WEBVTT

00:00:00.000 --> 00:00:05.000

- Mark Guido, welcome to Talk Python To Me.

00:00:05.000 --> 00:00:06.340

- Hi, hello.

00:00:06.340 --> 00:00:09.580

- It's fantastic to have you here.

00:00:09.580 --> 00:00:12.000

I'm so excited about all the things

00:00:12.000 --> 00:00:14.340

that are happening around Python performance.

00:00:14.340 --> 00:00:18.000

I feel like there's just a bunch of new ideas springing up

00:00:18.000 --> 00:00:21.660

and people working on it and yeah, it's exciting times.

00:00:21.660 --> 00:00:24.260

- Definitely.

00:00:24.260 --> 00:00:27.640

- Yeah, and you two are of course,

00:00:27.640 --> 00:00:29.120

right at the center of it.

00:00:29.120 --> 00:00:32.040

But before we talk about the performance work

00:00:32.040 --> 00:00:32.880

that you are all doing,

00:00:32.880 --> 00:00:35.800

as well as some of the other initiatives going along,

00:00:35.800 --> 00:00:37.040

maybe in parallel there,

00:00:37.040 --> 00:00:40.720

let's just get started with a little bit of background.

00:00:40.720 --> 00:00:42.200

Guido, you've been on the show before,

00:00:42.200 --> 00:00:44.080

you're a creator of Python,

00:00:44.080 --> 00:00:47.560

you hardly need an introduction to most people out there.

00:00:47.560 --> 00:00:50.960

But you have recently made a couple of big changes

00:00:50.960 --> 00:00:54.160

in your life, I thought I'd just ask you how that's going.

00:00:54.160 --> 00:00:56.780

You retired and we were all super happy for you on that.

00:00:56.780 --> 00:00:58.600

And then you said, you know what,

00:00:58.600 --> 00:01:00.140

I kind of want to play with code some more.

00:01:00.140 --> 00:01:01.360

And now you're at Microsoft.

00:01:01.360 --> 00:01:03.040

What's the story there?

00:01:03.040 --> 00:01:07.160

- Oh, I just liked the idea of retiring.

00:01:07.160 --> 00:01:12.160

So I tried to see how many times in a lifetime I can retire.

00:01:12.160 --> 00:01:15.240

Starting with my retirement from BDFL,

00:01:15.240 --> 00:01:20.920

which didn't stop me from staying super active

00:01:20.920 --> 00:01:21.940

in the community.

00:01:24.760 --> 00:01:29.460

But when I retired from Dropbox a little over two years ago,

00:01:29.460 --> 00:01:33.800

I really thought that that was it, that I believed it.

00:01:33.800 --> 00:01:34.640

- Yeah.

00:01:34.640 --> 00:01:36.200

- And everybody else believed it too.

00:01:36.200 --> 00:01:38.280

Dropbox certainly believed it.

00:01:38.280 --> 00:01:41.580

They were very sad to see me go.

00:01:41.580 --> 00:01:44.080

I was sad to go, but I thought there was time.

00:01:44.080 --> 00:01:49.080

And I sort of, I had a few great months decompressing,

00:01:54.400 --> 00:01:59.400

going on bike rides with my wife and family, fun stuff.

00:01:59.400 --> 00:02:02.600

And then the pandemic hit.

00:02:02.600 --> 00:02:03.440

- Yeah.

00:02:03.440 --> 00:02:05.960

- And a bunch of things got harder.

00:02:05.960 --> 00:02:09.040

The fortunately the bike rides eventually got restored

00:02:09.040 --> 00:02:12.120

but other activities like eating out

00:02:12.120 --> 00:02:14.020

was a lot more stressful.

00:02:14.020 --> 00:02:17.160

Basically just life was a lot more stressful in general.

00:02:17.160 --> 00:02:18.540

- Right, and human interaction

00:02:18.540 --> 00:02:22.060

was definitely strunken down to a kernel.

00:02:22.060 --> 00:02:29.420

Yeah, and I somehow I thought, well, I want to have something to do. I want to

00:02:29.420 --> 00:02:34.780

do more sort of software development in the team.

00:02:34.780 --> 00:02:42.460

And the Python core development team didn't really cut it for me because it's

00:02:42.460 --> 00:02:49.820

sort of diffuse and volunteer-based and sometimes you get stuck waiting for months for the steering

00:02:49.820 --> 00:02:59.180

council to sort of approve of or reject a certain idea that you've worked on.

00:02:59.180 --> 00:03:05.020

So I asked around and I found that Microsoft was super interested in hiring me

00:03:05.020 --> 00:03:13.260

and that was now, well tomorrow exactly a month, a year, tomorrow a year ago

00:03:14.380 --> 00:03:28.220

I started at Microsoft. In the beginning I just had to find my way around at Microsoft.

00:03:28.220 --> 00:03:35.180

Eventually I figured I should pick a project and after looking around and realizing I

00:03:35.180 --> 00:03:41.900

couldn't really turn the world of machine learning upside down. I figured I'd

00:03:42.940 --> 00:03:48.860

stay closer to home and see if Microsoft was interested in funding a team working on

00:03:48.860 --> 00:03:58.620

speeding up CPython. And I was actually inspired by Mark's proposals that were going around at the

00:03:58.620 --> 00:04:07.340

time. So I convinced Microsoft to sort of start a small team and get Mark on board.

00:04:08.460 --> 00:04:15.020

Yeah, that's fantastic. I also feel a little bit like machine learning is amazing, but I don't have

00:04:15.020 --> 00:04:20.460

a lot of experience with it. And whenever I work with it, I always kind of feel on the outside of

00:04:20.460 --> 00:04:27.100

it. But this core performance of Python, that helps everybody, right? Including even Microsoft,

00:04:27.100 --> 00:04:33.260

right? It maybe saves them energy on Azure when they're running Python workloads or whatever.

00:04:36.300 --> 00:04:38.300

So you're enjoying your time, you happy you're there?

00:04:38.300 --> 00:04:40.300

I'm very happy.

00:04:40.300 --> 00:04:45.300

Yeah, a lot of freedom to basically pursue what you are, right?

00:04:45.300 --> 00:04:50.300

And it's nice that the new Microsoft is very open source friendly.

00:04:50.300 --> 00:04:55.300

At least in many cases, obviously not everywhere.

00:04:55.300 --> 00:05:00.300

But sort of our department is very open source friendly.

00:05:00.300 --> 00:05:04.300

Things like Visual Studio Code are all open source.

00:05:04.300 --> 00:05:13.180

And so there was great support with management for sort of the way I said I wanted to do this project,

00:05:13.180 --> 00:05:15.860

which is completely out in the open.

00:05:15.860 --> 00:05:24.020

Everything we do is sort of just merged into main as soon as we can.

00:05:24.020 --> 00:05:27.980

We work with the core developers.

00:05:27.980 --> 00:05:31.900

we don't have like a private fork of Python

00:05:31.900 --> 00:05:34.100

where we do amazing stuff.

00:05:34.100 --> 00:05:37.620

And then we knock on the steering council's door

00:05:37.620 --> 00:05:41.820

and say, "Hey, we'd like to merge this."

00:05:41.820 --> 00:05:45.180

- Yeah, you're not gonna drop six months of work

00:05:45.180 --> 00:05:46.820

just in one block, right?

00:05:46.820 --> 00:05:49.240

It's there for everyone to see.

00:05:49.240 --> 00:05:50.080

- Exactly.

00:05:50.080 --> 00:05:53.320

- I think that's really, really positive.

00:05:53.320 --> 00:05:57.460

And wow, what a change, not just from Microsoft,

00:05:57.460 --> 00:06:00.140

but so many companies to work that way

00:06:00.140 --> 00:06:02.300

compared to 10, 15 years ago.

00:06:02.300 --> 00:06:05.820

Yeah, absolutely.

00:06:05.820 --> 00:06:08.460

Now, before I get to Mark, I just wanna,

00:06:08.460 --> 00:06:11.320

you know, a bunch of people are excited that you're here.

00:06:11.320 --> 00:06:13.820

And Luis out in the audience said,

00:06:13.820 --> 00:06:15.700

"Wow, it's Guido, I can't thank you enough

00:06:15.700 --> 00:06:18.020

"for your amazing Python and all the community."

00:06:18.020 --> 00:06:18.860

Awesome.

00:06:18.860 --> 00:06:21.060

- Great to hear.

00:06:21.060 --> 00:06:21.900

- Yeah.

00:06:21.900 --> 00:06:23.700

Mark, how about you?

00:06:23.700 --> 00:06:26.300

How'd you get into this Python performance thing?

00:06:26.300 --> 00:06:28.940

I know you did some stuff with HotPie back in the day.

00:06:28.940 --> 00:06:32.020

- Yeah, that was sort of my PhD work.

00:06:32.020 --> 00:06:36.460

So I guess I kind of got into the performance

00:06:36.460 --> 00:06:37.780

almost before the Python.

00:06:37.780 --> 00:06:42.020

So I was doing sort of compiler work, masters,

00:06:42.020 --> 00:06:44.700

and obviously just, you know,

00:06:44.700 --> 00:06:48.280

you need to write scripts and just get stuff done.

00:06:48.280 --> 00:06:51.300

And, you know, often just Python is just a language

00:06:51.300 --> 00:06:52.140

to get stuff done.

00:06:52.140 --> 00:06:55.540

And then it's that, I think I'm in Rigo,

00:06:55.540 --> 00:06:58.580

sort of, I think one of his sort of credits

00:06:58.580 --> 00:06:59.740

in one of his papers or something,

00:06:59.740 --> 00:07:01.060

he says, "Thank you for Python

00:07:01.060 --> 00:07:02.780

"for being such a great language to use

00:07:02.780 --> 00:07:05.220

"and such a challenge to optimize."

00:07:05.220 --> 00:07:08.540

So it's doubly good if you're coming at it from a sort of,

00:07:08.540 --> 00:07:10.720

so it provides this great intellectual challenge

00:07:10.720 --> 00:07:12.540

when you're actually trying to optimize it.

00:07:12.540 --> 00:07:14.380

And it's a really nice language to use as well.

00:07:14.380 --> 00:07:16.780

So it's doubly good.

00:07:16.780 --> 00:07:19.500

- It is doubly good, it's doubly good.

00:07:19.500 --> 00:07:21.220

Yeah, and before we move on really quick,

00:07:21.220 --> 00:07:22.860

Paul Everett says, "It's really impressive

00:07:22.860 --> 00:07:25.500

of how the in the open work has been done.

00:07:25.500 --> 00:07:27.020

Yeah, totally agree.

00:07:27.020 --> 00:07:28.060

- Hi, Paul.

00:07:28.060 --> 00:07:29.100

- Yeah, keep that going.

00:07:29.100 --> 00:07:30.800

Hey, Paul, happy to see you here.

00:07:30.800 --> 00:07:35.020

Now, we're gonna talk about making Python faster,

00:07:35.020 --> 00:07:38.380

but I wanna start this conversation

00:07:38.380 --> 00:07:41.180

with a bit of a hypothetical question,

00:07:41.180 --> 00:07:44.460

but sort of set the stage and ask,

00:07:44.460 --> 00:07:47.380

how much does Python really need to be faster?

00:07:47.380 --> 00:07:49.780

Because on one hand,

00:07:51.020 --> 00:07:53.260

Sure, there's a lot more performance we can do

00:07:53.260 --> 00:07:55.660

if you're going to say, well, we're going to solve the nbody

00:07:55.660 --> 00:08:00.940

problem using C++ or C# versus Python.

00:08:00.940 --> 00:08:04.540

It's going to be faster with the native value types and whatnot.

00:08:04.540 --> 00:08:06.780

On the other, people are building amazing software

00:08:06.780 --> 00:08:08.860

that runs really fast with Python already.

00:08:08.860 --> 00:08:12.900

We've got the C optimizations for things like NumPy

00:08:12.900 --> 00:08:16.780

and SQL alchemies, transformation layer,

00:08:16.780 --> 00:08:18.500

serialization layer, and so on.

00:08:18.500 --> 00:08:20.900

So a lot of times that kind of brings it back

00:08:20.900 --> 00:08:21.740

to C performance.

00:08:21.740 --> 00:08:25.380

So how much do you think Python really needs

00:08:25.380 --> 00:08:28.760

to be optimized already?

00:08:28.760 --> 00:08:32.220

Not that more is always better, faster is always better,

00:08:32.220 --> 00:08:34.400

but I just kind of want to set the stage

00:08:34.400 --> 00:08:36.100

and get your two thoughts on that.

00:08:36.100 --> 00:08:43.020

- Well, I always think back to my experience at Dropbox,

00:08:43.020 --> 00:08:48.080

where there was a large server called the meta server,

00:08:48.080 --> 00:09:00.480

which did sort of almost all the server side work, like anything that hits www.dropbox.com hits that server.

00:09:00.480 --> 00:09:10.080

And that server was initially a small prototype written in Python. The client was actually also a small prototype written in Python.

00:09:10.080 --> 00:09:16.480

And to this day, both the server and the client at Dropbox, as far as I know,

00:09:16.480 --> 00:09:20.720

unless in the last two years, they totally ripped it apart, but I don't think they did.

00:09:20.720 --> 00:09:27.120

They tweaked it, but it's still all now very large Python applications.

00:09:27.120 --> 00:09:36.960

And so Dropbox really sort of feels the speed of Python

00:09:36.960 --> 00:09:41.960

in its budget because they have thousands,

00:09:41.960 --> 00:09:44.440

I don't know how many thousands of machines

00:09:44.440 --> 00:09:48.760

that all run this enormous Python application.

00:09:48.760 --> 00:09:49.600

And yes-

00:09:49.600 --> 00:09:50.740

- If it was four times faster,

00:09:50.740 --> 00:09:54.180

that's not just for a quarter of the machines,

00:09:54.180 --> 00:09:58.120

that's less DevOps, less admin, all sorts of stuff, right?

00:09:58.120 --> 00:10:02.080

- Oh, even if it was 4% faster, they would notice.

00:10:02.080 --> 00:10:03.920

- Yeah.

00:10:04.920 --> 00:10:07.560

The other area where I think it's really relevant

00:10:07.560 --> 00:10:13.260

has to do with the multi-core side of things, right?

00:10:13.260 --> 00:10:17.220

I have a PC over there, 16 cores.

00:10:17.220 --> 00:10:19.020

My new laptop has 10 cores.

00:10:19.020 --> 00:10:23.700

Although with Python, it's hard to take true advantage

00:10:23.700 --> 00:10:27.160

of that side of modern CPU performance

00:10:27.160 --> 00:10:28.620

if it's not IO bound, right?

00:10:28.620 --> 00:10:33.800

- Yeah, so the, I don't know how deep you want me

00:10:33.800 --> 00:10:42.920

go into that and Mark can stop me if I'm going too deep too. But there are existing patterns

00:10:42.920 --> 00:10:53.080

that work reasonably well if you have a server application that sort of handles multiple

00:10:53.080 --> 00:11:00.120

fairly independent requests. Like if you're building a multi-core web application

00:11:03.080 --> 00:11:09.560

you can use multi-processing or pre-forking or a variety of ways of running a Python

00:11:09.560 --> 00:11:15.480

interpreter on each core that you have, each independently handling requests.

00:11:15.480 --> 00:11:21.160

And you can do that if you have 64 cores, you run 64 Python processes.

00:11:21.160 --> 00:11:24.920

Right, that's just a number in a micro-WSGI config file. It's nothing.

00:11:25.800 --> 00:11:35.240

Yeah, but it works for applications that are designed to handle

00:11:35.240 --> 00:11:43.880

multiple independent requests in a scalable fashion. There are other algorithms that

00:11:43.880 --> 00:11:52.120

you would want to execute where it's much more complicated to employ all your cores efficiently.

00:11:54.040 --> 00:11:56.040

Yeah, absolutely.

00:11:56.040 --> 00:12:00.040

That's still a nut that Python hasn't cracked.

00:12:00.040 --> 00:12:08.040

And I'm assuming you're asking this question because Sam Gross, a very smart developer at Facebook, claims that he has cracked it.

00:12:08.040 --> 00:12:16.040

Yeah, and perhaps he has. It's an interesting idea. We'll dive into that a little bit later.

00:12:16.040 --> 00:12:21.880

I'm more asking it just because I see a lot of people say that Python is too slow.

00:12:21.880 --> 00:12:25.640

And then I also see a lot of people being very successful with it.

00:12:25.640 --> 00:12:31.640

And it not being slow in practice, or not being much slower than other things.

00:12:31.640 --> 00:12:36.600

And so I more want to set the stage of like the context matters, right?

00:12:36.600 --> 00:12:39.880

This Dropbox example you have, it really matters to them.

00:12:39.880 --> 00:12:43.560

You know, my course website where people take courses,

00:12:43.560 --> 00:12:45.960

the response time of the pages is 40 milliseconds.

00:12:45.960 --> 00:12:48.560

If it was 38, it doesn't matter.

00:12:48.560 --> 00:12:50.640

It's really fast, it's fine.

00:12:50.640 --> 00:12:53.660

So I think, but if I was trying to do computational biology

00:12:53.660 --> 00:12:56.480

in Python, I would really want to be able to take advantage

00:12:56.480 --> 00:12:58.200

of those 16 cores, right?

00:12:58.200 --> 00:13:02.320

So there's just such a variety of perspectives

00:13:02.320 --> 00:13:04.320

where it matters.

00:13:04.320 --> 00:13:06.600

Mark, what are your thoughts on all this?

00:13:06.600 --> 00:13:10.120

- Well, it's just a case of saving energy, saving time.

00:13:10.120 --> 00:13:13.240

Just, it just makes the whole thing nicer to use.

00:13:13.240 --> 00:13:20.920

So I mean, there's a lot of just if development in data science and it's that responsiveness,

00:13:20.920 --> 00:13:26.000

the whole, you know, just breaking your train of thought because things take too long versus

00:13:26.000 --> 00:13:28.240

just keeping in the flow and all this sort of stuff.

00:13:28.240 --> 00:13:30.480

It's just nice to have something that's faster.

00:13:30.480 --> 00:13:33.480

I mean, it's not just the big companies saving money as well.

00:13:33.480 --> 00:13:36.280

I mean, it's just keeps everyone's server budgets down.

00:13:36.280 --> 00:13:41.320

I mean, if you just need a smaller virtual instance because you can serve the requests

00:13:41.320 --> 00:13:49.560

up fast enough because Python's faster. So I think it's just generally a sort of responsible

00:13:49.560 --> 00:13:53.460

thing to do. I mean, it's also just, you know, people expect technology to move forwards

00:13:53.460 --> 00:13:58.560

and there's this feeling of, you know, falling behind or not, you know, people wanting to

00:13:58.560 --> 00:14:01.320

move other languages because of the perceived performance.

00:14:01.320 --> 00:14:05.400

Right. I do think that that's an issue, you know, I'm moving to Go because it has better

00:14:05.400 --> 00:14:08.360

or async support, or we're rewriting this in Rust

00:14:08.360 --> 00:14:11.240

for whatever reason.

00:14:11.240 --> 00:14:13.200

Sometimes that might make sense, but other times

00:14:13.200 --> 00:14:16.080

I feel like that's just a shame, and it could be used better.

00:14:16.080 --> 00:14:17.800

A couple of questions from the audience

00:14:17.800 --> 00:14:21.120

I just want to throw out there.

00:14:21.120 --> 00:14:24.080

Let's see.

00:14:24.080 --> 00:14:29.720

One was, Guido, especially, you must

00:14:29.720 --> 00:14:31.640

be really proud to hear about the Mars

00:14:31.640 --> 00:14:35.600

helicopter and the lander and Python in space,

00:14:35.600 --> 00:14:39.160

how did you feel when you heard about the helicopter

00:14:39.160 --> 00:14:42.520

using Python and the lander using Python and Flask

00:14:42.520 --> 00:14:43.560

and things like that?

00:14:43.560 --> 00:14:53.880

- It wasn't really a surprise given how popular Python is

00:14:53.880 --> 00:14:55.400

amongst scientists.

00:14:57.240 --> 00:15:04.280

So I didn't throw a party but it made me feel good. I mean it's definitely

00:15:04.280 --> 00:15:12.040

sort of one of those accomplishments for a piece of technology if it's actually shot into space.

00:15:12.040 --> 00:15:22.040

You know you've made a difference. I remember like 30 years ago or more when I

00:15:23.480 --> 00:15:31.080

helped some coding on a European project called Amoeba, which was like a little distributed

00:15:31.080 --> 00:15:37.800

operating system. And one of the things that they always boasted was that

00:15:37.800 --> 00:15:44.040

our software runs on the European Space Station. And that's very important.

00:15:44.040 --> 00:15:50.920

Yeah, so yeah, I totally get the feeling and I hope that everyone who contributed to Python

00:15:51.800 --> 00:15:56.800

also sort of feels that their contribution has.

00:15:56.800 --> 00:16:00.200

- That sense of awe, if you look up in the night sky

00:16:00.200 --> 00:16:02.080

and you see that little, that bright star

00:16:02.080 --> 00:16:04.880

that's actually Mars and you think, yeah, it's up there.

00:16:04.880 --> 00:16:07.760

Yeah, fantastic.

00:16:07.760 --> 00:16:10.880

All right, let's dive into some of the performance stuff

00:16:10.880 --> 00:16:12.140

that you all have been doing.

00:16:12.140 --> 00:16:15.920

So maybe Guido, start us out with the team.

00:16:15.920 --> 00:16:18.760

So you've got a group of folks working together.

00:16:18.760 --> 00:16:21.400

It's not just you and also now Mark Shannon

00:16:21.400 --> 00:16:23.200

is working with you as well, right?