pydis

pydis is an experiment to disprove some of the falsehoods about performance and optimisation regarding software and interpreted languages in particular.

Below you will find a Redis clone, pydis, written in ~250 lines of idiomatic Python code, providing a subset of redis' functionality for which there are official benchmarks.

Briefly, pydis is ~50% as fast as Redis measured in number operations per second.

P.S. This is not a criticism of Redis, which is a brilliant project and a system-level software that powers thousands of infrastructures. It just happened to be one of the fastest software I could imagine and clone the same day.

Disclaimer

I have used the following libraries written in C for performance:

• uvloop

  uvloop is a fast, drop-in replacement of the built-in asyncio event loop. uvloop is implemented in Cython and uses libuv under the hood. >

• hiredis

  Python extension that wraps protocol parsing code in hiredis.

Discussion

• Hacker News

  1. 2020-11-15 -- Pydis – Redis clone in 250 lines of Python, for performance comparison
  2. 2019-03-02 -- Pydis: Redis clone in Python 3 to make points about performance

• Reddit

  1. 2019-03-01 -- pydis - A redis clone in Python 3 to disprove some falsehoods about performance

Results

redis-benchmark -q -t set,get,incr,lpush,rpush,lpop,rpop,sadd,hset,spop,lrange,mset -n 100000 -P 5 

• 100,000 requests in total per command.
• Requests are pipelined in groups of 5.

Benchmark	pydis	redis	Ratio
SET	271,947	467,361	0.582
GET	274,283	467,237	0.587
INCR	213,409	478,669	0.446
LPUSH	216,082	381,033	0.567
RPUSH	231,143	399,238	0.579
LPOP	248,527	384,332	0.647
RPOP	241,144	429,971	0.561
SADD	219,475	434,257	0.505
HSET	220,178	377,637	0.583
SPOP	288,068	477,705	0.603
LRANGE (100)	26,170	96,254	0.272
LRANGE (300)	9,163	24,768	0.370
LRANGE (500)	5,771	19,351	0.298
LRANGE (600)	4,705	13,869	0.339
MSET	125,215	195,121	0.642

Host System

• Ubuntu 20.04
• Python 3.8.5 (GCC 9.3.0)
• Redis 5.0.7 malloc=jemalloc-5.2.1 bits=64 build=636cde3b5c7a3923

Contributions

Contributions are very welcome, given that they fall into one of the following categories:

• Those that improve the performance.
  • The aim of this exercise is to prove that interpreted languages can be just as fast as C. So whilst using a faster parser in C with Python bindings is okay, rewriting pydis in Cython is not.
  • I will accept "minor" deviations from idioms only if the performance gains are worth it; stick to idiomatic Python otherwise!
• Those to achieve feature parity with Redis for which there are official benchmarks.
  • We are not trying to develop a full-featured Redis clone here so please do not implement commands for which there are no official benchmarks.
• Those that fix formatting etc.
  • Please do not invent your own style, use PEP 8.