
Right?


00:10:24.400 --> 00:10:26.900
So yeah.


00:10:26.900 --> 00:10:29.940
Yeah, definitely a lot going on.


00:10:29.940 --> 00:10:32.660
I think the key thing, though, that you are hinting at there


00:10:32.660 --> 00:10:37.020
is this is a runtime type of behavior.


00:10:37.020 --> 00:10:39.380
This is a runtime experience.


00:10:39.380 --> 00:10:40.700
Carl, you're shaking your head.


00:10:40.700 --> 00:10:42.500
What would you want to add to this?


00:10:42.500 --> 00:10:44.580
What Barry was saying here.


00:10:44.580 --> 00:10:48.020
Well, yeah, I mean, I think what a lot of people don't--


00:10:48.020 --> 00:10:51.380
I mean, maybe don't consciously realize initially


00:10:51.380 --> 00:10:55.580
about imports is that it's just executing some more code.


00:10:55.580 --> 00:10:58.740
So it's like, you can almost think of imports in Python


00:10:58.740 --> 00:11:00.780
as like syntax sugar for a function call


00:11:00.780 --> 00:11:04.180
where the body of the function is the module level code


00:11:04.180 --> 00:11:07.540
of some module and the return value is the module object


00:11:07.540 --> 00:11:10.080
that you end up getting back from the import.


00:11:10.080 --> 00:11:10.940
But like it really is,


00:11:10.940 --> 00:11:14.580
it's just causing some more code to be executed


00:11:14.580 --> 00:11:17.820
where the result is a module object that has stuff on it


00:11:17.820 --> 00:11:19.100
that you can use.


00:11:19.100 --> 00:11:21.500
And one of the consequences of that


00:11:21.500 --> 00:11:25.860
is that literally anything can happen when you import a module.