memcache

This is the Geni memcached client. It started out as a fork of fiveruns/memcache-client, which was a fork of seattle.rb’s memcache-client, but over time, our client has diverged, and I’ve rewritten the majority of the code. Of course, a lot of credit is due to those whose code served as a starting point for this code. Thanks to fauna for the idea to include the libmemcached source and the code to make it compile when installing the gem. Thanks to develon for the improvements to SegementedServer.

Usage

cache = Memcache.new(:server => "localhost:11211")
cache.set('stuff', [:symbol, 'String', 1, {:bar => 5}])
cache.get('stuff')
=> [:symbol, "String", 1, {:bar => 5}]

cache['things'] = {:foo => '1', :bar => [1,2,3]}
cache['things']
=> {:foo => "1", :bar => [1,2,3]}

How is this different from other memcache clients?

Like fiveruns/memcache-client and fauna/memcached, memcache (shown in italics when I am referring to this library) is a memcached client, but it differs significantly from these clients in several important ways.

Interface

I tried to keep the basic interface as similar as I could to memcache-client. In some cases, memcache can be a near drop-in replacement for memcache-client. However, I did rename the main class from MemCache to Memcache to prevent confusion and to force those switching to memcache to update their code. Here are the notable interface changes:

  • expiry and raw are specified as options in a hash now, instead of as unnamed parameters.

    cache.set('foo', :a,  :expiry => 10.minutes)
    cache.set('bar', :b,  :expiry => Time.parse('5:51pm Nov 24, 2018'))
    cache.set('baz', 'c', :expiry => 30.minutes, :raw => true)
    
  • get_multi has been replaced by a more versatile get interface. If the first argument is an array, then a hash of key/value pairs is returned. If the first argument is not an array, then the value alone is returned.

    cache.get('foo')          # => :a
    cache.get(['foo', 'bar']) # => {"foo"=>:a, "bar"=>:b}
    cache.get(['foo'])        # => {"foo"=>:a}
    
  • get also supports updating the expiry for a single key. this can be used to keep frequently accessed data in cache longer than less accessed data, though usually the memcached LRU algorithm will be sufficient.

    cache.get('foo', :expiry => 1.day)
    
  • Support for flags has been added to all methods. So you can store additional metadata on each value. Depending on which server version you are using, flags can be 16 bit or 32 bit unsigned integers (though it seems that memcache 1.4.1 returns signed values if the upper bit is set).

    cache.set('foo', :aquatic, :flags => 0b11101111)
    value = cache.get('foo')
    => :aquatic
    value.memcache_flags.to_s(2)
    => "11101111"
    
    cache.set('foo', 'aquatic', :raw => true, :flags => 0xff08)
    cache.get('foo', :raw => true).memcache_flags.to_s(2)
    => "1111111100001000"
    
  • In addition to add, which was already supported, support has been added for replace, append and prepend from the memcached protocol.

    cache.add('foo', 1)
    cache.add('foo', 0)
    cache.get('foo')
    => 1
    
    cache.replace('foo', 2)
    cache.get('foo')
    => 2
    
    cache.write('foo', 'bar')     ## shortcut for cache.set('foo', 'bar', :raw => true)
    cache.append('foo', 'none')   ## append and prepend only works on raw values
    cache.prepend('foo', 'foo')   ##
    cache.read('foo')             ## shortcut for cache.get('foo', :raw => true)
    => "foobarnone"
    
  • Support has also been added for cas (compare-and-set).

    value = cache.get('foo', :cas => true)
    cache.cas('foo', value.upcase, :cas => value.memcache_cas)
    cache.get('foo')
    => "FOOBARNONE"
    
    value = cache.get('foo', :cas => true)
    cache.set('foo', 'modified')
    cache.cas('foo', value.downcase, :cas => value.memcache_cas)
    cache.get('foo')
    => "modified"
    
  • Several additional convenience methods have been added including get_or_add, get_or_set, add_or_get, update, get_some, lock, unlock, and with_lock.

Implementation

The underlying architechture of memcache is more modular than memcache-client. A given Memcache instance has a group of servers, just like before, but much more of the functionality is encapsulated inside the Memcache::Server object. Really, a Server object is a thin wrapper around an remote memcached server that takes care of the socket and protocol details along with basic error handling. The Memcache class handles the partitioning algorithm, marshaling of ruby objects and various higher-level methods.

By encapsulating the protocol inside the Server object, it becomes very easy to plug-in alternate backend server implementations. Right now, there are three basic, alternate servers:

LocalServer

This is an in-process server for storing keys and values in local memory. It is good for testing, when you don’t want to spin up an instance of memcached, and also as a second level of caching. For example, in a web application, you can use this as a quick cache which lasts for the duration of a request.

PGServer

This is an implementation of memcached functionality using SQL. It stores all data in a single postgres table and uses PGconn to select and update this table. This works well as a permanent cache or in the case when your objects are very large. It can also be used in a multi-level cache setup with Memcache::Server to provide persistence without sacrificing speed.

NativeServer

This implementation uses native bindings to libmemcached. It is described in more detail in the “Native Bindings” section below.

Very Large Values

Memcached limits the size of values to 1MB. This is done to reduce memory usage, but it means that large data structures, which are also often costly to compute, cannot be stored easily. We solve this problem by providing an additional server called Memcache::SegmentedServer. It inherits from Memcache::Server, but includes code to segment and reassemble large values. Mike Stangel at Geni originally wrote this code as an extension to memcache-client and I adapted it for the new architecture.

You can use segmented values either by passing SegmentedServer objects to Memcache, or you can use the segment_large_values option.

server = Memcache::SegmentedServer.new(:host => 'localhost', :port => 11211)
cache = Memcache.new(:server => server)

cache = Memcache.new(:server => 'localhost:11211', :segment_large_values => true)

Error Handling and Recovery

We handle errors differently in memcache than memcache-client does. Whenever there is a connection error or other fatal error, memcache-client marks the offending server as dead for 30 seconds, and all calls that require that server fail for the next 30 seconds. This was unacceptable for us in a production environment. We tried changing the retry timeout to 1 second, but still found our exception logs filling up with failed web requests whenever a network connection was broken.

So, the default behavior in memcache is for reads to be stable even if the underlying server is unavailable. This means, that instead of raising an exception, a read will just return nil if the server is down. Of course, you need to monitor your memcached servers to make sure they aren’t down for long, but this allows your site to be resilient to minor network blips. Any error that occurs while unmarshalling a stored object will also return nil.

Writes, on the other hand, cannot just be ignored when the server is down. For this reason, every write operation is retried once by closing and reopening the connection before finally marking a server as dead and raising an exception. We will not attempt to read from a dead server for 5 seconds, but a write will always attempt to revive a dead server by attempting to connect.

Keys, Namespaces, and Prefixes

Unlike the other ruby memcache clients, keys in memcache can contain spaces. This is possible because the backend transparently enscapes all space characters, and is especially important if you are using method_cache or record_cache. Memcache::Server implements this escaping using gsub and it adds a slight performance penalty when escaping is necessary. NativeServer implements this escaping directly in C, and the performance overhead is negligible.

You can also partition your keys into different namespaces for convenience. This is done by prefixing all keys in the backend server with “namespace:”. However, the hash keys returned by multi gets do not contain the prefix. In this way, the namespace can be totally transparent to your code. You can also determine whether the prefix is used for hashing with the following option:

hash_with_prefix

Determines whether the prefix/namespace is used when hashing keys to determine which server to use. Defaults to true.

Native Bindings

The Memcache::NativeServer backend provides native bindings to libmecached. This is significantly faster than using Memcache::Server as demonstrated by runnning bench/benchmark.rb. NativeServer encapsulates a set of remote servers and allows you to use the various hashing methods in libmemcached.

You can use native bindings either by passing NativeServer objects to Memcache, or you can use the native option. Native bindings are compatible with segmented values through the SegmentedNativeServer object or by combining the native option with segment_large_values.

server = Memcache::NativeServer.new(:servers => ['localhost:11211', 'localhost:11212'])
cache = Memcache.new(:server => server)

cache = Memcache.new(:servers => ['localhost:11211', 'localhost:11212'], :native => true)

NativeServer also accepts a few other options:

hash

The libmemcached hashing method. See docs.tangent.org/libmemcached/index.html for more detail. One of:

:default :md5 :crc :fnv1_64 :fnv1a_64 :fnv1_32 :fnv1a_32 :jenkins :hsieh :murmur.

NOTE: Even though there is a libmemcached method named :default (which is actually Jenkins’s one-at-a-time hash), the default hashing method if you don’t specify one is :crc.

distribution

The libmemcached distribution method. See docs.tangent.org/libmemcached/index.html for more detail. One of:

:modula :consistent :ketama :ketama_spy

NOTE: :modula is the default. internally, :consistent is an alias for :ketama, and :ketama_spy provides compatibility with the SPY Memcached client for Java.

ketama

Sets the default distribution to :ketama and hash to :md5.

ketama_wieghted

Enables ketama weighting and sets the default distribution to :ketama and hash to :md5.

binary

A boolean value specifying whether to use memcached’s binary protocol instead of the default ascii protocol. This is slightly slower, but should allow you to use unicode keys.

Installation

$ sudo gem install memcache --source http://gemcutter.org

License:

Copyright © 2010 Justin Balthrop, Geni.com; Published under The MIT License, see the LICENSE file.

ext/extconf.rb Copyright © 2010 Cloudburst, LLC, licensed under the AFL3 license, and used with permission; see the ext/LICENSE_AFL3 file.