I just released jruby-memcached 0.1.0 gem, which is the fastest jruby memcached client so far and it is also compatible with memcached.gem. The following is the story why I created jruby-memcached gem.
We are trying to migrate our service from ree to jruby. It's a big project for us, as our repository is written from early 2009, it becomes bigger and bigger, and nobody can promise migrating it to jruby without any errors. Fortunately we are separating our service into different pools, like one pool to handle high scores requests, one pool to handle achievements requests, so our strategy is to migrate to jruby one pool by one pool, it makes migrating processes easier, everytime we only focus on one pool.
This raises one problem, we are using evan's memcached gem which is the fastest memcached client for MRI, but it isn't working in jruby, yes, it's a ruby gem with c extention. But we have to solve the situation that memcached client must work on both ree pool and jruby pool.
The first idea in my mind is to use a jruby memcached client, after googling I found jruby-memcached-client, but soon I get to know they can't be used together. jruby-memcached-client marshal dump the value, encode the value to base64 then save to memcached, and memcached gem only marshal dump, do not encode to base 64. Althrough I can fork and change jruby-memcached-client, but there is still a string issue passing from ruby to jave, I will mention it later.
So I have to give up jruby-memcached-client, then I try to use pure ruby memcached client like memcache-client or dalli. I pick up dalli which is faster than memcached-client, but after we deployed it on production, we found memcached misses jump too high which we can't afford, see
we have to revert the release.
I did some research about memcached, then I realized memcached and dalli are incompatible, why? You may think memcached is simple, just set key/value pair and get value based on key. It is right in common, but if you have more than 1 memcached servers, memcached client should know what key is on which memcached server, client must not fetch the key from server1 now but fetch it from server2 10 minutes later.
Let me introduce you 2 important client configurations
- Hashing, it is the algorithm to convert you string key to long hash.
- Consistent Hashing, aka distribution in memcached gem, it is the algorithm to map you generated hash to one of your memcached servers, it promises low ratio cache reassigns when you add or remove a memcached server. See more about consistent hashing on wikipedia.
memcached uses fnv1_32 as hash algorithm by default and dalli uses crc32, and their distribution algorithm are not compatible as well.
Finally I decided to write a jruby memcached based on a java memcached library by myself. After googling, I find 2 options: xmemcached and spymemcached. The author of xmemcached shows the benchmark which said xmemcached is a bit faster than spymemcached and it provides libmemcached compatible hasing algorithm (although I was cheated in the end :-) ), so I gave it a try first.
xmemcached's LibmemcachedMemcachedSessionLocator is not compatbile with libmemcached, at least not compatible with libmemcached 0.32 which is used by memcached gem. I have to dive into libmemcached 0.32 source code and override xmemcached MemcachedSessionLocator, and write a jruby gem to wrap the xmemcached. (writing a jruby gem is not difficult, I probably write a new post to introduce in the future) Then I released it on our reverse pool, sending high traffic to see the performance, I'm disappointed, memcached get time increased from 30+ ms to 60 ms, and it generated about 200 threads for xmemcached (we have 30 memcached servers and 2 memcached client instances).
Quickly I replaced xmemcached to spymemcached, and memcached get time decreased to 40+ ms and it only generates 2 threads, awesome. And its hash and distribution algorithms are 100% compatible to libmemcached 0.32. You can read the source code in src/main/java to see all hacks I did for spymemcached.
I mention a string issue above, it is when we passing a zlib deflated value, like "x\234c?P??/?I\001\000\b8\002a", it changes to "x?c?P??/?8a" in java, so we can't pass deflated string directly, instead we pass bytes.
I also did some benchmark between memcached, jruby-memcached and dalli.
in ruby-1.9.3 user system total real memcached set 1.110000 1.020000 2.130000 ( 4.592509) memcached get 0.970000 1.000000 1.970000 ( 4.172170) user system total real dalli set 8.330000 1.560000 9.890000 ( 10.094499) dalli get 8.530000 1.680000 10.210000 ( 10.331083) in jruby-126.96.36.199 user system total real jruby-memcached set 6.902000 0.000000 6.902000 ( 6.902000) jruby-memcached get 6.845000 0.000000 6.845000 ( 6.845000) user system total real dalli set 13.251000 0.000000 13.251000 ( 13.251000) dalli get 13.536000 0.000000 13.536000 ( 13.536000)
see more here, as you seen, both memcached and jruby-memcached are 2x faster than dalli.