Cassandra源代码分析：数据写入流程

org.apache.cassandra.thrift.CassandraServer类的add方法将接受客户端的请求，该函数定义如下：

public void add(ByteBuffer key, ColumnParent column_parent, CounterColumn column, ConsistencyLevel consistency_level)

throws InvalidRequestException, UnavailableException, TimedOutException, TException

{

// 数据验证

logger.debug("add");



state().hasColumnFamilyAccess(column_parent.column_family, Permission.WRITE);

String keyspace = state().getKeyspace();



CFMetaData metadata = ThriftValidation.validateColumnFamily(keyspace, column_parent.column_family, true);

ThriftValidation.validateKey(metadata, key);

ThriftValidation.validateCommutativeForWrite(metadata, consistency_level);

ThriftValidation.validateColumnParent(metadata, column_parent);

// SuperColumn field is usually optional, but not when we're adding

if (metadata.cfType == ColumnFamilyType.Super && column_parent.super_column == null)

{

throw new InvalidRequestException("missing mandatory super column name for super CF " + column_parent.column_family);

}

ThriftValidation.validateColumnNames(metadata, column_parent, Arrays.asList(column.name));



// 创建一个 RowMutation 对象,封装用户插入数据信息

RowMutation rm = new RowMutation(keyspace, key);

try

{

rm.addCounter(new QueryPath(column_parent.column_family, column_parent.super_column, column.name), column.value);

}

catch (MarshalException e)

{

throw new InvalidRequestException(e.getMessage());

}

// 插入数据

doInsert(consistency_level, Arrays.asList(new CounterMutation(rm, consistency_level)));

}

函数内部实现上首先将kv信息封装成RowMutation对象，之后创建QueryPath对象（主要是对数据进行封转），

最后调用doInsert方法执行插入动作，doInsert函数定义如下：

// 执行数据插入操作

private void doInsert(ConsistencyLevel consistency_level, List<? extends IMutation> mutations) throws UnavailableException, TimedOutException, InvalidRequestException

{

// 数据验证

ThriftValidation.validateConsistencyLevel(state().getKeyspace(), consistency_level);

if (mutations.isEmpty())

return;

try

{

schedule(DatabaseDescriptor.getRpcTimeout());

try

{

StorageProxy.mutate(mutations, consistency_level);

}

finally

{

release();

}

}

catch (TimeoutException e)

{

throw new TimedOutException();

}

}

函数内部首先进行数据检查，调用StorageProxy.mutate(mutations, consistency_level);执行数据的插入操作。

mute方法定义如下：

public static void mutate(List<? extends IMutation> mutations, ConsistencyLevel consistency_level) throws UnavailableException, TimeoutException

{

logger.debug("Mutations/ConsistencyLevel are {}/{}", mutations, consistency_level);

// 本地数据中心

final String localDataCenter = DatabaseDescriptor.getEndpointSnitch().getDatacenter(FBUtilities.getBroadcastAddress());



long startTime = System.nanoTime();

// 封装条件变量

List<IWriteResponseHandler> responseHandlers = new ArrayList<IWriteResponseHandler>();



IMutation mostRecentMutation = null;

try

{

for (IMutation mutation : mutations) // 对于每个Mutation

{

mostRecentMutation = mutation;

// CounterMutation:首先需要被写入到replicas leader中，之后在向其他的replicas中去分发

if (mutation instanceof CounterMutation)

{

responseHandlers.add(mutateCounter((CounterMutation)mutation, localDataCenter));

}

else

{

// WritePerformer：普通类型的数据分发

responseHandlers.add(performWrite(mutation, consistency_level, localDataCenter, standardWritePerformer));

}

}



// wait for writes. throws TimeoutException if necessary

for (IWriteResponseHandler responseHandler : responseHandlers)

{

// 等待任务结束或者是抛出异常

responseHandler.get();

}



}

catch (TimeoutException ex) // 捕获异常

{

if (logger.isDebugEnabled())

{

List<String> mstrings = new ArrayList<String>();

for (IMutation mutation : mutations)

mstrings.add(mutation.toString(true));

logger.debug("Write timeout {} for one (or more) of: ", ex.toString(), mstrings);

}

throw ex;

}

catch (IOException e)

{

assert mostRecentMutation != null;

throw new RuntimeException("error writing key " + ByteBufferUtil.bytesToHex(mostRecentMutation.key()), e);

}

finally

{

writeStats.addNano(System.nanoTime() - startTime);

}

}

对于每个Mutation对象，如果是CounterMutation类型的Mutation的话，首先要确保一个replica的写入成功，之后在向另外的N-1个replicas写入；其他类型的Mutation的话，没有这个要求，做法是首先得到N个replicas节点，向这个N个节点发送命令。

这两种类型的Mutation是通过两个函数mutateCounter和performWrite分别生成的，这里我们仅仅来看一下performWrite的实现：首先得到复制策略，通过复制策略得到所有replica的endpoints，将任务交给代理WritePerformer.apply执行。代码如下：



public static IWriteResponseHandler performWrite(IMutation mutation,

ConsistencyLevel consistency_level,

String localDataCenter,

WritePerformer performer)

throws UnavailableException, TimeoutException, IOException

{

// 得到复制策略

String table = mutation.getTable();

AbstractReplicationStrategy rs = Table.open(table).getReplicationStrategy();



// 得到所有replica的endpoints

Collection<InetAddress> writeEndpoints = getWriteEndpoints(table, mutation.key());



// 满足一致性的条件变量

IWriteResponseHandler responseHandler = rs.getWriteResponseHandler(writeEndpoints, consistency_level);



// exit early if we can't fulfill the CL at this time

// 如果已经能够确定不能满足一致性的条件，例如live的节点数量小于W，直接返回

responseHandler.assureSufficientLiveNodes();



// 代理给WritePerformer执行

performer.apply(mutation, writeEndpoints, responseHandler, localDataCenter, consistency_level);





return responseHandler;

}

同时需要注意的是在文件org.apache.cassandra.service.StorageProxy.java中有三个实现而来WritePerformer接口的类，WritePerformer接口定义如下：

private interface WritePerformer

{

public void apply(IMutation mutation, Collection<InetAddress> targets, IWriteResponseHandler responseHandler, String localDataCenter, ConsistencyLevel consistency_level) throws IOException, TimeoutException;

}

也就是说最终完成数据写入任务的是WritePerformer的apply方法。StorageProxy的三个实现该接口的类型如下：

// 最终的数据使用实现了WritePerformer接口的standardWritePerformer，counterWritePerformer

// 和counterWriteOnCoordinatorPerformer

standardWritePerformer = new WritePerformer()

{

public void apply(IMutation mutation,

Collection<InetAddress> targets,

IWriteResponseHandler responseHandler,

String localDataCenter,

ConsistencyLevel consistency_level)

throws IOException, TimeoutException

{

assert mutation instanceof RowMutation;

sendToHintedEndpoints((RowMutation) mutation, targets, responseHandler, localDataCenter, consistency_level);

}

};



/*

* We execute counter writes in 2 places: either directly in the coordinator node if it is a replica, or

* in CounterMutationVerbHandler on a replica othewise. The write must be executed on the MUTATION stage

* but on the latter case, the verb handler already run on the MUTATION stage, so we must not execute the

* underlying on the stage otherwise we risk a deadlock. Hence two different performer.

* 执行CounterMutation

*/

counterWritePerformer = new WritePerformer()

{

public void apply(IMutation mutation,

Collection<InetAddress> targets,

IWriteResponseHandler responseHandler,

String localDataCenter,

ConsistencyLevel consistency_level)

throws IOException

{

if (logger.isDebugEnabled())

logger.debug("insert writing local & replicate " + mutation.toString(true));



Runnable runnable = counterWriteTask(mutation, targets, responseHandler, localDataCenter, consistency_level);

runnable.run();

}

};



// 执行CounterMutation

counterWriteOnCoordinatorPerformer = new WritePerformer()

{

public void apply(IMutation mutation,

Collection<InetAddress> targets,

IWriteResponseHandler responseHandler,

String localDataCenter,

ConsistencyLevel consistency_level)

throws IOException

{

if (logger.isDebugEnabled())

logger.debug("insert writing local & replicate " + mutation.toString(true));



Runnable runnable = counterWriteTask(mutation, targets, responseHandler, localDataCenter, consistency_level);

StageManager.getStage(Stage.MUTATION).execute(runnable);

}

};

我们分别来看上面的几个实现，standardWritePerformer的实现方式比较简单，对于endpoints的集合，如果该节点还live，那么其发送写命令，如果该节点dead，那么这时执行hinted-handoff策略：

/**

* Send the mutations to the right targets, write it locally if it corresponds or writes a hint when the node

* is not available.

*

* Note about hints:

*

* | Hinted Handoff | Consist. Level |

* | on | >=1 | --> wait for hints. We DO NOT notify the handler with handler.response() for hints;

* | on | ANY | --> wait for hints. Responses count towards consistency.

* | off | >=1 | --> DO NOT fire hints. And DO NOT wait for them to complete.

* | off | ANY | --> DO NOT fire hints. And DO NOT wait for them to complete.

*

* @throws TimeoutException if the hints cannot be written/enqueued

*/

private static void sendToHintedEndpoints(final RowMutation rm,

Collection<InetAddress> targets,

IWriteResponseHandler responseHandler,

String localDataCenter,

ConsistencyLevel consistency_level)

throws IOException, TimeoutException

{

// Multimap that holds onto all the messages and addresses meant for a specific datacenter

Map<String, Multimap<Message, InetAddress>> dcMessages = new HashMap<String, Multimap<Message, InetAddress>>(targets.size());

MessageProducer producer = new CachingMessageProducer(rm);



for (InetAddress destination : targets) // 对于每个endpoint

{

if (FailureDetector.instance.isAlive(destination)) // 如果endpoint还live

{

String dc = DatabaseDescriptor.getEndpointSnitch().getDatacenter(destination);



if (destination.equals(FBUtilities.getBroadcastAddress()) && OPTIMIZE_LOCAL_REQUESTS)

{

// 如果当前机器就是replicas中的一个，直接写入到本地

insertLocal(rm, responseHandler);

}

else

{

// 否则需要向远程服务器发送命令

// belongs on a different server

if (logger.isDebugEnabled())

logger.debug("insert writing key " + ByteBufferUtil.bytesToHex(rm.key()) + " to " + destination);



Multimap<Message, InetAddress> messages = dcMessages.get(dc);

if (messages == null)

{

messages = HashMultimap.create();

dcMessages.put(dc, messages);

}



messages.put(producer.getMessage(Gossiper.instance.getVersion(destination)), destination);

}

}

else // 否则，这里的话，可能是需要使用hinted-handoff机制


{

if (!shouldHint(destination))

continue;



// Avoid OOMing from hints waiting to be written. (Unlike ordinary mutations, hint

// not eligible to drop if we fall behind.)

if (hintsInProgress.get() > maxHintsInProgress)

throw new TimeoutException();



// Schedule a local hint and let the handler know it needs to wait for the hint to complete too

Future<Void> hintfuture = scheduleLocalHint(rm, destination, responseHandler, consistency_level);

responseHandler.addFutureForHint(new CreationTimeAwareFuture<Void>(hintfuture));

}

}



// 向replicas发送message

sendMessages(localDataCenter, dcMessages, responseHandler);

}

到此我们已经完成了数据从StorageProxy到各个replicas的转发工作，当然这里还存在一些问题，会在下面的继续：

1. 首先replicas收到命令之后的处理动作

2. cassandra中如何生成replicas，如何发现endpoints的拓扑结构，这就涉及到cassandra中snitch的实现

3. cassandra中如何实现DHT？