缓存子系统如何设计(Cachable tag, Memcache/redis support, xml config support, LRU/LFU/本地缓存命中率)

大家对这段代码肯定很熟悉吧：

public List<UserInfo> SearchUsers(string userName)
        {
            string cacheKey=string.Format("SearchUsers_{0}", userName);
            List<UserInfo>  users = cache.Find(cacheKey) as List<UserInfo>;
            if (users == null)
            {
                users = repository.GetUsersByUserName(userName);
                cache.Set(cacheKey, users);
            }
            return users;
        }

class HttpRuntimeCache
    {
        public object Find(string key)
        {
            return HttpRuntime.Cache[key];
        }
        public void Set(string key, object value)
        {
            HttpRuntime.Cache[key] = value;
        }
    }

导致了如下这些问题：

1. 业务逻辑函数中引入了很多无关的缓存代码，导致DDD模型不够纯
2. 更换缓存Provider不方便
3. 加入缓存冗余机制不方便
4. 没办法同时使用多个缓存系统
5. 缓存大对象出现异常，比如Memcache有1M的value限制

有诸多问题，因此我们需要引入缓存子系统来解决上述问题，带来的好处：

1. DDD模型更加纯
2. 具体的Cache实现机制可以很灵活，比如HttpRuntimeCache, Memcache, Redis可以同时使用
3. 加入了Cache冗余机制，不会由于某一台Memcache或者Redis down机导致系统速度很慢，实际上，系统还是会保持飞快（除非backup也down了的情况）
4. 开发人员更加致力于核心业务，不会分散注意力
5. 缓存位置透明化，都会在xml配置文件中进行配置

解决方案，要用到这2篇文章的技术：C# 代理应用 - Cachable 和 聊聊Memcached的应用。 

主要的思路分2个：

模型端：通过代理来嵌入AOP方法，来判断是否需要缓存，有缓存value则直接返回value；缓存value的写入是通过AOP的后置方法写入的，因此不需要在业务函数中写代码，当然也支持代码调用。

Cache核心对象：这个对象要解决一致性hash算法、cache value大对象分解功能、冗余机制

代理嵌入AOP的方法，已经在这篇文章中说明了 C# 代理应用 - Cachable，有兴趣的看看，这里就不说了，我们来主要看看CacheCoordinator对象的实现

结构图如下：

先来看看UML图：

CacheCore代码（算法核心）：

public class CacheCore
    {
        private ICacheCoordinator cacheProvider = null;
        public CacheCore(ICacheCoordinator cacheProvider)
        {
            this.cacheProvider = cacheProvider;
        }

        public void Set(string location, string key, object value)
        {
            AssureSerializable(value);
            string xml = Serializer2XMLConvert(value);
            CacheParsedObject parsedObj = new CacheParsedObject();

            string classType = string.Format("{0}", value.GetType().FullName);
            if (xml.Length > CacheConfig.CacheConfiguration.MaxCacheEntitySize)
            {
                /*
                    key:1@3@ConcreteType
                    key_1:subvalue1
                    key_2:subvalue2
                    key_3:subvalue3
                */
                //拆分成更小的单元
                int splitCount = xml.Length / CacheConfig.CacheConfiguration.MaxCacheEntitySize;
                if (CacheConfig.CacheConfiguration.MaxCacheEntitySize * splitCount < xml.Length)
                    splitCount++;
                parsedObj.MainObject = new KeyValuePair<string, string>(key, string.Format("1@{0}@{1}", splitCount, classType));
                for (int i = 0; i < splitCount;i++ )
                {
                    if (i == splitCount - 1)  //最后一段，直接截取到最后，不用给出长度
                        parsedObj.SplittedElements.Add(xml.Substring(i * CacheConfig.CacheConfiguration.MaxCacheEntitySize));
                    else                      //其他，要给出长度
                        parsedObj.SplittedElements.Add(xml.Substring(i * CacheConfig.CacheConfiguration.MaxCacheEntitySize, CacheConfig.CacheConfiguration.MaxCacheEntitySize));
                }
            }
            else
            {
                /*
                    key:1@1@ConcreteType
                    key_1:value
                */
                parsedObj.MainObject = new KeyValuePair<string, string>(key, string.Format("1@1@{0}", classType));
                parsedObj.SplittedElements.Add(xml);
            }

            //针对CacheParsedObject进行逐项保存
            this.cacheProvider.Put(parsedObj.MainObject.Key, parsedObj.MainObject.Value);
            int curIndex = 0;
            foreach(string xmlValue in parsedObj.SplittedElements)
            {
                curIndex++;
                string tkey=string.Format("{0}_{1}", parsedObj.MainObject.Key, curIndex);
                this.cacheProvider.Put(tkey, xmlValue);
            }
        }

        public object Get(string location, string key)
        {
            string mainObjKeySetting = (string)cacheProvider.Get(key);
            if (mainObjKeySetting == null || mainObjKeySetting.Length == 0)
                return null;

            string classType;
            CacheParsedObject parsedObj;
            GetParsedObject(key, mainObjKeySetting, out classType, out parsedObj);

            string xmlValue=string.Empty;
            parsedObj.SplittedElements.ForEach(t=>xmlValue+=t);

            using (StringReader rdr = new StringReader(xmlValue))
            {
                //Assembly.Load("Core");
                Type t = Type.GetType(classType);
                XmlSerializer serializer = new XmlSerializer(t);
                return serializer.Deserialize(rdr);
            }
        }

        public void Remove(string location, string key)
        {
            string mainObjKeySetting = (string)cacheProvider.Get(key);
            if (mainObjKeySetting == null || mainObjKeySetting.Length == 0)
                return;

            string classType;
            CacheParsedObject parsedObj;
            GetParsedObject(key, mainObjKeySetting, out classType, out parsedObj);

            int i = 1;
            parsedObj.SplittedElements.ForEach(t => this.cacheProvider.Remove(string.Format("{0}_{1}", parsedObj.MainObject.Key, i++)));
            this.cacheProvider.Remove(parsedObj.MainObject.Key);
        }
        private void GetParsedObject(string key, string mainObjKeySetting, out string classType, out CacheParsedObject parsedObj)
        {
            int from = 1, end = 1;
            classType = string.Empty;
            if (mainObjKeySetting.IndexOf('@') > 0)
            {
                end = int.Parse(mainObjKeySetting.Split('@')[1]);
                classType = mainObjKeySetting.Split('@')[2];
            }

            parsedObj = new CacheParsedObject();
            parsedObj.MainObject = new KeyValuePair<string, string>(key, string.Format("1@{0}@{1}", end, classType));
            for (int i = from; i <= end; i++)
                parsedObj.SplittedElements.Add((string)this.cacheProvider.Get(string.Format("{0}_{1}", parsedObj.MainObject.Key, i)));
        }
        private string Serializer2XMLConvert(object value)
        {
            using (StringWriter sw = new StringWriter())
            {
                XmlSerializer xz = new XmlSerializer(value.GetType());
                xz.Serialize(sw, value);
                return sw.ToString();
            } 
        }
        private void AssureSerializable(object value)
        {
            if (value == null)
                throw new Exception("cache object must be Serializable");
            if (value.GetType().GetCustomAttributes(typeof(SerializableAttribute), true).Count()<=0)
                throw new Exception("cache object must be Serializable");
        }
    }

下面是CacheCoordinator的代码，这个类的加入目的是要加入缓存的冗余机制：

class CacheCoordinator : ICacheCoordinator
    {
        CacheServerWrapper backupCacheServer = new CacheServerWrapper(CacheConfig.CacheConfiguration.BackupCacheServer);
        CacheServersWrapper peerCacheServer = new CacheServersWrapper(CacheConfig.CacheConfiguration.PeerCacheServers);

        public void Put(string key, object value)
        {
            peerCacheServer.Put(key, value); 
            backupCacheServer.Put(key, value); //缓存冗余
        }

        public object Get(string key)
        {
            object o=peerCacheServer.Get(key);
            if (o != null)
                return o;
            return backupCacheServer.Get(key);
        }

        public void Remove(string key)
        {
            peerCacheServer.Remove(key);
            backupCacheServer.Remove(key);
        }
    }

剩下的就是具体的CacheProvider和CacheProviderWrapper类了：

public class CacheServerWrapper : ICacheExecutor
    {
        ICacheExecutor executor = null;
        private CacheServerInfo configInfo;
        public CacheServerWrapper(CacheServerInfo configInfo)
        {
            this.configInfo = configInfo;
            ICacheExecutor tmpExecutor = null;
            switch(this.configInfo.ServerType)
            {
                case CacheServerType.HttpRuntime:
                    tmpExecutor = new CacheProvider.HttpRuntimeCacheProvider(configInfo);
                    break;
                case CacheServerType.InMemory:
                    tmpExecutor = new CacheProvider.InMemoryCacheProvider(configInfo);
                    break;
                case CacheServerType.Memcached:
                    tmpExecutor = new CacheProvider.MemcachedCacheProvider(configInfo);
                    break;
                case CacheServerType.Redis:
                    tmpExecutor = new CacheProvider.RedisCacheProvider(configInfo);
                    break;
                default:
                    tmpExecutor = new CacheProvider.HttpRuntimeCacheProvider(configInfo);
                    break;
            }
            executor = tmpExecutor;
        }

        public string FullServerAddress
        {
            get
            {
                return this.configInfo.FullServerAddress;
            }
        }

        public void Put(string key, object value)
        {
            executor.Put(key, value);
        }

        public object Get(string key)
        {
            return executor.Get(key);
        }

        public void Remove(string key)
        {
            executor.Remove(key);
        }
    }

只贴出Memcache的操作类

class MemcachedCacheProvider : ICacheExecutor
    {
        private MemcachedClient mc = new MemcachedClient();
        private CacheServerInfo configInfo;
        public MemcachedCacheProvider(CacheServerInfo configInfo)
        {
            this.configInfo = configInfo;

            //初始化池  
            SockIOPool pool = SockIOPool.GetInstance();
            pool.SetServers(new string[] { string.Format("{0}:{1}", configInfo.ServerAddress, configInfo.ServerPort) });//设置连接池可用的cache服务器列表，server的构成形式是IP:PORT（如：127.0.0.1:11211）  
            pool.InitConnections = 3;//初始连接数  
            pool.MinConnections = 3;//最小连接数  
            pool.MaxConnections = 5;//最大连接数  
            pool.SocketConnectTimeout = 1000;//设置连接的套接字超时  
            pool.SocketTimeout = 3000;//设置套接字超时读取  
            pool.MaintenanceSleep = 30;//设置维护线程运行的睡眠时间。如果设置为0，那么维护线程将不会启动,30就是每隔30秒醒来一次  

            //获取或设置池的故障标志。  
            //如果这个标志被设置为true则socket连接失败，将试图从另一台服务器返回一个套接字如果存在的话。  
            //如果设置为false，则得到一个套接字如果存在的话。否则返回NULL，如果它无法连接到请求的服务器。  
            pool.Failover = true;

            pool.Nagle = false;//如果为false，对所有创建的套接字关闭Nagle的算法  
            pool.Initialize();
        }
        public void Put(string key, object value)
        {
            mc.Set(key, value);
        }

        public object Get(string key)
        {
            return mc.Get(key);
        }

        public void Remove(string key)
        {
            mc.Delete(key);
        }
    }

不能忘了可配置性，xml定义及代码如下：

<?xml version="1.0" encoding="utf-8" ?>
<CacheConfig>
  <MaxCacheEntitySize>1048576</MaxCacheEntitySize><!--1*1024*1024-->
  <PeerCacheServers>
    <CacheServer>
      <ServerType>InMemory</ServerType>
      <ServerAddress>127.0.0.1</ServerAddress>
      <ServerPort>11211</ServerPort>
    </CacheServer>
    <CacheServer>
      <ServerType>InMemory</ServerType>
      <ServerAddress>127.0.0.1</ServerAddress>
      <ServerPort>11212</ServerPort>
    </CacheServer>
  </PeerCacheServers>
  <BackupCacheServer>
    <CacheServer>
      <ServerType>InMemory</ServerType>
      <ServerAddress>127.0.0.1</ServerAddress>
      <ServerPort>11213</ServerPort>
    </CacheServer>
  </BackupCacheServer>
</CacheConfig>

读取配置信息的代码：

public static class CacheConfiguration
    {
        static CacheConfiguration()
        {
            Load();
        }

        private static void Load()
        {
            PeerCacheServers = new List<CacheServerInfo>();
            BackupCacheServer = null;

            XElement root = XElement.Load(System.IO.Path.Combine(AppDomain.CurrentDomain.BaseDirectory, "CacheConfig.xml"));

            MaxCacheEntitySize = int.Parse(root.Element("MaxCacheEntitySize").Value);
            foreach (var elm in root.Element("PeerCacheServers").Elements("CacheServer"))
            {
                CacheServerInfo srv = new CacheServerInfo();
                srv.ServerAddress = elm.Element("ServerAddress").Value;
                srv.ServerPort = int.Parse(elm.Element("ServerPort").Value);
                srv.ServerType = (CacheServerType)Enum.Parse(typeof(CacheServerType), elm.Element("ServerType").Value);
                PeerCacheServers.Add(srv);
            }
            foreach (var elm in root.Element("BackupCacheServer").Elements("CacheServer"))
            {
                CacheServerInfo srv = new CacheServerInfo();
                srv.ServerAddress = elm.Element("ServerAddress").Value;
                srv.ServerPort = int.Parse(elm.Element("ServerPort").Value);
                srv.ServerType = (CacheServerType)Enum.Parse(typeof(CacheServerType), elm.Element("ServerType").Value);
                BackupCacheServer = srv;
                break;
            }
            if (PeerCacheServers.Count <= 0)
                throw new Exception("Peer cache servers not found.");
            if (BackupCacheServer == null)
                throw new Exception("Backup cache server not found.");
            AssureDistinctFullServerAddress(PeerCacheServers);
        }

        private static void AssureDistinctFullServerAddress(List<CacheServerInfo> css)
        {
            Dictionary<string, int> map = new Dictionary<string, int>();
            foreach(CacheServerInfo csInfo in css)
            {
                if (map.ContainsKey(csInfo.FullServerAddress))
                    throw new Exception(string.Format("Duplicated server address found [{0}].", csInfo.FullServerAddress));
                else
                    map[csInfo.FullServerAddress] = 1;
            }
        }
        public static int MaxCacheEntitySize { get; set; }
        public static List<CacheServerInfo> PeerCacheServers { get; set; }
        public static CacheServerInfo BackupCacheServer { get; set; }
    }

代码下载

Append New

其实，我们忽略了一些重要的东西：

1. 如果Memcached, Redis服务器超过了5台以上，通信量上升很快，怎么办？
2. 由于取数据牵涉到网络I/O操作，因此速度依然比较慢，怎么办？

让我们来解决吧。

把新的UML图贴上（下图中左边红框中的是新增的）：

本地缓存替换策略：LFU/LRU，其他的有很多。

EventBus是分布式的，下面有讲为什么要分布式的。

当Domain层需要获取数据时的逻辑：

1. 先查看本地缓存中是否存在数据副本，存在则立刻返回（也没有网络I/O了）
2. 没有则去redis/memcached获取，有则返回；并且把数据放入本地cache中
3. 最后，实在没有数据，就db里取

当Domain层需要更新数据时的逻辑：

1. 在本地cache中进行更新操作
2. 更新分布式缓存
3. 发布分布式事件，通知其他app server的cache manager去主动拉数据到他们本地缓存

看得出来，加入这个新的角色后，能对下面2项有改善作用：

1. 降低网络间的通信流量
2. 增大本地缓存的命中率