Use native bytearray truncation

[lovelydinosaur]

Switching from .compress() to using native buffer truncation.

I figure despite some related work on #115 it's worth looking at this PR in isolation.

We probably want this to be blocked on #116.

Benchmarking on Python 3.7...

Before:

$ PYTHONPATH=. venv/bin/python bench/benchmarks/benchmarks.py
6901.9 requests/sec
7055.0 requests/sec
7084.2 requests/sec
7070.6 requests/sec
7107.9 requests/sec
7075.8 requests/sec
7079.4 requests/sec

$ PYTHONPATH=. venv/bin/python bench/benchmarks/benchmarks.py
6972.9 requests/sec
6996.9 requests/sec
6947.3 requests/sec
6999.9 requests/sec
7040.6 requests/sec
7030.0 requests/sec
6997.6 requests/sec

After:

$ PYTHONPATH=. venv/bin/python bench/benchmarks/benchmarks.py
7054.6 requests/sec
7133.3 requests/sec
7148.3 requests/sec
7116.4 requests/sec
7121.0 requests/sec
7128.2 requests/sec
7156.5 requests/sec

$ PYTHONPATH=. venv/bin/python bench/benchmarks/benchmarks.py
7018.9 requests/sec
7074.7 requests/sec
7075.0 requests/sec
7070.6 requests/sec
7028.2 requests/sec
7118.3 requests/sec
7131.2 requests/sec

[lovelydinosaur]

We also ought to follow up on the action in the comment "make sure PyPy also has the optimization", but I'm not sure where to start with that.

[bluetech]

I think it'd be helpful to keep the paragraph above.

[lovelydinosaur]

Fair point, yes. I've addressed it as an inline comment.

[bluetech]

I think a del would be a little clearer but maybe it's just me. (Also below).

[lovelydinosaur]

Agreed, yup!

[bluetech]

We also ought to follow up on the action in the comment "make sure PyPy also has the optimization", but I'm not sure where to start with that.

With this code:

size = 50 * 2**20
b = bytearray(size)
for i in range(size):
    del b[0]

Results:

python3.8:           4.257s
pypy3 version 7.3.2: 0.326s
pypy2 version 7.3.2: 0.326s
python2.7:           DNF

So it's either that pypy has this optimization, or it's doing some magic JIT handling of this specific code. I tried some tricks to foil any such optimization and it seems legit.

If we're willing to have a timing-based unit test for this, a size = 3 * 2**20 with a timeout of 5s would do the trick IMO, even with a super fast CPU running the quadratic case or a very slow CPU running the linear case. But it will probably be flaky anyway...

[njsmith]

I think this has been superseded by #115

[pgjones]

Agree with #115 superseding this. (Please reopen if not).