How do you get PyPy, Django and PostgreSQL to work together?

Question

What fork, or combination of packages should one to use to make PyPy, Django and PostgreSQL play nice together?

I know that PyPy and Django play nice together, but I am less certain about PyPy and PostgreSQL. I do see that Alex Gaynor has made a fork of PyPy called pypy-postgresql. I also know that some people are using psycopg2-ctypes.

Is there a difference between these forks? Or should we use the stable 1.9 PyPy and use psycopg2-ctypes? Using the ctypes options could hurt performance, see the comment below.

Also, has anyone experienced any pitfalls with using PyPy with pyscopg2? It seems easy enough to fall back on CPython if something isn't working right, but mostly I'm looking for things a programmer can do ahead of time to prepare.

I looked around, it doesn't seem that psycopg2 works natively with PyPy. Although, psycopg2-ctypes does seem to be working for some people, there was a discussion on pypy-dev. I work on Windows, and I don't think psycopg2-ctypes is ready for Windows yet, sadly.

Answer 1

psycopg2cffi (Updated 2015)

psycopg2cffi is yet another psycopg2-compatible replacement and should provide the best PostgreSQL performance with PyPy. Add this to your settings.py to remain compatible with both:

try:
    import psycopg2
except ImportError:
    # Fall back to psycopg2cffi
    from psycopg2cffi import compat
    compat.register()

psycopg2-ctypes (2012)

I also know that some people are using psycopg2-ctypes.

This is the easiest way; to stay compatible with both, just add this code in your Django settings.py:

try:
    import psycopg2
except ImportError:
    # Fall back to psycopg2-ctypes
    from psycopg2ct import compat
    compat.register()

I tested this a few releases ago; sadly in my experience, psycopg2-ctypes negates the small performance gains afforded by PyPy. But YMMV, it depends on how JIT-friendly your code is in general and what fraction of time you actually spend running Python code. And maybe PyPy has just improved since then.

and I don't think psycopg2-ctypes is ready for Windows yet

I haven't tried this, but ctypes is platform-independent. AFAICT you just have to make sure that the libpq.dll library is loadable (located in a directory in your PATH environment variable or local directory) and it should work on Windows just like in Linux.

pypy-postgresql

I do see that Alex Gaynor has made a fork of PyPy called pypy-postgresql.

I don't think this is a good choice in the long term. The branch hasn't been updated for more than a year and my attempts to build it have failed. And it seems wrong to hard-code a PostgreSQL driver in the interpreter anyway.

I believe there are no binaries out there of pypy-postgresql either, so if you want to use it, you'd need to build the whole PyPy branch yourself. Not for the faint of heart: it takes tens of minutes and a machine with at least 4 GB of memory. (Official instructions: http://pypy.org/download.html#building-from-source)

To build, you first need the source. If you have Mercurial installed, you can simply hg clone https://bitbucket.org/alex_gaynor/pypy-postgresql. If not, you can download the automagic "tip" zip file: https://bitbucket.org/alex_gaynor/pypy-postgresql/get/tip.zip

Open a command line, go into the decompressed directory, and then inside pypy/translator/goal

If you have PyPy installed, it's recommended to use that for building:

pypy translate.py -Ojit

Otherwise:

python translate.py -Ojit

Sadly this is where my knowledge ends. I get the error "BytecodeCorruption: unimplemented opcode, ofs=234, code=203, name=BUILD_LIST_FROM_ARG"