Patch meiheld

[waghanza]

Hi,

Following bottlepy/bottle#1091, this PR patch meinheld.

Thanks @agalera 🎉

Regards,

[waghanza]

I use this feature also for flask

Framework (Middleware)	Average	50% percentile	90% percentile	99% percentile	99.9% percentile	Standard deviation	Requests / s	Throughput
bottle (unpatched)	22.18 ms	18.03 ms	37.88 ms	75.21 ms	324.32 ms	15796.33	48441.00	119.39 MB
flask (unpatched)	52.62 ms	37.67 ms	110.96 ms	158.00 ms	632.17 ms	41189.67	20414.67	50.18 MB
bottle (patched)	21.00 ms	13.69 ms	41.86 ms	90.58 ms	282.19 ms	18187.67	53832.67	132.67 MB
flask (patched)	54.46 ms	38.40 ms	109.36 ms	166.15 ms	394.67 ms	36173.33	19572.00	48.09 MB

For bottle this patch is 🎉 but not sure if flask should be patched

[waghanza]

@agalera @benoitc any idea about the performance diff ? any help will be ❤️

[agalera]

in bottle should not be patched for errors in the response (bottlepy/bottle#896).
Bottle is much faster than Flask, this is normal.

If you want to improve the performance by returning answers in json format, you can install ujson is an optional requirement that makes bottle serialize much faster.

@defnull maybe I can contribute something more

[defnull]

Bottle is a lot faster than flask, this is also visible in other benchmarks. And yes, ujson is automatically picked up, if present, and used for faster json processing. That's a cheap boost for json-heavy APIs.

I'm not sure if this would be fair in a benchmark, though. Bottle is advertised as a zero-dependency framework, after all. And already faster than most ;)

[waghanza]

@defnull Why ? Flask and bottle are python framework and their are running both under gunicorn and meinheld

[waghanza]

@agalera In fact, we wan this benchmark as simple as possible, as for now we are just using 3 rules, see #221 (comment)

[defnull]

Bottle does only whats necessary and is written in a very dense, minimalistic way. I have also spent a lot of time to tune the important parts. For example, bottle bundles route regular expressions and can check up to 100 routes with a single re.find call. It's carefully optimized for speed from day one.

Flask, on the other hand, is based on werkzeug, which is a great peace of software, but kind of over-engineered in many places. Lots of classes, abstractions, small functions, deep call stacks. This overhead adds up. In a compiled language, optimizers would get rid of the abstraction overhead. In Python, abstractions always come at a cost.

There are faster WSGI frameworks than Bottle, but most of them 'cheat' by implementing parts in Cython or wrapping C libraries, or do way less than Bottle and are IMHO less useful.

[waghanza]

I understand the performance gap, but it's always useful to have some figures.