how is OOP doing in the "real world"

Stream: general

[Adam Kurkiewicz]

@Moses Schönfinkel @Patrick Massot @Kevin Buzzard

I've had a reality check in the "real world" (Skyscanner, a tech company, ~1000 software engineers, customer-facing money-making product).

Here are the results:

Screen-Shot-2018-04-06-at-3.02.37-PM.png

There was an interesting discussion in the thread, some highlights:

Basically the whole “tom is a type of cat is a type of animal”-style OO has (rightfully) disappeared

If you look at a sensible java/python codebase it doesn’t look so different from the structure of a c/rust/go program

I agree it's a change of mindset. I don't think OO is necessarily less used - but I think it's less often considered The Way to program. More people recognize that it's a tool that has its pros and cons.

Code-wise there's of course a lot of C#/python/java in the company (Skyscanner is historically a .net shop, there is still some ancient VB code running in production), and sadly not very much FP, although I know a few data scientists writing scala code (not sure if that counts as "real" FP), and certainly one python code base makes heavy use of functools. I don't think we have any Haskell whatsoever (could be wrong though, it's a big company!), but I could see it happening in a couple of years.

[Adam Kurkiewicz]

Incidentally, the only person who said that OOP is growing in popularity is an agile coach, not a software engineer...

[Kevin Buzzard]

Thank you very much for this. My remit on this committee is to decide which programming languages that my university should be teaching mathematics undergraduates, and we want to be strongly influenced by what people who hire mathematicians want to see. I am going to teach them Lean but only as an option; the issue is what we force them to learn.

[Kevin Buzzard]

I am pretty sure, however, that suggesting that they learn functional programming will not go down well. I suspect that many of our graduates end up working for banks / hedge funds rather than tech companies.

[Sebastian Ullrich]

Which may be the only sector with a significant amount of Haskell programmers... :)

[Simon Hudon]

I always find this debate a bit odd. Once you know a few languages, learning one more is not a big deal. Learning how to think about your programs though is not easy and is made much harder once you acquired bad thinking habits. Ideally, you'd like to teach a language that promotes good thinking so that you don't have to spend the whole semester making apologies for the language that you chose.

[Kevin Buzzard]

So you're saying I should choose...what? Erlang? ;-)

[Simon Hudon]

In short, is your goal to teach how to use a specific programming language or how to build a certain class of programs or how to program in general?

[Simon Hudon]

I think a case can be made for Erlang. Especially if you teach distributed systems. My personal number one choice is of course Haskell but it also has drawbacks in educational contexts

[Simon Hudon]

Maybe I should switch my #1 choice to Lean. I don't know if I want to do the paper work though :stuck_out_tongue_winking_eye:

[Simon Hudon]

I think OCaml can be a good choice. If you have to teach object oriented programming, I like Eiffel a lot. The tools are not the best though

[Simon Hudon]

@Sebastian Ullrich For imperative programming, would you recommend teaching Rust?

[Sebastian Ullrich]

No, I really don't think that would be a good idea. The learning curve is way too steep.

[Simon Hudon]

What would you recommend?

[Simon Hudon]

Also, is it steeper than Haskell?

[Sebastian Ullrich]

I don't know. We're teaching... Java, still. Well, our group is teaching Haskell and Prolog, but that's at the end of the bachelor's program

[Simon Hudon]

The big problem with teaching Haskell early is that the execution model is hard to grasp so you'd have to wait a few semesters to understand why your programs are slow

[Sebastian Ullrich]

Yes, I'd say it's steeper than Haskell. I feel like with Haskell you can start by using it as "executable math" and worry about things like laziness and monads later. In Rust, you can't do anything useful without understanding references and lifetimes first.

[Sebastian Ullrich]

You also really need the background of knowing other languages to appreciate the things Rust does differently

[Simon Hudon]

I was thinking about that recently, especially with regards to types. I was wondering if you should learn JavaScript before learning Haskell to appreciate how amazing the types are. And then I started wondering, should we do that for every useful invention in the history of programming languages?

[Adam Kurkiewicz]

I very much agree with this:

I always find this debate a bit odd. Once you know a few languages, learning one more is not a big deal. Learning how to think about your programs though is not easy and is made much harder once you acquired bad thinking habits

And that's precisely the reason why I think teaching "java OOP" or "python OOP" to novices is a bad idea. You can spend ages going through language-specific minutiae (mixins, multiple inheritance, generics, type erasure, etc.) without teaching things, which are more important (problem-solving, structuring programs, unit-testing, correctness).

[Adam Kurkiewicz]

I was listening to Simon Peyton Jones' talk in Canterbury last week, which was on high school CS education, and he said something along the lines (I'm paraphrasing): "It's good that the government recognises that we should teach programming in high schools, but what I'm afraid is going to happen is that we will just teach Python". His point was, again, that there is much more to computer science than programming in any specific language (sorted([2, 3, 1]) for example sorts a list in Python, but that fact is entirely educationally useless).

At UoG we cover Python in the first year, but without going too deep into the language, focusing mostly on problem-solving and applications. We do cover all the minutiae of java OOP in the second year, just because we're training future software engineers (and because we have 4 years to produce a graduate in Scotland). But we never cover all the minutiae of Python OOP, since that wouldn't be very educational (students can learn themselves if/when they need it).

[Moses Schönfinkel]

We teach C to the first years, sadly :). Then Java and then Haskell.

[Simon Hudon]

Just because we're training future software engineers

Do you mean by that that you want your graduates to know one common language in depth by the time they graduate so that they're ready to start their likely first job?

[Adam Kurkiewicz]

My specific advise for @Kevin Buzzard on language choice would be: "don't". Jump through any political fences you'll have to jump through and put together a course with your CS department, following their lead on language choice, and working together on curriculum specifics.

The benefits would be enormous:
1. You don't get an outsider to teach the subject. For a student nothing feels worse than a lecturer who doesn't really understand/ like the subject they're teaching.
2. You pave way for increased collaboration with the CS department. CS and maths really need one another. Don't know what the situation is at ICL, but in Glasgow we definitely don't have enough maths-CS collaboration, and it's not just my view, I've discussed this a few times with Muffy Calder for example.

[Adam Kurkiewicz]

@Simon Hudon I think that's more or less it. The pressure from the top is that our graduates should be "industry-ready" by 3rd year (Glasgow is very industry-focused, unlike for example Edinburgh, who get away with teaching Haskell to first years and Coq later in the course), and I imagine knowing one language in-depth is considered part of that "readiness". We do some cool stuff in the 4th year (FP, Constraints Programming, AI), and it's much more "bottom-up" with 4th year modules.

[Simon Hudon]

That's interesting, I didn't know about Edinburgh. Do you have any comparison about how well their graduates do on the job market? Or whether more of their students go to grad school?

[Simon Hudon]

At YorkU, our engineering school is pretty young and we're only now getting started with a software engineering degree. My supervisor was the one behind the push and we caught some flak when the time came for accreditation because too much attention was paid to formal specification. One of the guys doing the accreditation was an agile methodologist

[Simon Hudon]

The funny thing is that I think less and less that the two are in opposition. I feel like the way I use Haskell and Lean is pretty agile / hacker-ish

[Adam Kurkiewicz]

Sure, there is this website called UNISTATS, it uses data from "Destination of Leavers from Higher Education (DLHE)". Not sure if the links will work, but you should be able to search for Computing Science degrees from multiple universities:

1. https://unistats.ac.uk/subjects/employment/10007794FT-G400-2208/ReturnTo/Search
2. https://unistats.ac.uk/subjects/employment/10007790FT-UTCMPSIBENGH/ReturnTo/Search

The results are quite counter-intuitive. It appears that Edinburgh has slightly higher employment rate, and actually more industry employment and less further study.

Maybe students from Edinburgh are like "Enough of this theoretical nonsense, I want to do something real for a change" :D.

Also, the data is almost certainly skewed by JPMorgan. It's the biggest CS employer in Glasgow, easily 30-40% of our graduates go into JP, and they get a good starting salary of 32k or so, which I think is the reason why Glasgow graduates earn better money than Edinburgh.

[Simon Hudon]

Yeah, it's certainly hard to attribute a cause when the cities are not identical. I'd be tempted to say that students from Edinburgh are better equipped for understanding their programs.

[Simon Hudon]

Thanks for the stats

[Simon Hudon]

I was organizing a CS competition to pit various North American universities against each other. I personally studied in Sherbrooke QC which is more industry focused and Montreal (UdeM to be exact) has a lot of theory and math. As I recall, Montreal beat the pants off of everybody there

[Simon Hudon]

Lucky for me, I went in not wearing pants

[Adam Kurkiewicz]

@Kevin Buzzard sadly, the DLHE data can't settle the "hedge fund" vs "tech company" question. The data is not fine-grained enough, with only a very broad category of "Business and public service associate professionals", which is where 65% Mathematics ICL undergraduates land 6-months after their degree: https://unistats.ac.uk/subjects/employment/10003270FT-G100/ReturnTo/Search

Also, the data is 6 months post-graduation, what would be also nice to see is data ~5 years post-graduation, but I'm not sure if it exists.

Anecdotally I know of 1 mathematician in Skyscanner, doing great, and 2 mathematicians at UoG CS, both doing great, but that's not too scientific I'm afraid :/

[Simon Hudon]

That's still interesting. Anecdotally, my supervisor and I get a lot criticism about the way we run our software design class both by the faculty and the students taking but he has the experience that, 5 years down the line, students come back to thank him because it has finally made sense to them and they find our material useful

[Adam Kurkiewicz]

Personally, if I could go back 10 years in time, I would have advised my young and foolish self to consider going to Edinburgh. Although I did switch my main degree to JH in maths + CS at Glasgow, which worked quite well in the end. A fun fact: my wife got unconditional from Edinburgh, but she didn't go because none of her friends went. So I would have to find her too and advise her the same, or we wouldn't have met otherwise :D

[Patrick Massot]

My specific advise for @Kevin Buzzard on language choice would be: "don't". Jump through any political fences you'll have to jump through and put together a course with your CS department, following their lead on language choice, and working together on curriculum specifics.

I'm back from Köln (where I learned interesting stuff about Kevin's youth) and I can only contribute the following advice to Kevin: read the above quote thinking about what you would get if you ask a hardcore (∞, 1)-category theorist what should be taught in mathematics for biology course. And then do the reasonable thing.

[Patrick Massot]

That being said, I don't think you should teach python MRO in first year. There much to learn before that and, by the time you really need this, you can learn alone

[Patrick Massot]

I don't believe in teaching programming without using a programming language (like I don't believe in teaching category theory before linear algebra and elementary group theory)

[Patrick Massot]

So you need to choose a language. Why not going with something fun and actually used?

[Patrick Massot]

This only reason would be duck typing can encourage confused reasonning

[Patrick Massot]

But then you also have your Lean course

[Patrick Massot]

Teaching rigorous thinking

[Patrick Massot]

Also I really have a hard time imagining functional programming scaling to real world programming but I'd love to learn I'm wrong (I read a book on Haskell recently and it only seemed to be about solving problems created by the functional paradigm). My experience in real world is that you quickly get tired of adding millions of arguments to functions in order to get flexibility. Compare with the beginning of https://docs.djangoproject.com/fr/2.0/topics/class-based-views/intro/ for instance. But again this is only my amateur end-user view-point, I'm neither a computer scientist nor a professional programmer.

[Adam Kurkiewicz]

Re "real world" haskell usage: https://code.facebook.com/posts/745068642270222/fighting-spam-with-haskell/

Re (∞, 1)-category theorist -- not sure what ICL compsi looks like, but at Glasgow we have plenty of down-to-earth engineering-type people I would trust to do a decent job with co-teaching such course. That being said we have of course a fair share of "(∞, 1)-category theorists", so it's definitely something to be wary of.

[Moses Schönfinkel]

There are definitely examples of people using Haskell for real world stuff but OOP is in overwhelming majority. (Disclaimer: I hate OOP and I write functional exclusively except for when I teach where I am coerced to do Java :(.)

[Adam Kurkiewicz]

No disagreement that popularity(OOP) > popularity(FP).

But if we look at the derivative...

The sentiment in industry (or certainly the slice of industry I'm a part of) is that there is a swing from "rich OOP" (Tom is a Cat is an Animal or SimpleBeanFactoryAwareAspectInstanceFactory), back to more readable procedural/ imperative style. FP is still a bit in the shadows unfortunately, but pace of change in software engineering is very quick. I wouldn't be surprised if we saw internal Haskell products popping up in major tech companies few years down the line.

[Kevin Buzzard]

I guess I am interested in these derivatives as well as popularity, but I guess I am only interested in the industries that are hiring mathematicians.

[Kevin Buzzard]

(and then I am only interested in what they want the mathematicians to do)

[Adam Kurkiewicz]

I'm afraid you'll have to search more then. I have only 5 mathematicians doing non-mathematics in my professional network, 2.5 of them all seem to be doing what everybody else is doing (writing software, teaching web-dev, writing Integer Programming models, managing software developers), 2.5 of them work at the edge of pure maths/CS (i.e. proving theorems, not writing code).

[Kevin Buzzard]

What languages are the 2.5 writing software using?

[Kevin Buzzard]

Maybe that's not the right question.

[Kevin Buzzard]

Maybe I should just note that they are writing software

[Adam Kurkiewicz]

python, python, java. Certainly not. I will have a lot people writing python in my professional network :)

[Adam Kurkiewicz]

Likewise, I will have a lot of people writing software in my professional network. I think this is called Prosecutor's Fallacy or something like that.

[Adam Kurkiewicz]

What you need is better DLHE data. As it is DLHE is just too coarse: https://unistats.ac.uk/subjects/employment/10003270FT-G100/ReturnTo/Search