predis/FAQ.PERFORMANCES.md

Some frequently asked questions about the performances of Predis

---

Predis is a pure-PHP implementation: it can not be fast enough!

It really depends, but most of the times the answer is: yes, it is fast enough. I will give you a couple of easy numbers using a single Predis client with PHP 5.3.5 (custom build) and Redis 2.2 (localhost) under Ubuntu 11.04 (running on a Intel Q6600):

19600 SET/sec using 12 bytes for both key and value
18900 GET/sec while retrieving the very same values
0.200 seconds to fetch 30000 keys using _KEYS *_.

How does it compare with a nice C-based extension such as phpredis?

30500 SET/sec using 12 bytes for both key and value
31000 GET/sec while retrieving the very same values
0.030 seconds to fetch 30000 keys using "KEYS *"".

Wow, phpredis looks so much faster! Well we are comparing a C extension with a pure-PHP library so lower numbers are quite expected, but there is a fundamental flaw in them: is this really how you are going to use Redis in your application? Are you really going to send thousands of commands in a for-loop for each page request using a single client instance? If so, well I guess you are probably doing something wrong. Also, if you need to SET or GET multiple keys you should definitely use commands such as MSET and MGET. You can also use pipelining to get more performances when this technique can be used.

There is one more thing. We have tested the overhead of Predis by connecting on a localhost instance of Redis, but how these numbers change when we hit the network by connecting to instances of Redis that reside on other servers?

Using Predis:
3600 SET/sec using 12 bytes for both key and value
3600 GET/sec while retrieving the very same values
0.210 seconds to fetch 30000 keys using "KEYS *".

Using phpredis:
4000 SET/sec using 12 bytes for both key and value
4000 GET/sec while retrieving the very same values
0.051 seconds to fetch 30000 keys using "KEYS *".

There you go, you get almost the same average numbers and the reason is quite simple: network latency is a real performance killer and you cannot do (almost) anything about that. As a disclaimer, please remember that we are measuring the overhead of client libraries implementations and the effects of the network round-trip time, we are not really measuring how fast Redis is. Redis shines the best with thousands of concurrent clients doing requests! Also, actual performances should be measured according to how your application will use Redis.

I am convinced, but performances for multi-bulk replies (e.g. KEYS *) are still worse

Fair enough, but there is actually an option for you if you need even more speed and it consists on installing phpiredis (note the additional i in the name) and let Predis using it. phpiredis is a C-based extension that wraps hiredis (the official Redis C client library) with a thin layer that exposes its features to PHP. You will now get the benefits of a faster protocol parser just by adding a single line of code in your application:

$client = new Predis\Client('tcp://127.0.0.1', array(
    'connections' => array('tcp' => 'Predis\Network\PhpiredisConnection')
));

As simple as it is, nothing will really change in the way you use the library in your application. So, how fast is it now? There are not much improvements for inline or short bulk replies (e.g. SET or GET), but the speed for parsing multi-bulk replies is now on par with phpredis:

Using Predis with a phpiredis-based connection to fetch 30000 keys using _KEYS *_:

0.031 seconds from a local Redis instance
0.058 seconds from a remote Redis instance

If I need to install a C extension to get better performances, why not using phpredis?

Good question. Generically speaking, if you need absolute uber-speed using localhost instances of Redis and you do not care about abstractions built around some Redis features such as MULTI / EXEC, or if you do not need any kind of extensibility or guaranteed backwards compatibility with different versions of Redis (Predis currently supports from 1.2 up to 2.2, and even the current development version), then using phpredis can make sense for you. Otherwise, Predis is for you. Using phpiredis gives you a nice speed bump, but it is not mandatory.