Excellent! It hadn't occurred to me before that PyPy's parser error messages could be different from CPython's; I must have forgotten that they rewrote the parser [0].
PyPy is a very underrated project. I really pray that after the C-extensions improvements that the upcoming versions brings, a major wave of adoptions come.
I've been programming in Python for 15 years and I've known about PyPy for at least 10, and still haven't really found a reason to use it.
Part of it is because I don't choose what interpreter we use at work. At home, I just use CPython and it works fine for everything.
But I recently started learning more about PyPy, and it's technically crazy and ahead of its time [1]. It started in 2003 and I think something else like it didn't appear for more than 10 years [2], and is still research while PyPy is used in production.
So yeah it is hugely impressive, from a technical standpoint. But I see that there's not enough "carrot" to pay the cost of adoption, for most users. It's faster for some workloads, but predicting the speedup is nontrivial. And it also uses more memory.
It's unfortunate that the Python/C API is so complex and the extension ecosystem is so fragmented with wrappers as a result (SWIG, ctypes, cffi, etc.). I think there is a network effect with CPython that is very hard to break.
Everyone builds for CPython because it already has a large amount of useful C code in its ecosystem -- not to mention stuff like NumPy, Pandas, etc. which are a mix of C and Cython, Cython being a totally separate effort to write "faster Python".
At work, I use PyPy to run a Luigi server for our data pipeline. It gave us a pretty dramatic decrease in latency and allowed us to scale our workload with a trivial change. It's unfortunate you can't choose your interpreter, but when you can it's pretty easy to just test both and see which one is faster.
I also use it at home to solve Project Euler problems. For anything that takes more than half a second or so, it's usually a big win and can turn minutes into seconds.
Do any of the Project Euler questions actually require lengthy computation, or is that more a sign that one should be looking for some mathematical insight that leads to a more efficient computation?
It seems that you're saying that I should be finding mathematical insight instead of using PyPy, while xapata is saying that if my code runs fast in PyPy then I've already found enough mathematical insight.
Most of them don't require lengthy computation, but some of them did for me. I know at least some of the time that's because I didn't know enough math, but sometimes I had the same algorithm as the other people in the solution thread but with much slower runtimes. I think numpy gives a bigger speedup than PyPy for a lot of the problems.
Having a faster interpreter also allows me to play with the numbers more to get a better feel and understanding for the problem. I think that's the point of Project Euler, but I understand your point of view.
The main reason I'm not even considering PyPy is that it lags behind CPython by a couple of versions. I'm getting ready to move to 3.7 while PyPy is still on 3.5.
i'm still waiting for a 100% compatible pypy implementation. People will not switch if "python" can't run on it. Like it or not, but CPython is the standard python.
Anyone know where unladen swallow ended up?
I feel like the improved python interpreter is a pipe dream, but this is coming from someone who has long moved on from python to go due to pythons scaling troubles ( in my unique environment)
> People will not switch if "python" can't run on it.
For my use, it's been remarkably close to CPython. Where does it fall short for you?
I think it's a design goal to match CPython behavior, so I think it would be rare if they were to NOFIX a doesnt-match-CPython bug.
If you're writing anything other than I/O bound code in pure python, you're probably doing it wrong. Write an extension in C or rust for the inner loops and their critical data structures, and keep everything else in python.
In my experience, as someone who runs a fairly hefty Python application, the biggest source of incompatibility is the C extensions. One that's a thorn in my side is psycopg2. There's a version that works with cffi, but then you're crossing your fingers hoping that it matches the main psycopg2 package in terms of functionality and security updates.
I would gladly throw money at someone to figure out how to take the juicy compat bits from Pyston and bring them into Pypy.
I have had similar experiences teaching python; the error messages take a while to be able to decode. One idea I've had is to write an exception handler or wrapper that catches all exceptions and tries to display them in a more beginner friendly way (more human, and with some possible hints to how to fix). Unfortunately never had time to implement it among the actual teaching. This enhancement is a good start, but could be a bit better still with simpler/less technical language. Other common (runtime) errors (e.g. IndexError, TypeError) would also be a good target for this approach.
20 comments
[ 2.9 ms ] story [ 53.4 ms ] thread[0] https://morepypy.blogspot.com/2009/08/pypy-gets-new-compiler...
Part of it is because I don't choose what interpreter we use at work. At home, I just use CPython and it works fine for everything.
But I recently started learning more about PyPy, and it's technically crazy and ahead of its time [1]. It started in 2003 and I think something else like it didn't appear for more than 10 years [2], and is still research while PyPy is used in production.
So yeah it is hugely impressive, from a technical standpoint. But I see that there's not enough "carrot" to pay the cost of adoption, for most users. It's faster for some workloads, but predicting the speedup is nontrivial. And it also uses more memory.
It's unfortunate that the Python/C API is so complex and the extension ecosystem is so fragmented with wrappers as a result (SWIG, ctypes, cffi, etc.). I think there is a network effect with CPython that is very hard to break.
Everyone builds for CPython because it already has a large amount of useful C code in its ecosystem -- not to mention stuff like NumPy, Pandas, etc. which are a mix of C and Cython, Cython being a totally separate effort to write "faster Python".
[1] https://llvm.org/pubs/2006-10-DLS-PyPy.html
[2] Graal and Truffle seems like the closest thing, i.e. a meta-tracing JIT where you get a "JIT for free": https://blog.plan99.net/graal-truffle-134d8f28fb69
http://tratt.net/laurie/research/pubs/html/bolz_tratt__the_i...
I also use it at home to solve Project Euler problems. For anything that takes more than half a second or so, it's usually a big win and can turn minutes into seconds.
As a counter to this, I purposefully try to use CPython in order to force me to find algorithmically optimal solutions. It's a useful way to learn.
Anyone know where unladen swallow ended up?
I feel like the improved python interpreter is a pipe dream, but this is coming from someone who has long moved on from python to go due to pythons scaling troubles ( in my unique environment)
For my use, it's been remarkably close to CPython. Where does it fall short for you?
I think it's a design goal to match CPython behavior, so I think it would be rare if they were to NOFIX a doesnt-match-CPython bug.
If you're writing anything other than I/O bound code in pure python, you're probably doing it wrong. Write an extension in C or rust for the inner loops and their critical data structures, and keep everything else in python.
I would gladly throw money at someone to figure out how to take the juicy compat bits from Pyston and bring them into Pypy.
Then please donate!
https://pypy.org/py3donate.html
Building something compatible with modern Python (especially as new features and syntax seem to be piling up every month) is hard work.
> Anyone know where unladen swallow ended up?
Google killed it. Turns out it wasn't worth the effort.
Dropbox came to the same conclusion when they abandoned Pyston.