Simplifying Python's Async with Trio

00:00 Ever since Python 3.5 was released, we've had a really powerful way to write I/O bound async code

00:06 using the async and await keywords. On this episode, you'll meet Nathaniel Smith, who wrote

00:11 the Trio async framework that significantly simplifies complex coordinating operations

00:16 using async and await. This is Talk Python to Me, episode 167, recorded June 21st, 2018.

00:23 Welcome to Talk Python to Me, a weekly podcast on Python, the language, the libraries, the

00:42 ecosystem and the personalities. This is your host, Michael Kennedy. Follow me on Twitter

00:47 where I'm @mkennedy. Keep up with the show and listen to past episodes at talkpython.fm

00:52 and follow the show on Twitter via @talkpython. This episode is sponsored by Linode and Rollbar.

00:58 Please check out what they're offering during their segments. It really helps support the

01:01 show. Nathaniel, welcome to Talk Python.

01:05 Hey, thanks for having me.

01:07 It's great to have you here. I've wanted to have you on for a long time. I, you know,

01:10 listeners of the show probably know that I'm a fan of async programming and parallel programming.

01:18 It's definitely gotten some feedback and some attention, but I think it's super important

01:23 for Python. And I think what you're doing is incredible. Some of the stuff you showed at

01:27 PyCon this year was incredible. So I'm just, I'm really excited to talk about it. Yeah, it's

01:33 going to be fun. Before we get to that though, let's get your story into programming and Python.

01:36 So let's see. So for getting into programming, I mean, I guess I was fortunate in my choice of

01:41 parents. I was sort of, you know, born into sort of a relatively affluent family. And my mother was

01:47 also a professional programmer. So she mostly stopped, you know, switched to sort of part-time consulting

01:52 consulting when the kids came along. So she could be a stay at home mom. So actually I first learned a

01:58 program through in my elementary school, they did have this program, this teaching logo to us that my

02:05 mom created this project at my elementary school. Wow. That's pretty awesome. Many moms volunteer at

02:13 schools, but not any of them. Right. Learning programs. Yeah, no, you know, my mom is pretty

02:20 incredible. She also loves podcasts. So I mean, I guess, hey, hi, mom. Maybe she'll listen to this.

02:26 Yeah. I mean, she later went on to get a master's in education, and now teaches art and science and

02:34 programming. Yeah. But so yeah, so I got started early and was fortunate enough to have my own computer

02:39 starting around like age 10 and so on. So this was like in the days when I was like,

02:43 it was DOS. Windows 3.1 was pretty exciting. Like 386, 46 type thing. Oh yeah. Yeah. Turbo button to

02:50 like make it go faster. Yeah. Yeah. We don't have turbo buttons anymore. No, it just always goes fast.

02:55 Isn't that weird, right? Yeah. Yeah. Isn't there a little button to make your laptop half the speed?

02:59 Why would everyone want that? So, and then I, from there I installed Linux when I was like 13 on my

03:07 computer messing around. I had no idea what I was doing. It was Slackware 96 because like, was the name of it?

03:12 It was, it's like, you know, Windows 95 was the big thing then. So like, this was one better, right?

03:16 96.

03:17 Exactly. Yeah. Right. So yeah. So I've kind of been, you know, bumming around that kind of,

03:23 you know, programming open source world for something years now. It took me a little longer to get into

03:28 Python. I don't remember exactly when it probably, you know, five years later. So I know the first time

03:34 people were talking about, you know, Python was this cool new thing. It's, you know, kind of like

03:37 Perl, but so forth. first time I read the tutorial, I got to the point where it said that

03:43 Lambda was syntactically limited to a single expression and just like rage quit. I was like,

03:47 whatever, stupid language.

03:49 Where are the curly braces? Come on.

03:52 Yeah. Well, it wasn't, I mean, I guess also like, so like, but I had like sort of like logo,

03:55 I learned scheme and things like that. So like, this was like obviously terrible. I'd later came

04:01 back to it, got over that. And I, and under, you know, understand now why that makes sense in terms

04:05 of the block structure and everything, but the statement expression distinction that Python has,

04:10 but, it took me a little while, but yeah. So I, yeah, I guess I got into Python around

04:16 like 2.2 or so. And it sort of gradually become, you know, just sort of my main language.

04:22 In the meantime, I was also like, you know, I was getting a PhD, stuff like that. So using Python

04:27 for both like hobby hacking stuff. And also like for my, in my work, you know, for data analysis,

04:34 I wrote some various libraries there. What's your PhD in?

04:37 Yeah. Cognitive science.

04:38 Oh, right. Okay. Very awesome.

04:39 Yeah. Yeah. So it was like studying how people understand language and brains understand language.

04:44 Yeah. So yeah. So along the way, and then, yeah, getting embedded and all these different sort of,

04:49 you know, entangled in all these different sort of open source projects, like NumPy and so on.

04:53 Yeah. Yeah. That's really great. So what do you do these days?

04:57 So I'm actually kind of in a bit of transition right now. So the last few years, I've been

05:01 working at UC Berkeley at the Berkeley Institute for Data Science, which is a sort of new

05:07 data science center. That's what started up on campus. And so in my position has been sort of

05:12 this unique, one really where it's, been, I've been grant funded to sort of figure out how to,

05:19 you know, make sure the Python, you know, continues to work well, works better for scientists. There's

05:25 a, actually there's a blog post we can like put a link to there. I wrote just sort of talking about

05:30 all this sort of stuff I've done there, but some highlights being like, I got, some grants for

05:35 a NumPy development, $1.3 million over a few years, which is basically the first time NumPy

05:41 has been funded. We'll have full-time people working on it. Did I made a color map people like,

05:48 did some work on packaging? So like, many Linux sort of led that effort so we can have wheels on

05:54 Linux now. That sounds like a really fun job. Like you get to actually make a difference in

06:00 the open source space. Exactly. Yeah. I mean, it partly it's, it's sort of funny. Like this is what

06:04 I talk about in the blog post, but like, you know, open source is just sort of like desperately underfunded.

06:08 And like, it's amazing how well and effectively we can use the little bits of volunteer time and

06:14 things that people manage or, you know, people have spending, you know, a few hours a week.

06:18 But there's a lot of things that like larger projects that have huge impact that can't be

06:24 done in that mode where you actually need to like sit down for a little while and think about it and,

06:28 you know, understand the landscape, put some focused effort into something.

06:31 Yeah. It's a difference on like what you take as your goal. If you have a month of uninterrupted full

06:38 time, some project or a goal part of project versus I'm going to squeeze this in

06:44 Saturday morning before the kids get up, right? These are the same types of things you attack.

06:49 Exactly. Yeah. I mean, it's like a month of time is not that much in the grand scheme of things. One

06:53 person, one month is like compared to, you know, the amount of labor that like, you know, Google has

06:58 available. Yeah. But it's still, you know, enables all kinds of things that right now just

07:03 aren't happening. And so there's actually lots of low hanging fruit because no, there is nobody

07:07 almost who has that kind of time to worry about these sort of broad open source ecosystem kind of projects.

07:14 There's just, there's all kinds of... We saw how much of a difference was made with the Mozilla

07:19 grant, the Python package.

07:21 Or the PyPI. Yes.

07:22 Yeah. The PyPI, like that was like, hey, look, $170,000. Oh, actually this can happen. It's

07:27 been dragging on for years. Now it's, you know, it's real.

07:29 Yeah. Yeah. I mean, getting the team together to make that happen. Yeah. I had a little bit of

07:34 involvement on that. I'm on the PSF's packaging working group. I didn't, I wasn't heavily involved

07:39 in that. You know, I like made some introductions, gave some feedback on the grant and things.

07:43 So it was, yeah, it's just super exciting to see that up close. Because yeah, it was very successful.

07:48 And it wasn't...

07:49 Like it really got a lot out of how much investment was put into it.

07:53 Yeah. I mean, it was like, that had been dragging on for literally six years. The old PyPI was just

07:58 totally unworkable. But, you know.

08:03 It was like Kryptonon for people who wanted to contribute to open source. They looked at it and

08:08 like, oh, no, no, I'm not touching that. That's...

08:10 Yeah. I made one patch to the old PyPI. And it was just a trivial thing. It was just that we wanted it

08:16 to start to say that, oh yeah, many Linux is now a legal thing for a wheel to have in its name.

08:21 So it was just like, there's like a list of strings that allowed, I'm just adding one new string,

08:25 right? That was the most terrifying patch I've ever made. PR I've ever made. Because it's like,

08:30 if I have a typo, then PyPI dies, right? And there are no tests. There's nothing, right?

08:37 The consequences of failure is so high.

08:40 Yeah. Exactly.

08:41 Yeah. The new PyPI is way, way better. You should all go and help make it contribute. Because now

08:47 it's all set up. You can develop locally and there's a real test suite.

08:52 Yeah. It's really nice. I had the people involved on the show a couple episodes back too. So it

08:58 definitely got a chance to dig into that. Oh yeah. Right. That's right. Yeah. Yeah.

09:01 So let's talk a little bit about async. Yeah.

09:04 Yeah.

09:04 Because I think that that's one of the sort of revitalized areas of Python last,

09:09 since 3.4, 3.5, it really started coming along, right? So what, in 3.4, we got asyncio as a thing.

09:16 And in 3.5, it really, I feel like it really got more accessible with async and await.

09:22 Right. With the new syntax features to make it easier to use. Exactly. Yeah.

09:26 Yeah. Yeah, exactly. Like the keep it, the foundation was built previously, but it was,

09:31 it was still this sort of callback hell type of programming, right? Right.

09:37 Right. We should maybe take a little bit of a step back and just, you know, let people know about

09:42 the presentation you gave at PyCon, which introduces your project TRIA, which is what we're going to

09:46 focus on. But in your talk, you sort of set the stage and said, look, there's these two

09:51 philosophies or types of async programming that you might consider. So maybe, maybe you could touch on

09:58 that a little. Well, so, I mean, I think the first thing to say is for those who aren't kind of already,

10:03 you know, up on the jargon, async is sort of a subset of concurrent programming. And so concurrency,

10:08 meaning writing programs that do two things at the same time, which is very handy. You know,

10:14 it's something in our real life we do all the time, you know, you know, I'm working on one thing and my

10:18 friends working on another thing at the same time. It's very natural. But writing a program that does

10:22 that is a little bit trickier, especially, you know, in Python is kind of generally a sequential

10:27 language, right? It makes it easy to do, I want to do this and then that and then the other thing.

10:31 It doesn't directly have built into like the syntax, whatever ways to say, I want to do these

10:36 two things at the same time. And then when they're both done, do these other things.

10:39 So, so generally, there's this general question of like, how do you write concurrent programs in Python?

10:43 And then I think what you were thinking of, there's kind of two philosophies of concurrency,

10:48 which is one is the kind of preemptive concurrency that threads give you, where just everything

10:56 just kind of runs all the time interleaved kind of in arbitrary ways.

11:00 To be clear, this is in general threads, not Python threads, because we have this whole

11:04 thing called the gil.

11:05 From the programmer, from the user's point of view, the gil doesn't make, gil makes things

11:09 slower, but it doesn't make it really change what it feels like to use it compared to threads in

11:14 general. True.

11:15 So in general in threads, you might have like two threads actually running at the same time,

11:20 like on two different views. Because of Python, we have the global interpreter off the gil,

11:24 multiple, then it means that mostly only one thread can actually run at a time.

11:28 But because in Python interpreter controls that, it can decide at any moment to switch which thread

11:33 is running. From your point of view, as writing code in Python, it might as well be running multiple

11:38 things at the same time, basically, right? Because since it could switch at any moment,

11:42 you kind of have to act like it's just constantly switching, right?

11:45 And the reason this is kind of a challenge is because if you have a data structure, and you have two

11:50 different threads or whatever, you know, you have two different concurrent pieces of code acting on

11:55 the data structure at the same time, if they're not like, if you're not careful, you can make a big

11:59 mess. So like, you know, one, you have the classic examples are things like, you know, your program is

12:05 like managing like a bank. And so I'm going to withdraw money from this account, and then add it,

12:09 put it in this account. But if you don't do that, like atomically, so like ice, you know, first one,

12:14 one thread says, okay, does Alice have enough money? Yeah, great. Okay, I'll take that money out and put it

12:19 into Bob's account. Another thread says, oh, does Alice have enough money for this transfer to Carl?

12:24 And then if they're in a leave right, first both threads check and say, oh, yeah, Alice has plenty of money.

12:28 And then both threads take money out of the account. And now Alice has spent the same money twice, actually.

12:33 Yeah, great for Alice, but not so nice for the bank.

12:35 So yeah, someone's gonna be trouble running writing that software.

12:40 Yeah, right. So yeah, if you're writing that software, you need to sort of manage this.

12:44 Yeah. So one of the things that kind of helped me this like click in my mind was

12:48 thinking about how programs temporarily enter these invalid states.

12:54 Right.

12:55 What you're describing is basically that like at the beginning of that function,

12:59 the bank is in a valid state. At the end of the function, everything is valid. But somewhere in

13:04 between in order to accomplish that operations as a series of steps, it has to become invalid. And

13:09 long as nobody observes it in this invalid state, it's all good. But when they do like, like you're

13:16 describing halfway through, it's not okay anymore. It's kind of like transactions and databases,

13:21 but in time. And yeah, they're pretty similar in some ways, I guess.

13:25 Yeah. So yeah, exactly. So with threads, which the solution to this is you have to sort of explicitly,

13:31 like, make sure you notice all those places where you're passing through that invalid state and do

13:36 some kind of in like market somehow in your source code, like say, okay, I'm going to take a lock here.

13:40 That's going to make sure that anyone else who tries to use this has to wait for me to be done

13:44 and things get back to the valid state before they can look at it. Right?

13:48 Yeah, exactly. Okay. But that's really error pro.

13:51 And this is preemptive concurrency, right?

13:53 Yeah, exactly. It's still talking about kind of how threads work, right?

13:55 Yep. So you would have to like find all these places where you go through a temporarily

13:59 invalid state and mark them in your source code. And if you forget one, then you have this nasty bug

14:05 where Alice gets to suspend the money twice or all kinds of weird things can happen.

14:10 And it usually has to do with timing. And so it's very, super hard to debug.

14:13 Yeah. And it's like, yeah, it's like super subtle. Like, yeah, like it only happens one in a thousand times randomly. And it happens,

14:20 it depends how much memory you have. And it's only happens in production and not on your tests.

14:23 And just all kinds of, it's really, really, yeah, it's bad. Yeah.

14:27 Yeah. I've heard them described as Heisen bugs. And I just love that term.

14:30 Yes. Yes. Heisen bugs. Right. And it's just like, and so it means that when you're working with threads,

14:34 you just have to be like constantly vigilant, right? Like every line of code,

14:37 which we think, okay, is this the one that's going to introduce this terrible bug?

14:40 So that's, that sucks, right? You don't want to live like that.

14:44 So, yeah, right. You just have to be like constantly paranoid. So yeah. So the idea for

14:51 async concurrency, because we kind of flip it around instead of saying that like, okay,

14:56 we have to go through and find all the places in the source code where something dangerous is

15:01 happening and mark those. You say, you know what, let's be conservative. Let's assume something

15:05 dangerous could be happening anywhere. That's the default. So by default, only one thing is allowed to

15:09 happen at a time. And then we'll mark in the source code, you know, okay, here's a place where

15:14 I want to let other things happen. It's okay for, I'm not in the middle of, you know, just doing some

15:19 delicate operation, you know, adjusting someone's bank account. This is a fine place for that to happen.

15:27 This portion of talk Python to me is brought to you by Linode. Are you looking for bulletproof

15:31 hosting that's fast, simple, and incredibly affordable? Look past that bookstore and check

15:35 out Linode at talkpython.fm/linode. That's L I N O D E. Plans start at just $5 a month for a

15:43 dedicated server with a gig of RAM. They have 10 data centers across the globe. So no matter where you are,

15:48 there's a data center near you. Whether you want to run your Python web app, host a private get server

15:54 or file server, you'll get native SSDs on all the machines, a newly upgraded 200 gigabit network,

16:01 24 seven friendly support, even on holidays and a seven day money back guarantee. Do you need a little

16:06 help with your infrastructure? They even offer professional services to help you get started

16:11 with architecture, migrations, and more. Get a dedicated server for free for the next four months. Just visit

16:17 talkpython.fm/linode.

16:20 And you can still have bugs. You can still make mistakes and put that in the wrong place.

16:26 Because now there's only like a finite number of places to check when something goes wrong, right?

16:31 It's just much easier to reason about that kind of program.

16:34 Right. And that could be done with just straight up asyncio. But with three, five and above,

16:39 like basically the await keyword is the marker in your code for that.

16:44 Exactly. Yeah.

16:45 Yeah. Yeah. That's pretty beautiful. That's pretty beautiful. So I'm guessing,

16:50 I'm guessing most people probably under understand like how to use async and await the keywords,

16:54 but maybe just describe like what's, how to define like a super, super simple async method with the

17:01 keyword, just to the people understand this.

17:02 Yeah. Like you said, I'm going to be focusing on trio here and trio does kind of use only kind of a

17:07 simple subset. There's some extra complexity needed for like backwards compatibility with async

17:12 IOS sort of called that place thing. We can get, talk about that more later. But sort of for if,

17:17 especially if you're just, you know, if you're using trio, async await is very simple actually.

17:21 And so here's what you need to know is that there are two kinds of functions. There's the async ones,

17:26 which are these, the special functions that might let other code run during them. And so you kind of

17:31 need to be aware of what you call one of them that like, you know, the ground might shift under

17:35 feet, data structures might change while it's happening. And there's the regular functions,

17:40 like the synchronous functions where those always happen atomically. And so, and we're going to do is

17:45 we're going to make all the regular Python functions you already know, those are going to be the synchronous

17:49 ones. Because, you know, no one, you know, they're all written on the assumption that like, you know,

17:53 Python's a sequential language, it does things sequentially. And so they, no one like thought about

17:58 how to handle this async stuff when they were writing all these libraries that already exist.

18:03 So those are going to be all atomic. And then we need a way to mark these special async functions.

18:09 And we want to mark it at two places. So what we want to mark it when we define the function

18:12 so that we, you know, as part of the API, is this an async function or not? And we want to mark it at

18:18 the call sites. When you call one of these, when you're like reading through your code, you see, ah,

18:22 this is a point where I'm calling an async function. This is a point where other code might run in the

18:26 background, I need to make sure I'm not in the middle of some, you know, dangerous operation.

18:29 Right?

18:30 Right. Okay. So you've got like, maybe a call to a web service or something.

18:34 Yeah, that might return like a future, which then you hook in a call.

18:39 So there are no futures in Trio. That was one of the things I got rid of.

18:43 Yes. Yeah, that's beautiful. Yeah. But you got to, you have to wait on the result, basically.

18:48 Right?

18:49 So yeah. I mean, so yeah, so the way, yeah, the way you think about it for Trio is just that,

18:54 it's just like, there are these special functions and the way you call it is you type

18:57 a weight and then the function call. It's like a new kind of syntax for calling a function.

19:01 Right.

19:02 And that's sort of all you need to know. There is, yeah, like you said, like there's the

19:05 complexities around Python also supports a weighting on objects and having future objects,

19:09 like the function to return and so on and so forth. But it's kind of unnecessary clutter is

19:14 kind of how Trio's perspective on it. So we just don't do any of that.

19:17 Yeah. I feel like the AsyncIO layer is like, let's get something kind of working,

19:22 but it's not that delightful. And I feel like Trio really cleans it up. So there's a really great

19:28 example that you gave in your talk about happy eyeballs, which is like a way to try to use

19:33 basically DNS resolution to connect a socket to a server and some concurrency stuff around that,

19:39 which is, I don't want to talk about that yet. Maybe we'll have some time to talk about it later. But

19:43 basically there's a version in Twisted, which is how long was the one in Twisted? Hundreds?

19:49 And so, yeah, well, yeah, there's two different versions of Twisted I talked about in the talk.

19:54 One is the sort of the classic one that's in in master, which is like 120 lines long, I think,

20:00 roughly. I mean, not that it's not super meaningful to talk about lines of code like this, but like

20:04 just kind of give you a sense. Yeah. Yeah, exactly.

20:07 A better sense of complicated is to reason about that is it has inside this method. It has another

20:12 internal function. Inside that function, there's another function defined. And then inside that,

20:16 there's a fourth level of internal functions. So like, yeah, that's bad.

20:20 Because it's all these. Yeah. Yeah.

20:24 Almost like it's full of all these go to's and in some weird sense. And then, and you basically say,

20:30 now, if we apply the building blocks or the primitives of trio to it, oh, look, it's like 20 lines,

20:34 and it's totally straightforward, which is really great. So I think, you know, let's talk about trio.

20:40 Why don't you start by telling people what it is. Trio is a library for a sink concurrency in Python.

20:45 It's like an alternative to libraries like a sink or twisted or so on, which NATO or threads,

20:52 sort of all of them. And I think there's kind of, I think it was, there's kind of two pieces. So one

20:57 is there's sort of the idea trio, the ideas are like a research project or something where I have,

21:03 I sort of did some analysis of like, what is it about twisted and I think I own someone that makes

21:08 it so difficult to use sometimes, or why are things like happy eyeballs so complicated?

21:13 Where are some common errors sort of come from? It came up like, oh, that's actually I sort of,

21:18 as I dug into it, I realized actually, there's, you know, a small number of things that kind of

21:21 seem to cause a lot of the problems, and even ended up digging into some old literature from like the

21:26 60s, about early programming language design. On Univac or what was that one on? That was like

21:32 a really old computer. Oh, oh, Flowmatic. The, that language, which is, yeah, the Grace Hopper's

21:38 language, the precursor to COBOL, which is a really interesting language. That's going way back. Yeah.

21:42 Yeah. Yeah. And sort of talking about this transition, there was a lot of debates then about like,

21:47 how do you structure your language, not even getting into the concurrency part, just like,

21:51 how do you do a language at all that's like usable? And one of the big things that happened was

21:56 this switch from using go to as the main control structure to having things like if statements and

22:02 functions and for loops and so on. And I realized there's actually sort of an analogy there that's

22:08 surprisingly like precise between a lot of that, a lot of these async libraries are actually still

22:16 kind of in the go to stage, where sort of the basic primitives are in sort of a technical way,

22:21 kind of analogous to a go to, and they cause similar kinds of problems. And then if you look at,

22:25 okay, how did, you know, Dijkstra solve these problems back in the late sixties,

22:29 we can take those and apply them to concurrency. And that leads to something called,

22:34 I call it a nursery concept. Yeah. It's so interesting. And I had, yeah. And I had never

22:38 thought about the relationship between go to and many of these programming, these threaded programming

22:44 models, but you really hit it. Well, I think it's a super good analogy and it's really relevant.

22:50 Because so what Dijkstra said was, look, you should be able to treat these building blocks as black boxes.

22:56 Stuff goes in, stuff goes out. That's all. If you know kind of what it does, you don't need the details,

23:01 right? I mean, this is like a huge part of abstraction and programming, like functions,

23:05 classes, it's modules, et cetera. Right. Like, like even just think of like,

23:09 like in Python, think of the print function, right? Like in actually it does all kinds of complicated

23:17 stuff, right? It's got talking to the operating system interfaces to do IO and it's like buffer

23:22 a thing and character second version and blah, blah, blah. But like, you don't have to care about

23:26 that. You just type print hello world. And it, you know, it prints all the world.

23:30 You can just treat that as like this little atomic unit. And that's great.

23:33 That kind of. Yeah. It's such an important part of a building block of programming,

23:36 but these threads mean that stuff can still be going. You're just like all over the place. Right.

23:41 And it's very similar to the go-to's, which I thought that was a great analogy.

23:44 Yeah. I mean, specifically the issue is, so like, let's say the, the analog of print,

23:48 like in say twisted, you have this transport object. That's like your connection,

23:52 like a network connection and you call it's right method. And this is like,

23:55 I want to send this data to the remote site now. And that's like a function. You call it

23:59 it returns. That's all fine. But what's sort of confusing is that when you call it,

24:04 it returns, it hasn't actually finished yet. What it's actually done is sort of scheduled

24:08 that right to happen in the background. And that makes it hard to reason about the flow.

24:12 Cause it's like, Oh, the functions returned, but actually in a sense, it's like kind of still running.

24:17 Right. And now if I want to write and then do something else that happens after that,

24:21 that's hard to manage because I don't necessarily know when that's actually finished. I have to use some

24:26 other API to like check and ask, okay, has it actually written yet?

24:29 Yeah. And so yeah.

24:30 Did it succeed? Did it not succeed? Then how do you deal with that? And then how do you get that

24:34 back into your, yeah, your whole, it's yeah. It feels like almost like JavaScript.

24:37 Well, I mean, JavaScript, you know, is also in this general family of like, yeah,

24:42 it is as this async currency model that's sort of endemic to it. We use all over the place and it's

24:48 all callback based and it's all, yeah, it has that same kind of go-to problem.

24:52 Yeah, exactly. So let's talk about the architecture and then maybe like we can see about how that kind

24:58 of brings us back into something that Dykstra might be proud of.

25:01 Okay. Sure. Yeah. I guess I should also mention, so I don't know how, this is a discussion that really benefits from like diagrams, which is not something podcasts or

25:11 there is also, I have a blog post called notes on structured concurrency or go statement considered

25:18 harmful, which goes into this and sort of this set of ideas and much more detailed pictures and all

25:23 that. But so yeah, so the basically the idea is that, I mean, so it's what I said, right? That like,

25:27 when you call a function, it should be something you can treat as a black box. It does some stuff,

25:32 it returns. That's kind of what Dykstra's point. I mean, that's the problem with go to is also that

25:36 like, you know, the old go to world, you could call a function and maybe it returns, maybe it jumps

25:43 somewhere else, maybe some other function and some other point of your code suddenly returns instead

25:47 because the control go to just lets, you know, functions have a certain control flow they're supposed

25:52 to have, right? Where you have your function and then you call something. So control jumps into that

25:56 function. It executes that code and then it comes back again to your function. That's kind of this

26:00 nice little structure. That's why it's, you could treat it as a black box because you know, oh, it'll

26:03 do some stuff, but then it'll come back. Right. Go to isn't like that. It's just, it's just a one way

26:08 you leap off somewhere else and then maybe you come back, you could jump back, it's something you do manually.

26:13 Choose your own adventure book. You don't know when you're done, you don't know where to begin or

26:19 you're just like, it's always a mess. It's always a surprise. Yeah. Yeah. I mean, it's choose your own

26:23 adventure accepted. Then you're like, you know, actually I want to switch. I'm going to, it says

26:26 I can go to page five or 10, but I'm going to 17. Let's see what happens. So yeah. So in that,

26:32 and that breaks this function abstraction, right? So it means like you call a function and you know,

26:36 hopefully like if it's like a nice, well-written library and people aren't like messing with you,

26:41 it will come back normally. But there's no kind of guarantee of that in the language. Like if

26:46 someone decided to be clever, they could, it could do something else. It could jump somewhere else.

26:50 And then it becomes very hard to sort of reason about the overall structure of your program,

26:54 if that's true. And it also breaks sort of higher level structures that people are using the program.

26:59 But like, so, so think about exceptions, right? So it's very, a handy thing about exceptions is like,

27:04 oh, something goes wrong. Then, you know, and I didn't think about how to deal with it locally,

27:09 then the exception will propagate out in Python. And so until either someone catches it and knows how to

27:14 deal with it, or the program crashes, you get a trace back and like, that's not great. But at least,

27:18 you know, you didn't just like go blindly on doing the wrong thing. At least, you know,

27:22 you get a trace back, you can try and figure out what's going, what happened, right?

27:25 But not with, not with threading. Not if you care.

27:28 Well, okay, well, we'll get there in a moment, right? But if you have go to,

27:32 then it's where do you, when you raise an exception, it goes up the stack, right? It's like,

27:37 okay, does this function want to handle it? Okay. How about it's caller? How about it's caller's

27:40 caller? Right? Because you have this nice stack kind of, you know, you know who the caller is,

27:45 you know who the caller's caller is. If you have go to, then control just bounces around sort of at

27:50 random. So who wants to know about this exception? Like I don't have a well-defined sense of a caller

27:54 even, right? It's just like these basic things we take for granted growing up and you know, this,

27:59 you know, our ancestors struggled with and sort of fixed for us. And now we take for granted,

28:03 right? But like these basic assumptions just aren't there in the language of go to.

28:06 Yeah. With blocks have similar issues, right? So like a with block is supposed to like,

28:10 you say with open file and then you know, okay, inside the block, the file's open,

28:14 outside the block, the file's closed. That's nice. It's great. It makes it easy to manage resources.

28:20 But if you have go to, you like, you jump into the middle of a block, you jump out of the middle of

28:23 the block. Like what happens is the, where does the file open and close? Like how does that even,

28:27 I don't even, it doesn't work, right? Exactly. Yeah. So yeah. And then, then if you look at like problems we're dealing with threads or things that are a struggle and like

28:38 asyncio or twisted, but these are just kind of the things that are problematic, right? So you call

28:43 a function, is it, you know, it's still running maybe when it returns, it makes it hard to like

28:46 control sequencing, hard to treat things to black boxes. Like you don't know, like you need to like,

28:51 go read the source code when you use some random, asyncio like based library to find out like,

28:57 okay, I called this function, but did it actually do the thing then? Or is it scheduled something in

29:01 the background? Like you have to go, does the error go to the caller or does it go to the callback?

29:05 Right. Yeah. So if you spawn off this background thread or background task, as I think it calls it,

29:10 and it crashes, there's an error. There's an unhandled exception with that. Well, you've lost,

29:17 you don't have a nice sort of call stack that you've, you've spun up this new independent entity that's

29:21 executing and it's now split apart from your original code. And so there's sort of like, where does that

29:27 if there's an unhandled error, an exception, then what actually happens is that in threads and

29:32 they sync IO and twist or whatever, is it, you know, maybe print something on standard, like,

29:36 hey, I hope someone's looking, something went wrong. Then it throws away the exception and it carries on.

29:41 And you know, hopefully... It does make your software more reliable, because there's way fewer crashes when you don't actually... Well, for some value of reliable,

29:48 right? Yeah. Exactly. Nice. Okay. So what are the building blocks of Trio that like...

29:53 Yeah, exactly. Yeah. Yeah. So the main thing that is, in Trio, we have this thing we call a nursery.

29:59 And the idea, it's just a sort of a silly joke, right? About if you want to spawn a child task,

30:04 then it's not okay for it just to be go off independently. Like it has, you have to like,

30:08 put it somewhere and it'll be like taken care of. It's like a nursery is where children live.

30:13 Okay. So concretely, what it is, is if you want to spawn a child task in Trio, you have to first

30:19 have a nursery object. The way you get a nursery object is you write async with Trio.openNursery as

30:24 nursery. So it's like this open, you have like a with block that opens a nursery and then you get this

30:29 nursery object. And the object is tied to that with block. And so, and then once you have that,

30:34 you can say nursery, there's a method on that nursery to spawn a task into it. And all the tasks you

30:38 spawn to the nursery will run concurrently. But then the trick is that that with block,

30:44 its lifetime in the parent is tied to the lifetime of the child tasks. So that you can't exit the

30:50 with block while there's still children running. If you hit the end of the with block, then the

30:53 parent just stops and waits for them. The stuff within the with block could be interwoven

30:58 and like in certain ways, async and concurrent. But taken as a whole, the with block is like a black box.

31:05 You start it, it runs, and then out comes the answer at the end, and everything's done,

31:10 or it's gotten into its canceled state or whatever happens to it, right?

31:13 Exactly. Exactly. Yeah. And so right. And that so that lets us do things like so right. So now,

31:18 if you call a function in Trio, you know, well, okay, it might internally open a nursery and have

31:23 some currency. But by the time it finishes, it's done, it hasn't left anything running. Or if you have,

31:28 if something crashes, we have some child task, you know, has an exception that is no, it doesn't catch,

31:34 then we say, okay, what do we do with this? Oh, well, wait, the parent is just sitting there waiting

31:37 for the child. So that's what we do is the exception hops into the parent and continues executing. So

31:43 in there, sorry, continues like propagating out. So Trio sort of follows the normal Python rule of,

31:48 you know, you can catch an exception, but if you don't, then it will keep propagating until someone

31:52 does. Right. So that nursery is a really important block block here. And I think, you know, it's,

32:00 it's really cool to be able to start all these child tasks and they can even be sort of children

32:05 of the children, right? Like you could pass the nursery and the child task could spawn more child

32:09 tasks and whatever. And so it's all going to sort of be done at the end.

32:15 This portion of talk Python to me has been brought to you by roll bar. One of the frustrating things

32:19 about being a developer is dealing with errors, relying on users report errors, digging through

32:25 log files, trying to debug issues, or getting millions of alerts, just flooding your inbox and ruining your

32:30 day. With roll bars, full stack error monitoring, you get the context insight and control you need to

32:35 find and fix bugs faster. Adding roll bar to your Python app is as easy as pip install roll bar. You can

32:42 start tracking production errors and deployments in eight minutes or less. Are you considering

32:46 self hosting tools for security or compliance reasons? Then you should really check out roll bars,

32:51 compliance SAS option, get advanced security features and meet compliance without the hassle of self

32:57 hosting, including HIPAA ISO 27001 privacy shield and more. They'd love to give you a demo. Give roll bar a

33:05 try today. Go to talkpython.fm/roll bar and check them out.

33:09 The other thing that's really important in this type of threading is to not block forever, to not wait

33:17 on something forever, right?

33:19 Oh yeah, so you're talking about timeouts and cancel scopes.

33:22 Timeouts and cancel scopes and all that stuff.

33:24 Yeah, yeah, yeah.

33:24 Yeah, so this is a really basic problem. If you're doing anything, kind of network programming,

33:29 for example, or any kind of thing, your programs talking to other programs, other resources,

33:36 because sometimes those will just stop responding. So it's like, if you make an HTTP request,

33:42 then maybe it'll succeed. Maybe it'll fail, you'll get a 404 or something like that.

33:48 Or, but there's also a third possibility, which is that it just never finishes, right? Like the

33:54 network disappeared and it just like sits there forever. And your program, yeah. And like, you kind

34:00 of, and I mean, if it's just writing a little script or whatever, it's like that you're running

34:04 at the command line, it's fine. You know, you have some point you'll get bored and hit control C,

34:06 it's fine. But for sort of more complicated systems, you need to be robust against this.

34:11 Yeah.

34:11 Or unattended systems.

34:13 Yeah.

34:13 Like, well, we have 10 workers, but we're only getting this super low throughput. Why? Well,

34:17 because eight of them are like just blocked forever.

34:19 Yeah. You need to somehow, you know, detect this, be able to get out and say, okay, actually,

34:24 let's stop doing that. Try something else, raise an error or something, but like, you know,

34:29 not just sit there forever. So that means, yeah, there's just like, and this is just a problem

34:34 that's just endemic, right? It's just every time you do any kind of networking or other kinds of IPC,

34:40 you need to have this kind of timeout support. So that's one thing that makes it tricky.

34:43 The other thing that makes timeouts tricky is that usually the place where you need to set

34:47 the timeout is like buried way, way down deep inside the code. It's like, at some point your

34:52 HTTP client, you know, is using some socket library and the socket library is like trying to send some

34:58 bytes or read some bytes from the network, like a very, very low level operation. That's probably buried

35:02 under like five levels of abstraction. But then where you want to set the timeout is like way out at the

35:07 top where you say like, I just want to say, if this HTTP request isn't done in 10 seconds,

35:11 then give up. So timeouts are this problem that's like, you need them everywhere and you need

35:16 coordination across different layers of your system. So it's this really complicated problem that like

35:21 covers, you really kind of need your timeout framework that's like used universally by everyone

35:28 and kind of like coordinates, can coordinate across the system.

35:31 And it can be super tricky. Like imagine you want to call like two web services and do a database

35:37 transaction. And if one times out, you want them all to like roll. Yeah. Right. Which is like super,

35:43 like how do you coordinate that? Like there's just, you know, like, yeah, almost impossible. Right. Unless,

35:49 yeah. Unless you put it in trio. So yeah. So right. Yeah. So the traditional way to do this is like,

35:53 you have a timeout framework sort of inside each library, like inside, you know, your HTTP client has a

35:58 timeout argument that you pass or whatever, but that doesn't help with people just don't do that

36:03 reliably. And yeah. And it doesn't solve these coordination problems. So in trio, you say like,

36:07 no, this is something that we're just going to bake into the library, the IO library itself.

36:12 So all libraries that work with trio, if you have an issue that in trio, it uses trio's timeout system.

36:18 And so in this, this thing called cancel scope. So again, there's a lot more details on this in a

36:23 blog post that I wrote called timeouts and cancellation for humans. We'll put the link up there.

36:28 Podcast description, I guess. But in basically, yeah, it's, it's fairly simple to use. Basically,

36:34 the way it works is you can just anywhere in your code, say, you know, with timeout,

36:38 10 seconds, and then you put code inside that with block. And if it takes more than 10 seconds,

36:45 then trio raises a special exception called canceled that just sort of, you know, unwinds out of there.

36:51 And then that with block catches the exception. So, you know, it's a way of saying, okay,

36:55 whatever's happening with block, stop doing that, and then carry on at that point.

37:00 The other nice thing that's important about this, as compared to a lot of other systems we're doing,

37:04 working with timeouts and this cancellation, is that we have the with block to limits, like you can say,

37:10 this is the operation that I want to cancel, which is really important. So like asyncio doesn't work like this.

37:17 So when I think IO, you can say, I want to cancel something, and that injects the special exception,

37:21 but then there's nothing that keeps track of, you can't look at the exception, figure out, okay,

37:25 did I want to cancel this specific network operation, or this HTTP request, or this like entire program,

37:34 like, kind of, you don't keep track of that. When trio, because we said it's with block, we can say,

37:39 okay, I know this is the actual set, this is the operation that I'm trying to cancel right now.

37:44 And these can be nested.

37:46 Yeah, it's awesome. And you, if you want to do like a timeout, you can create a with block that says,

37:50 basically, do this to success or fail after 10 seconds. What is the syntax for that? It's like,

37:57 async with, I forget.

37:58 This one is just a regular with.

38:00 Okay.

38:00 Because, so async with is just like a regular with except that a regular with does, calls a method at

38:06 the beginning and the end of the block. And async with does an await call of a method at the beginning

38:10 of the block. And we talk about a weight being the special, you know, thing we call these special

38:14 functions. Yeah. And so for, it happens that for tree for timeouts, for nurseries, you have to use

38:20 an async with because at the end of the block, it might, the parent might have to wait for the

38:24 children. And you want to let them run while it's waiting. Right. So it has to be an await there.

38:30 So let's use async with for a timeout. You're just, you know, setting something at the beginning

38:34 of the end, but there's, it doesn't, it doesn't have to actually like stop and wait for, let other

38:37 things run there. It can be synchronous. That's just a little detail, but yeah. So it's, it's basically,

38:42 so with block, what we, the main one, sort of this basic one is called move on after to kind

38:46 of remind you that it's going to run the code. And then if the timeout happened, it doesn't

38:51 raise an exception. Like it, you can stop and see, okay, did it, was it canceled or not? But like,

38:55 it keeps executing after the block. So you can like, look at it, which is, this again has to do with,

39:00 like, like the simplest case for a timeout is just like, okay, if the timeout expires,

39:04 blow up everything, give up, raise an exception, right? But a lot of times you want to say, oh,

39:09 well, did this thing get canceled? If so, do some fallback, do something else. So the core thing in

39:16 Trio is, is not to raise an exception after that. It's like to provide some object you can look at to

39:20 see what happened and then figure out how you want to deal with it.

39:23 Yeah. The other thing that I thought you did well with these cancellation scopes is if there's an

39:30 exception on one of the tasks and there's a bunch running in one of these width blocks, you can

39:35 catch that and use that to cancel the other ones and just say, no, no, things are going wrong. We're

39:40 out of here. Every, all threads are done. We're out of here. Yeah. So yeah. So the nurseries actually

39:46 kind of need the cancel scopes because what, one of the things the nurseries do is this exception

39:51 family thing, right? So if an exception happens in a child, it isn't caught, then it has to hop

39:55 into the parent. So we have this kind of, I guess the way I think about it is normally you think of

39:59 like your call stack. It's like this little chain of like, you know, A calls B calls C calls D,

40:03 right? What nurseries do is they kind of make it that into like a tree. So like A calls B, B calls C and D

40:10 and E sort of simultaneously. And then maybe D calls several things at the same time, right? So you have

40:15 this kind of like this tree structure of your, for your call stack. And now some exception happens down

40:20 inside one of those branches. Well, what do you do with exceptions? You unwind the stack,

40:24 right? You go, you run any handlers inside that function. And then you go up to the parent,

40:28 you kill that functions call and you move in there and run any handlers there and so on.

40:33 But then we have this problem. If you're going up, unwinding the stack, and you hit one of these

40:36 places where branch, the stack now branches, how do you sort of unwind past that point? Right?

40:41 You don't want to like orphan the other children, like you just unwind the parent without

40:46 stopping them because that's our whole thing in Trio, right? We say we're not going to let

40:49 orphans run around unsupervised. That's a bad idea. Dijkstra doesn't like that. So what we do is,

40:58 but we have to unwind the stack, right? So what we do is we use the cancellation at that point to go and

41:02 we cancel all the other tasks that are inside that same nursery. So we sort of prune the tree,

41:08 unwind those tasks back up to this point. And then we can keep unwinding from there with the original

41:13 exception. Yeah, that's, that's really clever. I really like the way you put it all together.

41:18 So there's also some other pieces, like some file system bits. I love the, the whole trick of saying

41:26 like a wait sleep for a little bit of time, just go, I'm basically yielding this thread to let it do

41:32 other work, but then just carry on as opposed to say like a time.sleep and things like that.

41:37 So maybe tell us about the higher order of building blocks.

41:39 Yeah. So I guess, yeah, so far we've sort of been talking about kind of the core ideas in Trio, which

41:43 are, you know, not specific to Trio the project. In fact, there's so, I've had a, talking to a number

41:48 of other language designers or other languages, like, oh, this is interesting. We're also struggling with

41:52 these problems. But then, yeah, but Trio is also a specific project you can download. It has

41:57 documentation and API and all that. And one of the things I try to do with that project is to make

42:02 it sort of really sort of usable and accessible, kind of have this philosophy of like, the, you know,

42:07 kind of the buck stops here for developer experience. Like if there's something that you're trying to,

42:12 you're trying to write a program using Trio and there's something awkward or difficult or a problem

42:16 you have, then it's up to us to solve that. And even maybe it's, it's so we, it's, we have the project

42:22 itself. We also have things like that, like, you know, has the core, like networking and concurrency

42:27 stuff. We also have testing helpers and we have a documentation, like a plugins for Sphinx to make

42:34 it easier to document these things. And so there's something called pytest-Trio. How's that work with

42:40 that? So the main thing that pytest-Trio gives you is that you can have, when you write a test, you can say,

42:46 okay, I guess we need to give a little more information here. So Trio, there's sort of a

42:52 stereotype pattern that all Trio programs for, where you have like an async def main or whatever you want

42:57 to call it. That's like your top level async function. That's like, that's where you go into

43:02 that Trio's magic async mode. If you use all the Trio stuff.

43:04 Right. So maybe you start by saying trio.run main.

43:08 Yeah. Yeah. So the thing about async functions, you can only call them from other async functions.

43:12 So you have this problem. How do you call your first async function? And that's what trio.run.

43:17 So you have this little bit of ceremony at the very top of every Trio program. And this is also the case

43:22 for like asyncio and so on. But like, you say, I have to like switch into trio mode. And that's kind of

43:27 annoying in a test suite to do that on every single test. So that's the most basic thing that pytest-Trio

43:31 does for you is it makes it so you can write async def test, whatever, and it will take care of like

43:37 setting up trio and turning it on there. But it also has some other handy stuff. So one thing

43:42 is that it allows you to have async fixtures. And it does some sort of magic to like goes,

43:47 it switches into trio mode, and then sets up your fixtures and then calls your test,

43:50 and then tears down the fixtures. So it's sort of all within this async context.

43:54 And it's also integrated with some of the testing helpers that are built into trio itself.

43:59 So in particular, the one that's sort of the most, you know, gee whiz, awesome,

44:03 is that trio has this ability to use a fake clock. So this is an issue when you're writing like

44:10 networking programs. Like often you want to like have a test about, okay, what does happen if this

44:15 HE request just hangs forever? Like do my timeouts work correctly? Stuff like that, right? But it's

44:20 really annoying to run these tests because it's like, okay, and now I have to sit and wait for a

44:25 minute for the timeout to fire and make sure something happens. And that happens every time you run your

44:29 test. It just spends a minute sitting there doing nothing. And you're like, I really like, okay, but

44:34 like, this is really boring. I want my test suite to finish already. And I need like 100 of these

44:38 tests. And yeah.

44:40 So killing my build time.

44:42 Exactly. Yeah. Right. It's like really ruins your flow. So one thing I've done, actually,

44:48 what I did is when writing trio itself, I said, okay, I really want trio's own test suite to run

44:52 really quickly. I figured that'll force me to like come up with the tools that my users will need to

44:57 make their test suite run quickly. Yeah.

44:58 So if you just type pytest trio, you know, if you want to contribute to trio, we'd love to have

45:03 you. We're very friendly and all that. If you type pytest trio, it runs in like five seconds.

45:07 It has like 99 point something percent test coverage, which is like completely, it's very

45:12 difficult to get there because trio is this really complicated sort of networking library. It's all

45:16 this stuff that's usually hard to test. Part of that is that for all the timeout tests,

45:21 we have this magic clock. And so the way it works is you say, okay, trio, I don't want you to use,

45:26 I know it says like sleep 30 seconds or whatever. I don't want you to actually use sleep 30 real

45:31 seconds. I only want you, I want you to sleep 30 virtual seconds. And so it's a special thing you

45:36 sort of pass the tree at our run to say, every time you have timeouts, sleeping, anything inside this call,

45:41 I want you to sort of use this virtual clock instead. And the way the virtual clock works,

45:45 it starts out at time zero and it just stays there. And you can like advance it manually if you

45:50 want or things like that. But normally what you do is you just use the automatic behavior,

45:54 which is, it's just, it's, it's at time zero. And then it sort of watches what your program is

45:59 doing. Anytime that your program sort of like finishes running everything and just stops and

46:04 is waiting for stuff to happen, then it looks to see, okay, looks like the next thing that would

46:08 wake up, there's like a sleep 10 and sleep 20. Okay. So in 10 seconds, that's the next one that'll wake

46:13 up. I'm just going to jump the clock forward 10 seconds and then start running again. Right.

46:17 So anytime it knows it's going to be waiting for a certain amount, it's like, all right,

46:21 we'll just, we'll let the wait start and then we'll just go right past that.

46:24 So it's basically, yeah, you just write your test the way you normally would with like for use timeouts

46:29 regularly, test your real code, but sleeps, whatever is easiest. And then what's annoying about that

46:34 normally is then your, your test takes like 30 seconds, a minute, whatever it is to run. Most of

46:39 which the time is just sitting there doing nothing, waiting for time to pass. So if you flip the switch to

46:44 use the special clock, then it does this exactly the same things, but it just skips over all those

46:48 times when it's sitting doing nothing. And so suddenly your test runs in like a few milliseconds.

46:52 Oh, that's awesome. Yeah. It's pretty awesome. And then it tests trio is hooked up to that.

46:57 So you can just turn this on with just like flip a switch on any test.

47:00 Oh, that's great. Yeah. So one of the things that makes me a little bit sad about Python's async

47:07 loops and stuff is like the asyncio based apps and the trio based apps. Those are not exactly the same

47:16 and they're not exactly compatible. It's not like the core you're using the same core. And so it just

47:22 keeps running like the asyncio loop and the trio loop. These are not the same. They got to be like

47:26 brought together with different APIs. Right. Yeah.

47:30 But you seem to have, you do have some interoperability. So like trio can work with

47:35 libraries that maybe assume asyncio is there or something, right?

47:38 Trio itself is just like a totally different library than asyncio. I've looked at, you know,

47:43 could I build it on top of asyncio? And there's sort of a number of reasons why that didn't sort of

47:47 make sense. And yes, and there is this big problem because it's just because of technical things about

47:55 how these async concurrency systems work. There has to be like one library ultimately that controls

48:00 all the actual networking like asyncio or trio or whatever, or twisted or tornado or something.

48:06 And that means to like, if you have a, like a say an HTTP client that's written to work on async

48:11 IO, it will necessarily work on trio because it's using a different set of networking primitives underneath

48:16 or vice versa. And this is sort of a, a larger sort of ecosystem problem, right? So there used to be,

48:24 there's twisted and tornado and G event and they all, none of them could interoperate. You'd have to

48:28 like pick which one they're using. And asyncio was sort of one of the reasons it exists is to try

48:33 and solve that problem and become the standard one, but then twisted and tornado and everyone can use.

48:38 And now they can all work on top of asyncio. And now all those libraries written for twisted and tornado,

48:43 you can mix and match however you like. And then here comes trio and kind of ruins that

48:48 by being here's this new thing you should use. So to try and kind of mitigate that,

48:53 there is this library called a trio asyncio, which lets you use asyncio libraries on top of trio.

49:00 The way it does this is kind of, it creates like a virtual asyncio loop that internally uses trios primitives

49:06 under the cover. And it kind of lets you, you know, kind of cordon them off and kind of a little container,

49:11 sort of all the weird stuff they sink IO can do. You can do that stuff, but, the kind of in a little box

49:17 that won't like leak out to pollute the rest of your program, your trio program.

49:22 I think this is really encouraging because that means if you maybe have already invested in AC IO and

49:27 you've already got some code written on it, like you could still Yeah.

49:30 Trio without going, I'm rewriting trio. And is that worth it? Is that a good idea?

49:34 Yeah. Or again, it gives you sort of an incremental path. You can say like, well, okay, I can at least

49:40 get it running on trio first of all, and then I can start porting one piece at a time and eventually

49:44 end up all in trio. Hopefully.

49:46 Exactly.

49:47 Exactly.

49:47 Now, the reason it's not, you can't just magically make this all work because

49:51 trio and Cinco really have fundamentally different ideas about things. Now, obviously, I think trio's

49:56 ideas are better. They're kind of the new thing that I try to fix all these problems, but it's not that

50:02 it's the differences aren't just like in terms of the internal implementation. The differences are in

50:06 terms of just like the fundamental concepts that are exposed.

50:09 Right. Like the philosophy of it.

50:10 Yeah. Right. It totally changes how you write the library on top.

50:14 Right. So it's not something you can just sort of magically switch.

50:17 But there's a little bit of an incremental aspect to it. So we're almost out of time.

50:22 Right.

50:22 Just really quickly, what's the future of trio? Like where is it going? What you got planned?

50:27 And is it production ready?

50:29 So yeah. So I should be clear. Yeah. Right now, the trio library itself is very solid,

50:33 but there is not much of an ecosystem around it. So like there is not currently an HTTP client or

50:39 an HTTP server that you can just use out of the box and it's like mature and all that

50:43 for trio. There are some solutions for these kinds of issues. And I don't want to say too

50:47 much because, you know, this will change quickly. We have a chat channel. If you go to our

50:51 documentation or whatever, you can like find out what the latest news is about what you should use.

50:57 But it's not something that, you know, is ready today to run big websites or something like that.

51:01 Okay.

51:02 Just because the libraries aren't there yet. If you'd like to help, you know, write those libraries and

51:05 make it happen. I'd love to have you. We have a really solid contributing policy and things like

51:11 that. You can check it out. The other thing that's happening is asyncio. So I also, I spent a lot of

51:17 time. I am a core Python developer. I talked to Yuri Selvanov as the main asyncio developer and Guido

51:22 about all this stuff. And so there is this, Yuri is quite keen on saying, oh, well, right. You know,

51:27 trios ideas are better. We should add them all into I sync IO. This is quite complicated. There's a lot

51:32 of, I mean, we could probably do a whole other podcast about all the trade offs there. And maybe

51:36 we should, I don't know. It's pretty interesting. It is interesting.

51:38 Maybe relevant. So that's something that's also happening is that Yuri is going to be trying to

51:42 add nurseries and cancel scopes and things to I sync IO. So I think there's a lot going to be a lot of

51:47 limitations. It's a lot of the value in trios of things people can't do. And I sync IO has already

51:52 got like six layers of abstraction built in there or I don't know. It's not actually six. It's like

51:57 four. And they're all totally doing all the things that trio says, these are, these are things that

52:02 should never be done. It shouldn't be possible. So you can't, that's also, you could fix just adding

52:07 a new layer on top, but you know, it's still better than nothing, right? Like, you know,

52:11 I think it would continue to exist. So we do want to make it as good as possible.

52:15 By these ideas. Yeah, absolutely.

52:16 And ultimately we don't, I mean, maybe like no one knows for sure whether like the make a new thing

52:21 plus a compatibility layer, like trio, the trio, the single thing I mentioned, is that going to be

52:25 the best thing or is making asyncio better going to be the best thing? We'd, none of us know for

52:28 sure. So we are trying both versions. I will sort of see.

52:32 I'm super excited just to hear that that collaboration is happening. I think that's great. All right.

52:38 I think that we're out of time for trio. It's super interesting project. I really love what you've

52:42 done there. I think it's, I think it's brilliant. So, people should definitely check it out.

52:46 Thanks a lot.

52:47 Yeah, you're welcome. So quick two final questions. If you're going to write some Python code,

52:50 what editor do you use?

52:51 I use Emacs. I've been using it for 20 years. I'm stuck.

52:55 It's great. It's not any, I don't know that it works for other people or not just because,

53:00 yeah.

53:01 Yeah, sure. I definitely, I started on Emacs as well. And notable pipe effect.

53:06 Yeah. Well, trio, obviously.

53:07 And, pytest.strio?

53:10 Yeah. pytest trios, Sphinx controlled trio. There's, if you go to the, you know,

53:15 github.com/python-trio to see all the different projects under the trio organization. And they're

53:21 sort of trying to build up that ecosystem. Like I said, so.

53:24 Yeah. Sounds cool. Yeah. So final call to action. People are excited. They want to try trio. Maybe

53:28 they want to contribute to it. What do they do?

53:30 Yeah. So check out, start with the documentation trio.readthedocs.io. That also will give you

53:36 links to our chat is, sort of a place to hang out. it has our contributing docs. If you want

53:41 to get involved like that, we give out commitments on your first pull request acceptance. So there's

53:47 lots of people. Yeah, we, we want, you know, this is a project for everyone. I don't want to just be my,

53:52 you know, personal little thing.

53:53 Yeah, that sounds great. Awesome.

53:55 Yeah.

53:56 All right. Nathaniel, thank you for sharing your project and creating it. It's quite great. And,

54:00 we may have to come back and dig into this a little bit more. This is fun.

54:02 Yeah. Thanks for having me. Yeah. Yeah. Talk to you later.

54:05 You too. Bye bye.

54:05 This has been another episode of talk Python to me. Our guest on this episode was Nathaniel Smith,

54:13 and it was brought to you by Linode and Rollbar. Linode is bulletproof hosting for whatever you're

54:18 building with Python. Get four months free at talkpython.fm/linode. That's L I N O D E.

54:26 Rollbar takes the pain out of errors. They give you the context insight you need to quickly locate and

54:32 fix errors that might have gone unnoticed until your users complain. Of course, as talk Python to me,

54:38 listeners track a ridiculous number of errors for free at rollbar.com/talkpython to me.

54:43 Want to level up your Python? If you're just getting started, try my Python jumpstart by building 10

54:49 apps or our brand new 100 days of code in Python. And if you're interested in more than one course,

54:54 be sure to check out the everything bundle. It's like a subscription that never expires.

54:58 Be sure to subscribe to the show. Open your favorite podcatcher and search for Python. We should be right at the top.

55:04 You can also find the iTunes feed at /itunes, Google Play feed at /play, and direct RSS feed at /rss on talkpython.fm.

55:13 This is your host, Michael Kennedy. Thanks so much for listening. I really appreciate it. Now get out there and write some Python code.

55:19 I'll see you soon.