HiveMetastore 的架构简析
HiveMetastore 的架构简析
Hive Metastore 是 Hive 元数据管理的服务。可以把元数据存储在数据库中。对外通过 api 访问。
hive_metastore.thrift
对外提供的 Thrift 接口定义在文件 standalone-metastore/src/main/thrift/hive_metastore.thrift 中。
内容包括用到的结构体和枚举，和常量，和 rpc Service。
如分区定义如下：
struct Partition {1: list<string> values // string value is converted to appropriate partition key type2: string dbName,3: string tableName,4: i32 createTime,5: i32 lastAccessTime,6: StorageDescriptor sd,7: map<string, string> parameters,8: optional PrincipalPrivilegeSet privileges,9: optional string catName
}
Service的定义了 client 和 server 的 RPC 请求。如增加分区的定义如下：
service ThriftHiveMetastore extends fb303.FacebookService
{Partition add_partition(1:Partition new_part)throws(1:InvalidObjectException o1, 2:AlreadyExistsException o2, 3:MetaException o3)
}
ThriftHiveMetastore.java
hive_metastore.thrift 编译之后生成文件 ThriftHiveMetastore.java
ThriftHiveMetastore 结构如下：
public class ThriftHiveMetastore {// Iface 定义了 Service 的所有的接口。仅列出了 add_partitionpublic interface Iface extends com.facebook.fb303.FacebookService.Iface {public Partition add_partition(Partition new_part) throws InvalidObjectException, AlreadyExistsException, MetaException, org.apache.thrift.TException;// omit other methods}// AsyncIface 定义了异步接口。public interface AsyncIface extends com.facebook.fb303.FacebookService .AsyncIface {public void add_partition(Partition new_part, org.apache.thrift.async.AsyncMethodCallback resultHandler) throws org.apache.thrift.TException;// omit other methods}// Client 的实现public static class Client extends com.facebook.fb303.FacebookService.Client implements Iface {public Partition add_partition(Partition new_part) throws InvalidObjectException, AlreadyExistsException, MetaException, org.apache.thrift.TException{send_add_partition(new_part);return recv_add_partition();}// receiveBase 调用 result.read 方法，从 protocalpublic Partition recv_add_partition() throws InvalidObjectException, AlreadyExistsException, MetaException, org.apache.thrift.TException{add_partition_result result = new add_partition_result();receiveBase(result, "add_partition");if (result.isSetSuccess()) {return result.success;}if (result.o1 != null) {throw result.o1;}if (result.o2 != null) {throw result.o2;}if (result.o3 != null) {throw result.o3;}throw new org.apache.thrift.TApplicationException(org.apache.thrift.TApplicationException.MISSING_RESULT, "add_partition failed: unknown result");}}// Processor 是服务端处理框架，把所有要处理的 rpc 的名称和处理的映射放到 map里，客户端请求 rpc，先输出 rpc 的名称。public static class Processor<I extends Iface> extends com.facebook.fb303.FacebookService.Processor<I> implements org.apache.thrift.TProcessor {private static <I extends Iface> Map<String, org.apache.thrift.ProcessFunction<I, ? extends org.apache.thrift.TBase>> getProcessMap(Map<String, org.apache.thrift.ProcessFunction<I, ? extends org.apache.thrift.TBase>> processMap) {processMap.put("add_partition", new add_partition());}}// AsyncProcessorpublic static class AsyncProcessor<I extends AsyncIface> extends com.facebook.fb303.FacebookService.AsyncProcessor<I> {}
}
IHMSHandler
IHMSHandler 是服务器端需要实现的接口。除了 ThriftHiveMetastore.Iface 外，还包括其他一些方法。
IHMSHandler extends ThriftHiveMetastore.Iface, Configurable
HMSHandler
HMSHandler 是服务器端的具体实现。创建服务时，创建 HMSHandler，多个线程调用同一个 HMSHandler 对象来处理 client 的请求。
public class HMSHandler extends FacebookBase implements IHMSHandler {@Overridepublic Partition add_partition(final Partition part)throws InvalidObjectException, AlreadyExistsException, MetaException {return add_partition_with_environment_context(part, null);}// omit other methods@Overridepublic Partition add_partition_with_environment_context(final Partition part, EnvironmentContext envContext)throws InvalidObjectException, AlreadyExistsException,MetaException {startTableFunction("add_partition",part.getCatName(), part.getDbName(), part.getTableName());Partition ret = null;Exception ex = null;try {ret = add_partition_core(getMS(), part, envContext);} catch (Exception e) {ex = e;if (e instanceof MetaException) {throw (MetaException) e;} else if (e instanceof InvalidObjectException) {throw (InvalidObjectException) e;} else if (e instanceof AlreadyExistsException) {throw (AlreadyExistsException) e;} else {throw newMetaException(e);}} finally {endFunction("add_partition", ret != null, ex, part != null ? part.getTableName(): null);}return ret;}
}
getMS 从 ThreadLocal 里，每个线程单独的。
@Overridepublic RawStore getMS() throws MetaException {Configuration conf = getConf();return getMSForConf(conf);}public static RawStore getMSForConf(Configuration conf) throws MetaException {RawStore ms = threadLocalMS.get();if (ms == null) {ms = newRawStoreForConf(conf);ms.verifySchema();threadLocalMS.set(ms);ms = threadLocalMS.get();}return ms;}
为什么 Handler 总是一个线程处理一个 client 的请求
如果不是一个线程处理一个 client 的请求，那么 client 先发送一个请求，然后再发送第二个请求时， RawStore ms = threadLocalMS.get(); 就有可能拿到的是其他线程的 ms。
因为 org.apache.thrift.server.TThreadPoolServer.serve方法中。为每个 socket 创建一个 client 对象，并且把 client 的所有请求有 WorkerProcess 进行处理。WorkerProcess 是一个 Runnable。最终提交到 executorService_ 中。
while(!this.stopped_) {try {TTransport client = this.serverTransport_.accept();WorkerProcess wp = new WorkerProcess(client);while(true) {try {this.executorService_.execute(wp);break;} catch (Throwable var13) {// omit}}} catch (TTransportException var14) {// }}
- WorkerProcess
WorkerProcess 也是除了设置停止标志外死循环。
do {if (eventHandler != null) {eventHandler.processContext(connectionContext, inputTransport, outputTransport);}} while(!TThreadPoolServer.this.stopped_ && processor.process(inputProtocol, outputProtocol));
processor 类型是 TUGIBasedProcessor。
- 当客户端正常退出时。
client 会调用 metastore 的 shutdown 方法。此方法里，清除所有的 threadlocal 对象。
public void shutdown() {cleanupRawStore();PerfLogger.getPerfLogger(false).cleanupPerfLogMetrics();}protected static void cleanupRawStore() {try {RawStore rs = HMSHandler.getRawStore();if (rs != null) {HMSHandler.logInfo("Cleaning up thread local RawStore...");rs.shutdown();}} finally {HMSHandler handler = HMSHandler.threadLocalHMSHandler.get();if (handler != null) {handler.notifyMetaListenersOnShutDown();}HMSHandler.threadLocalHMSHandler.remove();HMSHandler.threadLocalConf.remove();HMSHandler.threadLocalModifiedConfig.remove();HMSHandler.removeRawStore();HMSHandler.logInfo("Done cleaning up thread local RawStore");}}
- 异常退出时
WorkerProcess 的 finally 处理不论是否当前连接调用 shutdown，都执行eventHandler.deleteContext
finally {if (eventHandler != null) {eventHandler.deleteContext(connectionContext, inputProtocol, outputProtocol);
}
在 HiveMetaStore.java里定义了 eventHandler, 也调用了 cleanupRawStore，和 shutdown 方法里调用的一样。
@Override
public void deleteContext(ServerContext serverContext, TProtocol tProtocol, TProtocol tProtocol1) {openConnections.decrementAndGet();// If the IMetaStoreClient#close was called, HMSHandler#shutdown would have already// cleaned up thread local RawStore. Otherwise, do it now.HMSHandler.cleanupRawStore();
}
threadLocal 对象的 remove 方法多次调用是没有副作用的。
其他考虑的点
在异常退出时，没有调用 PerfLogger.getPerfLogger(false).cleanupPerfLogMetrics();, 是否会内存溢出。
调用 PerfLogger.getPerfLogger(false). 当参数是 false 时，如果 ThreadLocal 里已经有，则不会创建对象。处理线程的个数是固定的。不会导致内存问题