从上一篇对Hive metastore表结构的简要分析中,我再根据数据设计的实体对象,再进行整个代码结构的总结。那么我们先打开metadata的目录,其目录结构:
可以看到,整个hivemeta的目录包含metastore(客户端与服务端调用逻辑)、events(事件目录包含table生命周期中的检查、权限认证等listener实现)、hooks(这里的hooks仅包含了jdo connection的相关接口)、parser(对于表达树的解析)、spec(partition的相关代理类)、tools(jdo execute相关方法)及txn及model,下来我们从整个metadata分逐一进行代码分析及注释:
没有把包打开,很多类?是不是感觉害怕很想死?我也想死,咱们继续。。一开始,我们可能觉得一团乱麻烦躁,这是啥玩意儿啊这。。冷静下来,我们从Hive这个大类开始看,因为它是metastore元数据调用的入口。整个生命周期分析流程为: HiveMetaStoreClient客户端的创建及加载、HiveMetaStore服务端的创建及加载、createTable、dropTable、AlterTable、createPartition、dropPartition、alterPartition。当然,这只是完整metadata的一小部分。
1、HiveMetaStoreClient客户端的创建及加载
那么我们从Hive这个类一点点开始看:
1 private HiveConf conf = null; 2 private IMetaStoreClient metaStoreClient; 3 private UserGroupInformation owner; 4 5 // metastore calls timing information 6 private final MapmetaCallTimeMap = new HashMap (); 7 8 private static ThreadLocal hiveDB = new ThreadLocal () { 9 @Override 10 protected synchronized Hive initialValue() { 11 return null; 12 } 13 14 @Override 15 public synchronized void remove() { 16 if (this.get() != null) { 17 this.get().close(); 18 } 19 super.remove(); 20 } 21 };
这里声明的有hiveConf对象、metaStoreClient 、操作用户组userGroupInfomation以及调用时间Map,这里存成一个map,用来记录每一个动作的运行时长。同时维护了一个本地线程hiveDB,如果db为空的情况下,会重新创建一个Hive对象,代码如下:
1 public static Hive get(HiveConf c, boolean needsRefresh) throws HiveException { 2 Hive db = hiveDB.get(); 3 if (db == null || needsRefresh || !db.isCurrentUserOwner()) { 4 if (db != null) { 5 LOG.debug("Creating new db. db = " + db + ", needsRefresh = " + needsRefresh + 6 ", db.isCurrentUserOwner = " + db.isCurrentUserOwner()); 7 } 8 closeCurrent(); 9 c.set("fs.scheme.class", "dfs"); 10 Hive newdb = new Hive(c); 11 hiveDB.set(newdb); 12 return newdb; 13 } 14 db.conf = c; 15 return db; 16 }
随后我们会发现,在创建Hive对象时,便已经将function进行注册,什么是function呢,通过上次的表结构分析,可以理解为所有udf等jar包的元数据存储。代码如下:
1 // register all permanent functions. need improvement 2 static { 3 try { 4 reloadFunctions(); 5 } catch (Exception e) { 6 LOG.warn("Failed to access metastore. This class should not accessed in runtime.",e); 7 } 8 } 9 10 public static void reloadFunctions() throws HiveException {
//获取 Hive对象,用于后续方法的调用 11 Hive db = Hive.get();
//通过遍历每一个dbName 12 for (String dbName : db.getAllDatabases()) {
//通过dbName查询挂在该db下的所有function的信息。 13 for (String functionName : db.getFunctions(dbName, "*")) { 14 Function function = db.getFunction(dbName, functionName); 15 try {
//这里的register便是将查询到的function的数据注册到Registry类中的一个Map中,以便计算引擎在调用时,不必再次查询数据库。 16 FunctionRegistry.registerPermanentFunction( 17 FunctionUtils.qualifyFunctionName(functionName, dbName), function.getClassName(), 18 false, FunctionTask.toFunctionResource(function.getResourceUris())); 19 } catch (Exception e) { 20 LOG.warn("Failed to register persistent function " + 21 functionName + ":" + function.getClassName() + ". Ignore and continue."); 22 } 23 } 24 } 25 }
调用getMSC()方法,进行metadataClient客户端的创建,代码如下:
1 1 private IMetaStoreClient createMetaStoreClient() throws MetaException { 2 2 3 //这里实现接口HiveMetaHookLoader 4 3 HiveMetaHookLoader hookLoader = new HiveMetaHookLoader() { 5 4 @Override 6 5 public HiveMetaHook getHook( 7 6 org.apache.hadoop.hive.metastore.api.Table tbl) 8 7 throws MetaException { 9 8 10 9 try { 11 10 if (tbl == null) { 12 11 return null; 13 12 } 14 //根据tble的kv属性加载不同storage的实例,比如hbase、redis等等拓展存储,作为外部表进行存储 15 13 HiveStorageHandler storageHandler = 16 14 HiveUtils.getStorageHandler(conf, 17 15 tbl.getParameters().get(META_TABLE_STORAGE)); 18 16 if (storageHandler == null) { 19 17 return null; 20 18 } 21 19 return storageHandler.getMetaHook(); 22 20 } catch (HiveException ex) { 23 21 LOG.error(StringUtils.stringifyException(ex)); 24 22 throw new MetaException( 25 23 "Failed to load storage handler: " + ex.getMessage()); 26 24 } 27 25 } 28 26 }; 29 27 return RetryingMetaStoreClient.getProxy(conf, hookLoader, metaCallTimeMap, 30 28 SessionHiveMetaStoreClient.class.getName()); 31 29 }
2、HiveMetaStore服务端的创建及加载
在HiveMetaStoreClient初始化时,会初始化HiveMetaStore客户端,代码如下:
1 public HiveMetaStoreClient(HiveConf conf, HiveMetaHookLoader hookLoader) 2 throws MetaException { 3 4 this.hookLoader = hookLoader; 5 if (conf == null) { 6 conf = new HiveConf(HiveMetaStoreClient.class); 7 } 8 this.conf = conf; 9 filterHook = loadFilterHooks(); 10 //根据hive-site.xml中的hive.metastore.uris配置,如果配置该参数,则认为是远程连接,否则为本地连接 11 String msUri = conf.getVar(HiveConf.ConfVars.METASTOREURIS); 12 localMetaStore = HiveConfUtil.isEmbeddedMetaStore(msUri); 13 if (localMetaStore) { //本地连接直接连接HiveMetaStore 16 client = HiveMetaStore.newRetryingHMSHandler("hive client", conf, true); 17 isConnected = true; 18 snapshotActiveConf(); 19 return; 20 } 21 22 //获取配置中的重试次数及timeout时间 23 retries = HiveConf.getIntVar(conf, HiveConf.ConfVars.METASTORETHRIFTCONNECTIONRETRIES); 24 retryDelaySeconds = conf.getTimeVar( 25 ConfVars.METASTORE_CLIENT_CONNECT_RETRY_DELAY, TimeUnit.SECONDS); 26 27 //拼接metastore uri 28 if (conf.getVar(HiveConf.ConfVars.METASTOREURIS) != null) { 29 String metastoreUrisString[] = conf.getVar( 30 HiveConf.ConfVars.METASTOREURIS).split(","); 31 metastoreUris = new URI[metastoreUrisString.length]; 32 try { 33 int i = 0; 34 for (String s : metastoreUrisString) { 35 URI tmpUri = new URI(s); 36 if (tmpUri.getScheme() == null) { 37 throw new IllegalArgumentException("URI: " + s 38 + " does not have a scheme"); 39 } 40 metastoreUris[i++] = tmpUri; 41 42 } 43 } catch (IllegalArgumentException e) { 44 throw (e); 45 } catch (Exception e) { 46 MetaStoreUtils.logAndThrowMetaException(e); 47 } 48 } else { 49 LOG.error("NOT getting uris from conf"); 50 throw new MetaException("MetaStoreURIs not found in conf file"); 51 } 52 调用open方法创建连接 53 open(); 54 }
从上面代码中可以看出,如果我们是远程连接,需要配置hive-site.xml中的hive.metastore.uri,是不是很熟悉?加入你的client与server不在同一台机器,就需要配置进行远程连接。那么我们继续往下面看,创建连接的open方法:
1 private void open() throws MetaException { 2 isConnected = false; 3 TTransportException tte = null;
//是否使用Sasl 4 boolean useSasl = conf.getBoolVar(ConfVars.METASTORE_USE_THRIFT_SASL);
//If true, the metastore Thrift interface will use TFramedTransport. When false (default) a standard TTransport is used. 5 boolean useFramedTransport = conf.getBoolVar(ConfVars.METASTORE_USE_THRIFT_FRAMED_TRANSPORT);
//If true, the metastore Thrift interface will use TCompactProtocol. When false (default) TBinaryProtocol will be used 具体他们之间的区别我们后续再讨论 6 boolean useCompactProtocol = conf.getBoolVar(ConfVars.METASTORE_USE_THRIFT_COMPACT_PROTOCOL);
//获取socket timeout时间 7 int clientSocketTimeout = (int) conf.getTimeVar( 8 ConfVars.METASTORE_CLIENT_SOCKET_TIMEOUT, TimeUnit.MILLISECONDS); 9 10 for (int attempt = 0; !isConnected && attempt < retries; ++attempt) { 11 for (URI store : metastoreUris) { 12 LOG.info("Trying to connect to metastore with URI " + store); 13 try { 14 transport = new TSocket(store.getHost(), store.getPort(), clientSocketTimeout); 15 if (useSasl) { 16 // Wrap thrift connection with SASL for secure connection. 17 try {
//创建HadoopThriftAuthBridge client 18 HadoopThriftAuthBridge.Client authBridge = 19 ShimLoader.getHadoopThriftAuthBridge().createClient(); 20 //权限认证相关 21 // check if we should use delegation tokens to authenticate 22 // the call below gets hold of the tokens if they are set up by hadoop 23 // this should happen on the map/reduce tasks if the client added the 24 // tokens into hadoop's credential store in the front end during job 25 // submission. 26 String tokenSig = conf.get("hive.metastore.token.signature"); 27 // tokenSig could be null 28 tokenStrForm = Utils.getTokenStrForm(tokenSig); 29 if(tokenStrForm != null) { 30 // authenticate using delegation tokens via the "DIGEST" mechanism 31 transport = authBridge.createClientTransport(null, store.getHost(), 32 "DIGEST", tokenStrForm, transport, 33 MetaStoreUtils.getMetaStoreSaslProperties(conf)); 34 } else { 35 String principalConfig = 36 conf.getVar(HiveConf.ConfVars.METASTORE_KERBEROS_PRINCIPAL); 37 transport = authBridge.createClientTransport( 38 principalConfig, store.getHost(), "KERBEROS", null, 39 transport, MetaStoreUtils.getMetaStoreSaslProperties(conf)); 40 } 41 } catch (IOException ioe) { 42 LOG.error("Couldn't create client transport", ioe); 43 throw new MetaException(ioe.toString()); 44 } 45 } else if (useFramedTransport) { 46 transport = new TFramedTransport(transport); 47 } 48 final TProtocol protocol;
//后续详细说明两者的区别(因为俺还没看,哈哈) 49 if (useCompactProtocol) { 50 protocol = new TCompactProtocol(transport); 51 } else { 52 protocol = new TBinaryProtocol(transport); 53 }
//创建ThriftHiveMetastore client 54 client = new ThriftHiveMetastore.Client(protocol); 55 try { 56 transport.open(); 57 isConnected = true; 58 } catch (TTransportException e) { 59 tte = e; 60 if (LOG.isDebugEnabled()) { 61 LOG.warn("Failed to connect to the MetaStore Server...", e); 62 } else { 63 // Don't print full exception trace if DEBUG is not on. 64 LOG.warn("Failed to connect to the MetaStore Server..."); 65 } 66 } 67 //用户组及用户的加载 68 if (isConnected && !useSasl && conf.getBoolVar(ConfVars.METASTORE_EXECUTE_SET_UGI)){ 69 // Call set_ugi, only in unsecure mode. 70 try { 71 UserGroupInformation ugi = Utils.getUGI(); 72 client.set_ugi(ugi.getUserName(), Arrays.asList(ugi.getGroupNames())); 73 } catch (LoginException e) { 74 LOG.warn("Failed to do login. set_ugi() is not successful, " + 75 "Continuing without it.", e); 76 } catch (IOException e) { 77 LOG.warn("Failed to find ugi of client set_ugi() is not successful, " + 78 "Continuing without it.", e); 79 } catch (TException e) { 80 LOG.warn("set_ugi() not successful, Likely cause: new client talking to old server. " 81 + "Continuing without it.", e); 82 } 83 } 84 } catch (MetaException e) { 85 LOG.error("Unable to connect to metastore with URI " + store 86 + " in attempt " + attempt, e); 87 } 88 if (isConnected) { 89 break; 90 } 91 } 92 // Wait before launching the next round of connection retries. 93 if (!isConnected && retryDelaySeconds > 0) { 94 try { 95 LOG.info("Waiting " + retryDelaySeconds + " seconds before next connection attempt."); 96 Thread.sleep(retryDelaySeconds * 1000); 97 } catch (InterruptedException ignore) {} 98 } 99 } 100 101 if (!isConnected) { 102 throw new MetaException("Could not connect to meta store using any of the URIs provided." + 103 " Most recent failure: " + StringUtils.stringifyException(tte)); 104 } 105 106 snapshotActiveConf(); 107 108 LOG.info("Connected to metastore."); 109 }
本篇先对对protocol的原理放置一边。从代码中可以看出HiveMetaStore服务端是通过ThriftHiveMetaStore创建,它本是一个class类,但其中定义了接口Iface、AsyncIface,这样做的好处是利于继承实现。那么下来,我们看一下HMSHandler的初始化。如果是在本地调用的过程中,直接调用newRetryingHMSHandler,便会直接进行HMSHandler的初始化。代码如下:
1 public HMSHandler(String name, HiveConf conf, boolean init) throws MetaException { 2 super(name); 3 hiveConf = conf; 4 if (init) { 5 init(); 6 } 7 }
下俩我们继续看它的init方法:
1 public void init() throws MetaException {
//获取与数据交互的实现类className,该类为objectStore,是RawStore的实现,负责JDO与数据库的交互。 2 rawStoreClassName = hiveConf.getVar(HiveConf.ConfVars.METASTORE_RAW_STORE_IMPL);
//加载Listeners,来自hive.metastore.init.hooks,可自行实现并加载 3 initListeners = MetaStoreUtils.getMetaStoreListeners( 4 MetaStoreInitListener.class, hiveConf, 5 hiveConf.getVar(HiveConf.ConfVars.METASTORE_INIT_HOOKS)); 6 for (MetaStoreInitListener singleInitListener: initListeners) { 7 MetaStoreInitContext context = new MetaStoreInitContext(); 8 singleInitListener.onInit(context); 9 } 10 //初始化alter的实现类 11 String alterHandlerName = hiveConf.get("hive.metastore.alter.impl", 12 HiveAlterHandler.class.getName()); 13 alterHandler = (AlterHandler) ReflectionUtils.newInstance(MetaStoreUtils.getClass( 14 alterHandlerName), hiveConf);
//初始化warehouse 15 wh = new Warehouse(hiveConf); 16 //创建默认db以及用户,同时加载currentUrl 17 synchronized (HMSHandler.class) { 18 if (currentUrl == null || !currentUrl.equals(MetaStoreInit.getConnectionURL(hiveConf))) { 19 createDefaultDB(); 20 createDefaultRoles(); 21 addAdminUsers(); 22 currentUrl = MetaStoreInit.getConnectionURL(hiveConf); 23 } 24 } 25 //计数信息的初始化 26 if (hiveConf.getBoolean("hive.metastore.metrics.enabled", false)) { 27 try { 28 Metrics.init(); 29 } catch (Exception e) { 30 // log exception, but ignore inability to start 31 LOG.error("error in Metrics init: " + e.getClass().getName() + " " 32 + e.getMessage(), e); 33 } 34 } 35 //Listener的PreListener的初始化 36 preListeners = MetaStoreUtils.getMetaStoreListeners(MetaStorePreEventListener.class, 37 hiveConf, 38 hiveConf.getVar(HiveConf.ConfVars.METASTORE_PRE_EVENT_LISTENERS)); 39 listeners = MetaStoreUtils.getMetaStoreListeners(MetaStoreEventListener.class, hiveConf, 40 hiveConf.getVar(HiveConf.ConfVars.METASTORE_EVENT_LISTENERS)); 41 listeners.add(new SessionPropertiesListener(hiveConf)); 42 endFunctionListeners = MetaStoreUtils.getMetaStoreListeners( 43 MetaStoreEndFunctionListener.class, hiveConf, 44 hiveConf.getVar(HiveConf.ConfVars.METASTORE_END_FUNCTION_LISTENERS)); 45 //针对partitionName的正则校验,可自行设置,根据hive.metastore.partition.name.whitelist.pattern进行设置 46 String partitionValidationRegex = 47 hiveConf.getVar(HiveConf.ConfVars.METASTORE_PARTITION_NAME_WHITELIST_PATTERN); 48 if (partitionValidationRegex != null && !partitionValidationRegex.isEmpty()) { 49 partitionValidationPattern = Pattern.compile(partitionValidationRegex); 50 } else { 51 partitionValidationPattern = null; 52 } 53 54 long cleanFreq = hiveConf.getTimeVar(ConfVars.METASTORE_EVENT_CLEAN_FREQ, TimeUnit.MILLISECONDS); 55 if (cleanFreq > 0) { 56 // In default config, there is no timer. 57 Timer cleaner = new Timer("Metastore Events Cleaner Thread", true); 58 cleaner.schedule(new EventCleanerTask(this), cleanFreq, cleanFreq); 59 } 60 }
它初始化了与数据库交互的rawStore的实现类、物理操作的warehouse以及Event与Listener。从而通过接口调用相关meta生命周期方法进行表的操作。
3、createTable
从createTable方法开始。上代码:
1 public void createTable(String tableName, Listcolumns, List partCols, 2 Class extends InputFormat> fileInputFormat, 3 Class> fileOutputFormat, int bucketCount, List bucketCols, 4 Map parameters) throws HiveException { 5 if (columns == null) { 6 throw new HiveException("columns not specified for table " + tableName); 7 } 8 9 Table tbl = newTable(tableName);
//SD表属性,设置该表的input及output class名,在计算引擎计算时,拉取相应的ClassName 通过反射进行input及output类的加载 10 tbl.setInputFormatClass(fileInputFormat.getName()); 11 tbl.setOutputFormatClass(fileOutputFormat.getName()); 12
//封装FileSchema对象,该为每个column的名称及字段类型,并加入到sd对象的的column属性中 13 for (String col : columns) { 14 FieldSchema field = new FieldSchema(col, STRING_TYPE_NAME, "default"); 15 tbl.getCols().add(field); 16 } 17
//如果在创建表时,设置了分区信息,比如dt字段为该分区。则进行分区信息的记录,最终写入Partition表中 18 if (partCols != null) { 19 for (String partCol : partCols) { 20 FieldSchema part = new FieldSchema(); 21 part.setName(partCol); 22 part.setType(STRING_TYPE_NAME); // default partition key 23 tbl.getPartCols().add(part); 24 } 25 }
//设置序列化的方式 26 tbl.setSerializationLib(LazySimpleSerDe.class.getName());
//设置分桶信息 27 tbl.setNumBuckets(bucketCount); 28 tbl.setBucketCols(bucketCols);
//设置table额外添加的kv信息 29 if (parameters != null) { 30 tbl.setParamters(parameters); 31 } 32 createTable(tbl); 33 }
从代码中可以看到,Hive 构造了一个Table的对象,该对象可以当做是一个model,包含了几乎所有以Tbls表为主表的所有以table_id为的外键表属性(具体可参考hive metastore表结构),封装完毕后在进行createTable的调用,接下来的调用如下:
public void createTable(Table tbl, boolean ifNotExists) throws HiveException { try {
//这里再次获取SessionState中的CurrentDataBase进行setDbName(安全) if (tbl.getDbName() == null || "".equals(tbl.getDbName().trim())) { tbl.setDbName(SessionState.get().getCurrentDatabase()); }
//这里主要对每一个column属性进行校验,比如是否有非法字符等等 if (tbl.getCols().size() == 0 || tbl.getSd().getColsSize() == 0) { tbl.setFields(MetaStoreUtils.getFieldsFromDeserializer(tbl.getTableName(), tbl.getDeserializer())); }
//该方法对table属性中的input、output以及column属性的校验 tbl.checkValidity(); if (tbl.getParameters() != null) { tbl.getParameters().remove(hive_metastoreConstants.DDL_TIME); } org.apache.hadoop.hive.metastore.api.Table tTbl = tbl.getTTable();
//这里开始进行权限认证,牵扯到的便是我们再hive中配置的 hive.security.authorization.createtable.user.grants、hive.security.authorization.createtable.group.grants、
hive.security.authorization.createtable.role.grants配置参数,来自于hive自己封装的 用户、角色、组的概念。
PrincipalPrivilegeSet principalPrivs = new PrincipalPrivilegeSet(); SessionState ss = SessionState.get(); if (ss != null) { CreateTableAutomaticGrant grants = ss.getCreateTableGrants(); if (grants != null) { principalPrivs.setUserPrivileges(grants.getUserGrants()); principalPrivs.setGroupPrivileges(grants.getGroupGrants()); principalPrivs.setRolePrivileges(grants.getRoleGrants()); tTbl.setPrivileges(principalPrivs); } }
//通过客户端链接服务端进行table的创建 getMSC().createTable(tTbl); } catch (AlreadyExistsException e) { if (!ifNotExists) { throw new HiveException(e); } } catch (Exception e) { throw new HiveException(e); } }
那么下来,我们来看一下受到调用的HiveMetaClient中createTable方法,代码如下:
1 public void createTable(Table tbl, EnvironmentContext envContext) throws AlreadyExistsException, 2 InvalidObjectException, MetaException, NoSuchObjectException, TException {
//这里获取HiveMeetaHook对象,针对不同的存储引擎进行创建前的加载及验证 3 HiveMetaHook hook = getHook(tbl); 4 if (hook != null) { 5 hook.preCreateTable(tbl); 6 } 7 boolean success = false; 8 try { //随即调用HiveMetaStore进行服务端与数据库的创建交互 10 create_table_with_environment_context(tbl, envContext); 11 if (hook != null) { 12 hook.commitCreateTable(tbl); 13 } 14 success = true; 15 } finally {
如果创建失败的话,进行回滚操作 16 if (!success && (hook != null)) { 17 hook.rollbackCreateTable(tbl); 18 } 19 } 20 }
这里简要说下Hook的作用,HiveMetaHook为接口,接口方法包括preCreate、rollbackCreateTable、preDropTable等等操作,它的实现为不同存储类型的预创建加载及验证,以及失败回滚等动作。代码如下:
1 public interface HiveMetaHook { 2 /** 3 * Called before a new table definition is added to the metastore 4 * during CREATE TABLE. 5 * 6 * @param table new table definition 7 */ 8 public void preCreateTable(Table table) 9 throws MetaException; 10 11 /** 12 * Called after failure adding a new table definition to the metastore 13 * during CREATE TABLE. 14 * 15 * @param table new table definition 16 */ 17 public void rollbackCreateTable(Table table) 18 throws MetaException; 35 public void preDropTale(Table table) 36 throws MetaException;
...............................
随后,我们再看一下HiveMetaStore服务端的createTable方法,如下:
1 private void create_table_core(final RawStore ms, final Table tbl,
2 final EnvironmentContext envContext) 3 throws AlreadyExistsException, MetaException, 4 InvalidObjectException, NoSuchObjectException { 5 //名称正则校验,校验是否含有非法字符 6 if (!MetaStoreUtils.validateName(tbl.getTableName())) { 7 throw new InvalidObjectException(tbl.getTableName() 8 + " is not a valid object name"); 9 }
//改端代码属于校验代码,对于column的名称及column type类型j及partitionKey的名称校验 10 String validate = MetaStoreUtils.validateTblColumns(tbl.getSd().getCols()); 11 if (validate != null) { 12 throw new InvalidObjectException("Invalid column " + validate); 13 }
14 if (tbl.getPartitionKeys() != null) { 15 validate = MetaStoreUtils.validateTblColumns(tbl.getPartitionKeys()); 16 if (validate != null) { 17 throw new InvalidObjectException("Invalid partition column " + validate); 18 } 19 } 20 SkewedInfo skew = tbl.getSd().getSkewedInfo(); 21 if (skew != null) { 22 validate = MetaStoreUtils.validateSkewedColNames(skew.getSkewedColNames()); 23 if (validate != null) { 24 throw new InvalidObjectException("Invalid skew column " + validate); 25 } 26 validate = MetaStoreUtils.validateSkewedColNamesSubsetCol( 27 skew.getSkewedColNames(), tbl.getSd().getCols()); 28 if (validate != null) { 29 throw new InvalidObjectException("Invalid skew column " + validate); 30 } 31 } 32 33 Path tblPath = null; 34 boolean success = false, madeDir = false; 35 try {
//创建前的事件调用,metastore已实现的listner事件包含DummyPreListener、AuthorizationPreEventListener、AlternateFailurePreListener以及MetaDataExportListener。
//这些Listener是干嘛的呢?详细解释由分析meta设计模式时,详细说明。 36 firePreEvent(new PreCreateTableEvent(tbl, this)); 37
//打开事务 38 ms.openTransaction(); 39
//如果db不存在的情况下,则抛异常 40 Database db = ms.getDatabase(tbl.getDbName()); 41 if (db == null) { 42 throw new NoSuchObjectException("The database " + tbl.getDbName() + " does not exist"); 43 } 44 45 // 校验该db下,table是否存在 46 if (is_table_exists(ms, tbl.getDbName(), tbl.getTableName())) { 47 throw new AlreadyExistsException("Table " + tbl.getTableName() 48 + " already exists"); 49 } 50 // 如果该表不为视图表,则组装完整的tbleParth ->fs.getUri().getScheme()+fs.getUri().getAuthority()+path.toUri().getPath())
51 if (!TableType.VIRTUAL_VIEW.toString().equals(tbl.getTableType())) { 52 if (tbl.getSd().getLocation() == null 53 || tbl.getSd().getLocation().isEmpty()) { 54 tblPath = wh.getTablePath( 55 ms.getDatabase(tbl.getDbName()), tbl.getTableName()); 56 } else {
//如果该表不是内部表同时tbl的kv中storage_handler为空时,则只是警告 57 if (!isExternal(tbl) && !MetaStoreUtils.isNonNativeTable(tbl)) { 58 LOG.warn("Location: " + tbl.getSd().getLocation() 59 + " specified for non-external table:" + tbl.getTableName()); 60 } 61 tblPath = wh.getDnsPath(new Path(tbl.getSd().getLocation())); 62 }
//将拼接完的tblPath set到sd的location中 63 tbl.getSd().setLocation(tblPath.toString()); 64 } 65 //创建table的路径 66 if (tblPath != null) { 67 if (!wh.isDir(tblPath)) { 68 if (!wh.mkdirs(tblPath, true)) { 69 throw new MetaException(tblPath 70 + " is not a directory or unable to create one"); 71 } 72 madeDir = true; 73 } 74 }
// hive.stats.autogather 配置判断 75 if (HiveConf.getBoolVar(hiveConf, HiveConf.ConfVars.HIVESTATSAUTOGATHER) && 76 !MetaStoreUtils.isView(tbl)) { 77 if (tbl.getPartitionKeysSize() == 0) { // Unpartitioned table 78 MetaStoreUtils.updateUnpartitionedTableStatsFast(db, tbl, wh, madeDir); 79 } else { // Partitioned table with no partitions. 80 MetaStoreUtils.updateUnpartitionedTableStatsFast(db, tbl, wh, true); 81 } 82 } 83 84 // set create time 85 long time = System.currentTimeMillis() / 1000; 86 tbl.setCreateTime((int) time); 87 if (tbl.getParameters() == null || 88 tbl.getParameters().get(hive_metastoreConstants.DDL_TIME) == null) { 89 tbl.putToParameters(hive_metastoreConstants.DDL_TIME, Long.toString(time)); 90 }
执行createTable数据库操作 91 ms.createTable(tbl); 92 success = ms.commitTransaction(); 93 94 } finally { 95 if (!success) { 96 ms.rollbackTransaction();
//如果由于某些原因没有创建,则进行已创建表路径的删除 97 if (madeDir) { 98 wh.deleteDir(tblPath, true); 99 } 100 }
//进行create完成时的listener类发送 比如 noftify通知 101 for (MetaStoreEventListener listener : listeners) { 102 CreateTableEvent createTableEvent = 103 new CreateTableEvent(tbl, success, this); 104 createTableEvent.setEnvironmentContext(envContext); 105 listener.onCreateTable(createTableEvent); 106 } 107 } 108 }
这里的listener后续会详细说明,那么我们继续垂直往下看,这里的 ms.createTable方法。ms便是RawStore接口对象,这个接口对象包含了所有生命周期的统一方法调用,部分代码如下:
1 public abstract Database getDatabase(String name) 2 throws NoSuchObjectException; 3 4 public abstract boolean dropDatabase(String dbname) throws NoSuchObjectException, MetaException; 5 6 public abstract boolean alterDatabase(String dbname, Database db) throws NoSuchObjectException, MetaException; 7 8 public abstract ListgetDatabases(String pattern) throws MetaException; 9 10 public abstract List getAllDatabases() throws MetaException; 11 12 public abstract boolean createType(Type type); 13 14 public abstract Type getType(String typeName); 15 16 public abstract boolean dropType(String typeName); 17 18 public abstract void createTable(Table tbl) throws InvalidObjectException, 19 MetaException; 20 21 public abstract boolean dropTable(String dbName, String tableName) 22 throws MetaException, NoSuchObjectException, InvalidObjectException, InvalidInputException; 23 24 public abstract Table getTable(String dbName, String tableName) 25 throws MetaException; 26 ..................
那么下来我们来看一下具体怎么实现的,首先hive metastore会通过调用getMS()方法,获取本地线程中的RawStore的实现,代码如下:
1 public RawStore getMS() throws MetaException {
//获取本地线程中已存在的RawStore 2 RawStore ms = threadLocalMS.get();
//如果不存在,则创建该对象的实现,并加入到本地线程中 3 if (ms == null) { 4 ms = newRawStore(); 5 ms.verifySchema(); 6 threadLocalMS.set(ms); 7 ms = threadLocalMS.get(); 8 } 9 return ms; 10 }
看到这里,是不是很想看看newRawStore它干嘛啦?那么我们继续:
1 public static RawStore getProxy(HiveConf hiveConf, Configuration conf, String rawStoreClassName, 2 int id) throws MetaException { 3 //通过反射,创建baseClass,随后再进行该实现对象的创建 4 Class extends RawStore> baseClass = (Class extends RawStore>) MetaStoreUtils.getClass( 5 rawStoreClassName); 6 7 RawStoreProxy handler = new RawStoreProxy(hiveConf, conf, baseClass, id); 8 9 // Look for interfaces on both the class and all base classes. 10 return (RawStore) Proxy.newProxyInstance(RawStoreProxy.class.getClassLoader(), 11 getAllInterfaces(baseClass), handler); 12 }
那么问题来了,rawstoreClassName从哪里来呢?它是在HiveMetaStore进行初始化时加载的,来源于HiveConf中的METASTORE_RAW_STORE_IMPL,配置参数,也就是RawStore的实现类ObjectStore。好了,既然RawStore的实现类已经创建,那么我们继续深入ObjectStore,代码如下:
1 @Override 2 public void createTable(Table tbl) throws InvalidObjectException, MetaException { 3 boolean commited = false; 4 try {
//创建事务 5 openTransaction();
//这里再次进行db 、table的校验,代码不再贴出来,具体为什么又要做一次校验,还需要深入思考 6 MTable mtbl = convertToMTable(tbl);
这里的pm为ObjectStore创建时,init的JDO PersistenceManage对象。这里便是提交Table对象的地方,具体可研究下JDO module对象与数据库的交互 7 pm.makePersistent(mtbl);
//封装权限用户、角色、组对象并写入 8 PrincipalPrivilegeSet principalPrivs = tbl.getPrivileges(); 9 List
4、dropTable
二话不说上从Hive类中上代码:
1 public void dropTable(String tableName, boolean ifPurge) throws HiveException {
//这里Hive 将dbName与TableName合并成一个数组 2 String[] names = Utilities.getDbTableName(tableName); 3 dropTable(names[0], names[1], true, true, ifPurge); 4 }
为什么要进行这样的处理呢,其实是因为 drop table的时候 我们的sql语句会是drop table dbName.tableName 或者是drop table tableName,这里进行tableName和DbName的组装,如果为drop table tableName,则获取当前session中的dbName,代码如下:
1 public static String[] getDbTableName(String dbtable) throws SemanticException {
//获取当前Session中的DbName 2 return getDbTableName(SessionState.get().getCurrentDatabase(), dbtable); 3 } 4 5 public static String[] getDbTableName(String defaultDb, String dbtable) throws SemanticException { 6 if (dbtable == null) { 7 return new String[2]; 8 } 9 String[] names = dbtable.split("\\."); 10 switch (names.length) { 11 case 2: 12 return names;
//如果长度为1,则重新组装 13 case 1: 14 return new String [] {defaultDb, dbtable}; 15 default: 16 throw new SemanticException(ErrorMsg.INVALID_TABLE_NAME, dbtable); 17 } 18 }
随后通过getMSC()调用HiveMetaStoreClient中的dropTable,代码如下:
1 public void dropTable(String dbname, String name, boolean deleteData, 2 boolean ignoreUnknownTab, EnvironmentContext envContext) throws MetaException, TException, 3 NoSuchObjectException, UnsupportedOperationException { 4 Table tbl; 5 try {
//通过dbName与tableName获取正个Table对象,也就是通过dbName与TableName获取该Table存储的所有元数据 6 tbl = getTable(dbname, name); 7 } catch (NoSuchObjectException e) { 8 if (!ignoreUnknownTab) { 9 throw e; 10 } 11 return; 12 }
//根据table type来判断是否为IndexTable,如果为索引表则不允许删除 13 if (isIndexTable(tbl)) { 14 throw new UnsupportedOperationException("Cannot drop index tables"); 15 }
//这里的getHook 与create时getHook一致,获取对应table存储的hook 16 HiveMetaHook hook = getHook(tbl); 17 if (hook != null) { 18 hook.preDropTable(tbl); 19 } 20 boolean success = false; 21 try {
调用HiveMetaStore服务端的dropTable方法 22 drop_table_with_environment_context(dbname, name, deleteData, envContext); 23 if (hook != null) { 24 hook.commitDropTable(tbl, deleteData); 25 } 26 success=true; 27 } catch (NoSuchObjectException e) { 28 if (!ignoreUnknownTab) { 29 throw e; 30 } 31 } finally { 32 if (!success && (hook != null)) { 33 hook.rollbackDropTable(tbl); 34 } 35 } 36 }
下面我们重点看下服务端HiveMetaStore干了些什么,代码如下:
1 private boolean drop_table_core(final RawStore ms, final String dbname, final String name, 2 final boolean deleteData, final EnvironmentContext envContext, 3 final String indexName) throws NoSuchObjectException, 4 MetaException, IOException, InvalidObjectException, InvalidInputException { 5 boolean success = false; 6 boolean isExternal = false; 7 Path tblPath = null; 8 ListpartPaths = null; 9 Table tbl = null; 10 boolean ifPurge = false; 11 try { 12 ms.openTransaction(); 13 // 获取正个Table的对象属性 14 tbl = get_table_core(dbname, name); 15 if (tbl == null) { 16 throw new NoSuchObjectException(name + " doesn't exist"); 17 }
//如果sd数据为空,则认为该表数据损坏 18 if (tbl.getSd() == null) { 19 throw new MetaException("Table metadata is corrupted"); 20 } 21 ifPurge = isMustPurge(envContext, tbl); 22 23 firePreEvent(new PreDropTableEvent(tbl, deleteData, this)); //判断如果该表存在索引,则需要先删除该表的索引 25 boolean isIndexTable = isIndexTable(tbl); 26 if (indexName == null && isIndexTable) { 27 throw new RuntimeException( 28 "The table " + name + " is an index table. Please do drop index instead."); 29 } //如果不是索引表,则删除索引元数据 31 if (!isIndexTable) { 32 try { 33 Listindexes = ms.getIndexes(dbname, name, Short.MAX_VALUE); 34 while (indexes != null && indexes.size() > 0) { 35 for (Index idx : indexes) { 36 this.drop_index_by_name(dbname, name, idx.getIndexName(), true); 37 } 38 indexes = ms.getIndexes(dbname, name, Short.MAX_VALUE); 39 } 40 } catch (TException e) { 41 throw new MetaException(e.getMessage()); 42 } 43 }
//判断是否为外部表 44 isExternal = isExternal(tbl); 45 if (tbl.getSd().getLocation() != null) { 46 tblPath = new Path(tbl.getSd().getLocation()); 47 if (!wh.isWritable(tblPath.getParent())) { 48 String target = indexName == null ? "Table" : "Index table"; 49 throw new MetaException(target + " metadata not deleted since " + 50 tblPath.getParent() + " is not writable by " + 51 hiveConf.getUser()); 52 } 53 } 54 56 checkTrashPurgeCombination(tblPath, dbname + "." + name, ifPurge); 57 //获取所有partition的location path 这里有个奇怪的地方,为什么不将Table对象直接传入,而是又在该方法中重新getTable,同时校验上级目录的读写权限 58 partPaths = dropPartitionsAndGetLocations(ms, dbname, name, tblPath, 59 tbl.getPartitionKeys(), deleteData && !isExternal); 60 //调用ObjectStore进行meta数据的删除 61 if (!ms.dropTable(dbname, name)) { 62 String tableName = dbname + "." + name; 63 throw new MetaException(indexName == null ? "Unable to drop table " + tableName: 64 "Unable to drop index table " + tableName + " for index " + indexName); 65 } 66 success = ms.commitTransaction(); 67 } finally { 68 if (!success) { 69 ms.rollbackTransaction(); 70 } else if (deleteData && !isExternal) { //删除物理partition 73 deletePartitionData(partPaths, ifPurge); 74 //删除Table路径 75 deleteTableData(tblPath, ifPurge); 76 // ok even if the data is not deleted 77
//Listener 处理 78 for (MetaStoreEventListener listener : listeners) { 79 DropTableEvent dropTableEvent = new DropTableEvent(tbl, success, deleteData, this); 80 dropTableEvent.setEnvironmentContext(envContext); 81 listener.onDropTable(dropTableEvent); 82 } 83 } 84 return success; 85 }
我们继续深入ObjectStore中的dropTable,会发现 再一次通过dbName与tableName获取整个Table对象,随后逐一删除。也许代码并不是同一个人写的也可能是由于安全性考虑?很多可以通过接口传入的Table对象,都重新获取了,这样会不会加重数据库的负担呢?ObjectStore代码如下:
1 public boolean dropTable(String dbName, String tableName) throws MetaException, 2 NoSuchObjectException, InvalidObjectException, InvalidInputException { 3 boolean success = false; 4 try { 5 openTransaction();
//重新获取Table对象 6 MTable tbl = getMTable(dbName, tableName); 7 pm.retrieve(tbl); 8 if (tbl != null) { 9 //下列代码查询并删除所有的权限 10 ListtabGrants = listAllTableGrants(dbName, tableName); 11 if (tabGrants != null && tabGrants.size() > 0) { 12 pm.deletePersistentAll(tabGrants); 13 }
14 ListtblColGrants = listTableAllColumnGrants(dbName, 15 tableName); 16 if (tblColGrants != null && tblColGrants.size() > 0) { 17 pm.deletePersistentAll(tblColGrants); 18 } 19 20 List partGrants = this.listTableAllPartitionGrants(dbName, tableName); 21 if (partGrants != null && partGrants.size() > 0) { 22 pm.deletePersistentAll(partGrants); 23 } 24 25 List partColGrants = listTableAllPartitionColumnGrants(dbName, 26 tableName); 27 if (partColGrants != null && partColGrants.size() > 0) { 28 pm.deletePersistentAll(partColGrants); 29 } 30 // delete column statistics if present 31 try {
//删除column统计表数据 32 deleteTableColumnStatistics(dbName, tableName, null); 33 } catch (NoSuchObjectException e) { 34 LOG.info("Found no table level column statistics associated with db " + dbName + 35 " table " + tableName + " record to delete"); 36 } 37 //删除mcd表数据 38 preDropStorageDescriptor(tbl.getSd()); 39 //删除整个Table对象相关表数据 40 pm.deletePersistentAll(tbl); 41 } 42 success = commitTransaction(); 43 } finally { 44 if (!success) { 45 rollbackTransaction(); 46 } 47 } 48 return success; 49 }
5、AlterTable
下来我们看下AlterTable,AlterTable包含的逻辑较多,因为牵扯到物理存储上的路径修改等,那么我们来一点点查看。还是从Hive类中开始,上代码:
1 public void alterTable(String tblName, Table newTbl, boolean cascade) 2 throws InvalidOperationException, HiveException { 3 String[] names = Utilities.getDbTableName(tblName); 4 try { 5 //删除table kv中的DDL_TIME 因为要alterTable所以,该事件会被改变 6 if (newTbl.getParameters() != null) { 7 newTbl.getParameters().remove(hive_metastoreConstants.DDL_TIME); 8 }
//进行相关校验,包含dbName、tableName、column、inputOutClass、outputClass的校验等,如果校验不通过则抛出HiveException 9 newTbl.checkValidity();
//调用alterTable 10 getMSC().alter_table(names[0], names[1], newTbl.getTTable(), cascade); 11 } catch (MetaException e) { 12 throw new HiveException("Unable to alter table. " + e.getMessage(), e); 13 } catch (TException e) { 14 throw new HiveException("Unable to alter table. " + e.getMessage(), e); 15 } 16 }
对于HiveMetaClient,并没有做相应处理,所以我们直接来看HiveMetaStore服务端做了些什么呢?
1 private void alter_table_core(final String dbname, final String name, final Table newTable, 2 final EnvironmentContext envContext, final boolean cascade) 3 throws InvalidOperationException, MetaException { 4 startFunction("alter_table", ": db=" + dbname + " tbl=" + name 5 + " newtbl=" + newTable.getTableName()); 6 7 //更新DDL_Time 8 if (newTable.getParameters() == null || 9 newTable.getParameters().get(hive_metastoreConstants.DDL_TIME) == null) { 10 newTable.putToParameters(hive_metastoreConstants.DDL_TIME, Long.toString(System 11 .currentTimeMillis() / 1000)); 12 } 13 boolean success = false; 14 Exception ex = null; 15 try {
//获取已有Table的整个对象 16 Table oldt = get_table_core(dbname, name);
//进行Event处理 17 firePreEvent(new PreAlterTableEvent(oldt, newTable, this));
//进行alterTable处理,后面详细说明 18 alterHandler.alterTable(getMS(), wh, dbname, name, newTable, cascade); 19 success = true; 20
//进行Listener处理 21 for (MetaStoreEventListener listener : listeners) { 22 23 AlterTableEvent alterTableEvent = 24 new AlterTableEvent(oldt, newTable, success, this); 25 alterTableEvent.setEnvironmentContext(envContext); 26 listener.onAlterTable(alterTableEvent); 27 } 28 } catch (NoSuchObjectException e) { 29 // thrown when the table to be altered does not exist 30 ex = e; 31 throw new InvalidOperationException(e.getMessage()); 32 } catch (Exception e) { 33 ex = e; 34 if (e instanceof MetaException) { 35 throw (MetaException) e; 36 } else if (e instanceof InvalidOperationException) { 37 throw (InvalidOperationException) e; 38 } else { 39 throw newMetaException(e); 40 } 41 } finally { 42 endFunction("alter_table", success, ex, name); 43 } 44 }
那么,我们重点看下alterHandler具体所做的事情,在这之前简要说下alterHandler的初始化,它是在HiveMetaStore init时获取的hive.metastore.alter.impl参数的className,也就是HiveAlterHandler的name,那么具体,我们来看下它alterTable时的实现,前方高能,小心火烛:)
1 public void alterTable(RawStore msdb, Warehouse wh, String dbname, 2 String name, Table newt, boolean cascade) throws InvalidOperationException, MetaException { 3 if (newt == null) { 4 throw new InvalidOperationException("New table is invalid: " + newt); 5 } 6 //校验新的tableName是否合法 7 if (!MetaStoreUtils.validateName(newt.getTableName())) { 8 throw new InvalidOperationException(newt.getTableName() 9 + " is not a valid object name"); 10 }
//校验新的column Name type是否合法 11 String validate = MetaStoreUtils.validateTblColumns(newt.getSd().getCols()); 12 if (validate != null) { 13 throw new InvalidOperationException("Invalid column " + validate); 14 } 15 16 Path srcPath = null; 17 FileSystem srcFs = null; 18 Path destPath = null; 19 FileSystem destFs = null; 20 21 boolean success = false; 22 boolean moveData = false; 23 boolean rename = false; 24 Table oldt = null; 25 List> altps = new ArrayList >(); 26 27 try { 28 msdb.openTransaction();
//这里直接转换小写,可以看出 代码不是一个人写的 29 name = name.toLowerCase(); 30 dbname = dbname.toLowerCase(); 31 32 //校验新的tableName是否存在 33 if (!newt.getTableName().equalsIgnoreCase(name) 34 || !newt.getDbName().equalsIgnoreCase(dbname)) { 35 if (msdb.getTable(newt.getDbName(), newt.getTableName()) != null) { 36 throw new InvalidOperationException("new table " + newt.getDbName() 37 + "." + newt.getTableName() + " already exists"); 38 } 39 rename = true; 40 } 41 42 //获取老的table对象 43 oldt = msdb.getTable(dbname, name); 44 if (oldt == null) { 45 throw new InvalidOperationException("table " + newt.getDbName() + "." 46 + newt.getTableName() + " doesn't exist"); 47 } 48 //alter Table时 获取 METASTORE_DISALLOW_INCOMPATIBLE_COL_TYPE_CHANGES配置项,如果为true的话,将改变column的type类型,这里为false 49 if (HiveConf.getBoolVar(hiveConf, 50 HiveConf.ConfVars.METASTORE_DISALLOW_INCOMPATIBLE_COL_TYPE_CHANGES, 51 false)) { 52 // Throws InvalidOperationException if the new column types are not 53 // compatible with the current column types. 54 MetaStoreUtils.throwExceptionIfIncompatibleColTypeChange( 55 oldt.getSd().getCols(), newt.getSd().getCols()); 56 } 57 //cascade参数由调用Hive altertable方法穿过来的,也就是引擎调用时参数的设置,这里用来查看是否需要alterPartition信息 58 if (cascade) { 59 //校验新的column是否与老的column一致,如不一致,说明进行了column的添加或删除操作 60 if(MetaStoreUtils.isCascadeNeededInAlterTable(oldt, newt)) {
//根据dbName与tableName获取整个partition的信息 61 Listparts = msdb.getPartitions(dbname, name, -1); 62 for (Partition part : parts) { 63 List oldCols = part.getSd().getCols(); 64 part.getSd().setCols(newt.getSd().getCols()); 65 String oldPartName = Warehouse.makePartName(oldt.getPartitionKeys(), part.getValues());
//如果columns不一致,则删除已有的column统计信息 66 updatePartColumnStatsForAlterColumns(msdb, part, oldPartName, part.getValues(), oldCols, part);
//更新整个Partition的信息 67 msdb.alterPartition(dbname, name, part.getValues(), part); 68 } 69 } else { 70 LOG.warn("Alter table does not cascade changes to its partitions."); 71 } 72 } 73 74 //判断parititonkey是否改变,也就是dt 或 hour等partName是否改变 76 boolean partKeysPartiallyEqual = checkPartialPartKeysEqual(oldt.getPartitionKeys(), 77 newt.getPartitionKeys()); 78
//如果已有表为视图表,同时发现老的partkey与新的partKey不一致,则报错 79 if(!oldt.getTableType().equals(TableType.VIRTUAL_VIEW.toString())){ 80 if (oldt.getPartitionKeys().size() != newt.getPartitionKeys().size() 81 || !partKeysPartiallyEqual) { 82 throw new InvalidOperationException( 83 "partition keys can not be changed."); 84 } 85 } 86 //如果该表不为视图表,同时,该表的location信息并未发生变化,同时新的location信息并不为空,并且已有的该表不为外部表,说明用户是想要移动数据到新的location地址,那么该操作
// 为alter table rename操作 91 if (rename 92 && !oldt.getTableType().equals(TableType.VIRTUAL_VIEW.toString()) 93 && (oldt.getSd().getLocation().compareTo(newt.getSd().getLocation()) == 0 94 || StringUtils.isEmpty(newt.getSd().getLocation())) 95 && !MetaStoreUtils.isExternalTable(oldt)) { 96 //获取新的location信息 97 srcPath = new Path(oldt.getSd().getLocation()); 98 srcFs = wh.getFs(srcPath); 99 100 // that means user is asking metastore to move data to new location 101 // corresponding to the new name 102 // get new location 103 Database db = msdb.getDatabase(newt.getDbName()); 104 Path databasePath = constructRenamedPath(wh.getDatabasePath(db), srcPath); 105 destPath = new Path(databasePath, newt.getTableName()); 106 destFs = wh.getFs(destPath); 107 //设置新的table location信息 用于后续更新动作 108 newt.getSd().setLocation(destPath.toString()); 109 moveData = true; 110 //校验物理目标地址是否存在,如果存在则会override所有数据,是不允许的。 114 if (!FileUtils.equalsFileSystem(srcFs, destFs)) { 115 throw new InvalidOperationException("table new location " + destPath 116 + " is on a different file system than the old location " 117 + srcPath + ". This operation is not supported"); 118 } 119 try { 120 srcFs.exists(srcPath); // check that src exists and also checks 121 // permissions necessary 122 if (destFs.exists(destPath)) { 123 throw new InvalidOperationException("New location for this table " 124 + newt.getDbName() + "." + newt.getTableName() 125 + " already exists : " + destPath); 126 } 127 } catch (IOException e) { 128 throw new InvalidOperationException("Unable to access new location " 129 + destPath + " for table " + newt.getDbName() + "." 130 + newt.getTableName()); 131 } 132 String oldTblLocPath = srcPath.toUri().getPath(); 133 String newTblLocPath = destPath.toUri().getPath(); 134 135 //获取old table中的所有partition信息 136 Listparts = msdb.getPartitions(dbname, name, -1); 137 for (Partition part : parts) { 138 String oldPartLoc = part.getSd().getLocation();
//这里,便开始新老partition地址的变换,修改partition元数据信息 139 if (oldPartLoc.contains(oldTblLocPath)) { 140 URI oldUri = new Path(oldPartLoc).toUri(); 141 String newPath = oldUri.getPath().replace(oldTblLocPath, newTblLocPath); 142 Path newPartLocPath = new Path(oldUri.getScheme(), oldUri.getAuthority(), newPath); 143 altps.add(ObjectPair.create(part, part.getSd().getLocation())); 144 part.getSd().setLocation(newPartLocPath.toString()); 145 String oldPartName = Warehouse.makePartName(oldt.getPartitionKeys(), part.getValues()); 146 try { 147 //existing partition column stats is no longer valid, remove them 148 msdb.deletePartitionColumnStatistics(dbname, name, oldPartName, part.getValues(), null); 149 } catch (InvalidInputException iie) { 150 throw new InvalidOperationException("Unable to update partition stats in table rename." + iie); 151 } 152 msdb.alterPartition(dbname, name, part.getValues(), part); 153 } 154 }
//更新stats相关信息 155 } else if (MetaStoreUtils.requireCalStats(hiveConf, null, null, newt) && 156 (newt.getPartitionKeysSize() == 0)) { 157 Database db = msdb.getDatabase(newt.getDbName()); 158 // Update table stats. For partitioned table, we update stats in 159 // alterPartition() 160 MetaStoreUtils.updateUnpartitionedTableStatsFast(db, newt, wh, false, true); 161 } 162 updateTableColumnStatsForAlterTable(msdb, oldt, newt); 163 // now finally call alter table 164 msdb.alterTable(dbname, name, newt); 165 // commit the changes 166 success = msdb.commitTransaction(); 167 } catch (InvalidObjectException e) { 168 LOG.debug(e); 169 throw new InvalidOperationException( 170 "Unable to change partition or table." 171 + " Check metastore logs for detailed stack." + e.getMessage()); 172 } catch (NoSuchObjectException e) { 173 LOG.debug(e); 174 throw new InvalidOperationException( 175 "Unable to change partition or table. Database " + dbname + " does not exist" 176 + " Check metastore logs for detailed stack." + e.getMessage()); 177 } finally { 178 if (!success) { 179 msdb.rollbackTransaction(); 180 } 181 if (success && moveData) { //开始更新hdfs路径,进行老路径的rename到新路径 ,调用fileSystem的rename操作 185 try { 186 if (srcFs.exists(srcPath) && !srcFs.rename(srcPath, destPath)) { 187 throw new IOException("Renaming " + srcPath + " to " + destPath + " failed"); 188 } 189 } catch (IOException e) { 190 LOG.error("Alter Table operation for " + dbname + "." + name + " failed.", e); 191 boolean revertMetaDataTransaction = false; 192 try { 193 msdb.openTransaction();
//这里会发现,又一次进行了alterTable元数据动作,或许跟JDO的特性有关?还是因为安全? 194 msdb.alterTable(newt.getDbName(), newt.getTableName(), oldt); 195 for (ObjectPairpair : altps) { 196 Partition part = pair.getFirst(); 197 part.getSd().setLocation(pair.getSecond()); 198 msdb.alterPartition(newt.getDbName(), name, part.getValues(), part); 199 } 200 revertMetaDataTransaction = msdb.commitTransaction(); 201 } catch (Exception e1) { 202 // we should log this for manual rollback by administrator 203 LOG.error("Reverting metadata by HDFS operation failure failed During HDFS operation failed", e1); 204 LOG.error("Table " + Warehouse.getQualifiedName(newt) + 205 " should be renamed to " + Warehouse.getQualifiedName(oldt)); 206 LOG.error("Table " + Warehouse.getQualifiedName(newt) + 207 " should have path " + srcPath); 208 for (ObjectPair pair : altps) { 209 LOG.error("Partition " + Warehouse.getQualifiedName(pair.getFirst()) + 210 " should have path " + pair.getSecond()); 211 } 212 if (!revertMetaDataTransaction) { 213 msdb.rollbackTransaction(); 214 } 215 } 216 throw new InvalidOperationException("Alter Table operation for " + dbname + "." + name + 217 " failed to move data due to: '" + getSimpleMessage(e) + "' See hive log file for details."); 218 } 219 } 220 } 221 if (!success) { 222 throw new MetaException("Committing the alter table transaction was not successful."); 223 } 224 }
6、createPartition
在分区数据写入之前,会先进行partition的元数据注册及物理文件路径的创建(内部表),Hive类代码如下:
1 public Partition createPartition(Table tbl, MappartSpec) throws HiveException { 2 try {
//new出来一个Partition对象,传入Table对象,调用Partition的构造方法来initialize Partition的信息 3 return new Partition(tbl, getMSC().add_partition( 4 Partition.createMetaPartitionObject(tbl, partSpec, null))); 5 } catch (Exception e) { 6 LOG.error(StringUtils.stringifyException(e)); 7 throw new HiveException(e); 8 } 9 }
这里的createMetaPartitionObject作用在于整个Partition传入对象的校验对对象的封装,代码如下:
1 public static org.apache.hadoop.hive.metastore.api.Partition createMetaPartitionObject( 2 Table tbl, MappartSpec, Path location) throws HiveException { 3 List pvals = new ArrayList ();
//遍历整个PartCols,并且校验partMap中是否一一对应 4 for (FieldSchema field : tbl.getPartCols()) { 5 String val = partSpec.get(field.getName()); 6 if (val == null || val.isEmpty()) { 7 throw new HiveException("partition spec is invalid; field " 8 + field.getName() + " does not exist or is empty"); 9 } 10 pvals.add(val); 11 } 12 //set相关的属性信息,包括DbName、TableName、PartValues、以及sd信息 13 org.apache.hadoop.hive.metastore.api.Partition tpart = 14 new org.apache.hadoop.hive.metastore.api.Partition(); 15 tpart.setDbName(tbl.getDbName()); 16 tpart.setTableName(tbl.getTableName()); 17 tpart.setValues(pvals); 18 19 if (!tbl.isView()) { 20 tpart.setSd(cloneS d(tbl)); 21 tpart.getSd().setLocation((location != null) ? location.toString() : null); 22 } 23 return tpart; 24 }
随之MetaDataClient对于该对象调用MetaDataService的addPartition,并进行了深拷贝,这里不再详细说明,那么我们直接看下服务端干了什么:
1 private Partition add_partition_core(final RawStore ms, 2 final Partition part, final EnvironmentContext envContext) 3 throws InvalidObjectException, AlreadyExistsException, MetaException, TException { 4 boolean success = false; 5 Table tbl = null; 6 try { 7 ms.openTransaction();
//根据DbName、TableName获取整个Table对象信息 8 tbl = ms.getTable(part.getDbName(), part.getTableName()); 9 if (tbl == null) { 10 throw new InvalidObjectException( 11 "Unable to add partition because table or database do not exist"); 12 } 13 //事件处理 14 firePreEvent(new PreAddPartitionEvent(tbl, part, this)); 15 //在创建Partition之前,首先会校验元数据中该partition是否存在 16 boolean shouldAdd = startAddPartition(ms, part, false); 17 assert shouldAdd; // start would throw if it already existed here
//创建Partition路径 18 boolean madeDir = createLocationForAddedPartition(tbl, part); 19 try {
//加载一些kv信息 20 initializeAddedPartition(tbl, part, madeDir);
//写入元数据 21 success = ms.addPartition(part); 22 } finally { 23 if (!success && madeDir) {
//如果没有成功,便删除物理路径 24 wh.deleteDir(new Path(part.getSd().getLocation()), true); 25 } 26 } 27 // we proceed only if we'd actually succeeded anyway, otherwise, 28 // we'd have thrown an exception 29 success = success && ms.commitTransaction(); 30 } finally { 31 if (!success) { 32 ms.rollbackTransaction(); 33 } 34 fireMetaStoreAddPartitionEvent(tbl, Arrays.asList(part), envContext, success); 35 } 36 return part; 37 }
这里提及一个设计上的点,从之前的表结构设计上,没有直接存储PartName,而是将key与value单独存在与kv表中,这里我们看下createLocationForAddedPartition:
1 private boolean createLocationForAddedPartition( 2 final Table tbl, final Partition part) throws MetaException { 3 Path partLocation = null; 4 String partLocationStr = null;
//如果sd不为null,则将sd的location信息作为表跟目录赋给partLocationStr 5 if (part.getSd() != null) { 6 partLocationStr = part.getSd().getLocation(); 7 } 8 //如果为null,则重新拼接part Location 9 if (partLocationStr == null || partLocationStr.isEmpty()) { 10 // set default location if not specified and this is 11 // a physical table partition (not a view) 12 if (tbl.getSd().getLocation() != null) {
//如果不为null,则继续拼接文件路径及part的路径,组成完成的Partition location 13 partLocation = new Path(tbl.getSd().getLocation(), Warehouse 14 .makePartName(tbl.getPartitionKeys(), part.getValues())); 15 } 16 } else { 17 if (tbl.getSd().getLocation() == null) { 18 throw new MetaException("Cannot specify location for a view partition"); 19 } 20 partLocation = wh.getDnsPath(new Path(partLocationStr)); 21 } 22 23 boolean result = false;
//将location信息写入sd表 24 if (partLocation != null) { 25 part.getSd().setLocation(partLocation.toString()); 26 27 // Check to see if the directory already exists before calling 28 // mkdirs() because if the file system is read-only, mkdirs will 29 // throw an exception even if the directory already exists. 30 if (!wh.isDir(partLocation)) { 31 if (!wh.mkdirs(partLocation, true)) { 32 throw new MetaException(partLocation 33 + " is not a directory or unable to create one"); 34 } 35 result = true; 36 } 37 } 38 return result; 39 }
总结:
7、dropPartition
删除partition就不再从Hive开始了,我们直接看HiveMetaStore服务端做了什么:
1 private boolean drop_partition_common(RawStore ms, String db_name, String tbl_name, 2 Listpart_vals, final boolean deleteData, final EnvironmentContext envContext) 3 throws MetaException, NoSuchObjectException, IOException, InvalidObjectException, 4 InvalidInputException { 5 boolean success = false; 6 Path partPath = null; 7 Table tbl = null; 8 Partition part = null; 9 boolean isArchived = false; 10 Path archiveParentDir = null; 11 boolean mustPurge = false; 12 13 try { 14 ms.openTransaction();
//根据dbName、tableName、part_values获取整个part信息 15 part = ms.getPartition(db_name, tbl_name, part_vals);
//获取所有Table对象 16 tbl = get_table_core(db_name, tbl_name); 17 firePreEvent(new PreDropPartitionEvent(tbl, part, deleteData, this)); 18 mustPurge = isMustPurge(envContext, tbl); 19 20 if (part == null) { 21 throw new NoSuchObjectException("Partition doesn't exist. " 22 + part_vals); 23 } 24 //这一片还没有深入看Arrchived partition 25 isArchived = MetaStoreUtils.isArchived(part); 26 if (isArchived) { 27 archiveParentDir = MetaStoreUtils.getOriginalLocation(part); 28 verifyIsWritablePath(archiveParentDir); 29 checkTrashPurgeCombination(archiveParentDir, db_name + "." + tbl_name + "." + part_vals, mustPurge); 30 } 31 if (!ms.dropPartition(db_name, tbl_name, part_vals)) { 32 throw new MetaException("Unable to drop partition"); 33 } 34 success = ms.commitTransaction(); 35 if ((part.getSd() != null) && (part.getSd().getLocation() != null)) { 36 partPath = new Path(part.getSd().getLocation()); 37 verifyIsWritablePath(partPath); 38 checkTrashPurgeCombination(partPath, db_name + "." + tbl_name + "." + part_vals, mustPurge); 39 } 40 } finally { 41 if (!success) { 42 ms.rollbackTransaction(); 43 } else if (deleteData && ((partPath != null) || (archiveParentDir != null))) { 44 if (tbl != null && !isExternal(tbl)) { 45 if (mustPurge) { 46 LOG.info("dropPartition() will purge " + partPath + " directly, skipping trash."); 47 } 48 else { 49 LOG.info("dropPartition() will move " + partPath + " to trash-directory."); 50 }
//删除partition 51 // Archived partitions have har:/to_har_file as their location. 52 // The original directory was saved in params 53 if (isArchived) { 54 assert (archiveParentDir != null); 55 wh.deleteDir(archiveParentDir, true, mustPurge); 56 } else { 57 assert (partPath != null); 58 wh.deleteDir(partPath, true, mustPurge); 59 deleteParentRecursive(partPath.getParent(), part_vals.size() - 1, mustPurge); 60 } 61 // ok even if the data is not deleted 62 } 63 } 64 for (MetaStoreEventListener listener : listeners) { 65 DropPartitionEvent dropPartitionEvent = 66 new DropPartitionEvent(tbl, part, success, deleteData, this); 67 dropPartitionEvent.setEnvironmentContext(envContext); 68 listener.onDropPartition(dropPartitionEvent); 69 } 70 } 71 return true; 72 }
8、alterPartition
alterPartition牵扯的校验及文件目录的修改,我们直接从HiveMetaStore中的rename_partition中查看:
1 private void rename_partition(final String db_name, final String tbl_name, 2 final Listpart_vals, final Partition new_part, 3 final EnvironmentContext envContext) 4 throws InvalidOperationException, MetaException, 5 TException {
//日志记录 6 startTableFunction("alter_partition", db_name, tbl_name); 7 8 if (LOG.isInfoEnabled()) { 9 LOG.info("New partition values:" + new_part.getValues()); 10 if (part_vals != null && part_vals.size() > 0) { 11 LOG.info("Old Partition values:" + part_vals); 12 } 13 } 14 15 Partition oldPart = null; 16 Exception ex = null; 17 try { 18 firePreEvent(new PreAlterPartitionEvent(db_name, tbl_name, part_vals, new_part, this)); 19 //校验PartName的规范性 20 if (part_vals != null && !part_vals.isEmpty()) { 21 MetaStoreUtils.validatePartitionNameCharacters(new_part.getValues(), 22 partitionValidationPattern); 23 } 24 调用alterHandler的alterPartition进行partition物理上的rename,以及元数据修改 25 oldPart = alterHandler.alterPartition(getMS(), wh, db_name, tbl_name, part_vals, new_part); 26 27 // Only fetch the table if we actually have a listener 28 Table table = null; 29 for (MetaStoreEventListener listener : listeners) { 30 if (table == null) { 31 table = getMS().getTable(db_name, tbl_name); 32 } 33 AlterPartitionEvent alterPartitionEvent = 34 new AlterPartitionEvent(oldPart, new_part, table, true, this); 35 alterPartitionEvent.setEnvironmentContext(envContext); 36 listener.onAlterPartition(alterPartitionEvent); 37 } 38 } catch (InvalidObjectException e) { 39 ex = e; 40 throw new InvalidOperationException(e.getMessage()); 41 } catch (AlreadyExistsException e) { 42 ex = e; 43 throw new InvalidOperationException(e.getMessage()); 44 } catch (Exception e) { 45 ex = e; 46 if (e instanceof MetaException) { 47 throw (MetaException) e; 48 } else if (e instanceof InvalidOperationException) { 49 throw (InvalidOperationException) e; 50 } else if (e instanceof TException) { 51 throw (TException) e; 52 } else { 53 throw newMetaException(e); 54 } 55 } finally { 56 endFunction("alter_partition", oldPart != null, ex, tbl_name); 57 } 58 return; 59 }
这里我们着重看一下,alterHandler.alterPartition方法,前方高能:
1 public Partition alterPartition(final RawStore msdb, Warehouse wh, final String dbname, 2 final String name, final Listpart_vals, final Partition new_part) 3 throws InvalidOperationException, InvalidObjectException, AlreadyExistsException, 4 MetaException { 5 boolean success = false; 6 7 Path srcPath = null; 8 Path destPath = null; 9 FileSystem srcFs = null; 10 FileSystem destFs = null; 11 Partition oldPart = null; 12 String oldPartLoc = null; 13 String newPartLoc = null; 14 15 //修改新的partition的DDL时间 16 if (new_part.getParameters() == null || 17 new_part.getParameters().get(hive_metastoreConstants.DDL_TIME) == null || 18 Integer.parseInt(new_part.getParameters().get(hive_metastoreConstants.DDL_TIME)) == 0) { 19 new_part.putToParameters(hive_metastoreConstants.DDL_TIME, Long.toString(System 20 .currentTimeMillis() / 1000)); 21 } 22 //根据dbName、tableName获取整个Table对象 23 Table tbl = msdb.getTable(dbname, name); 24 //如果传入的part_vals为空或为0,说明修改的只是partition的其他元数据信息而不牵扯到partKV,则直接元数据,在msdb.alterPartition会直接更新 25 if (part_vals == null || part_vals.size() == 0) { 26 try { 27 oldPart = msdb.getPartition(dbname, name, new_part.getValues()); 28 if (MetaStoreUtils.requireCalStats(hiveConf, oldPart, new_part, tbl)) { 29 MetaStoreUtils.updatePartitionStatsFast(new_part, wh, false, true); 30 } 31 updatePartColumnStats(msdb, dbname, name, new_part.getValues(), new_part); 32 msdb.alterPartition(dbname, name, new_part.getValues(), new_part); 33 } catch (InvalidObjectException e) { 34 throw new InvalidOperationException("alter is not possible"); 35 } catch (NoSuchObjectException e){ 36 //old partition does not exist 37 throw new InvalidOperationException("alter is not possible"); 38 } 39 return oldPart; 40 } 41 //rename partition 42 try { 43 msdb.openTransaction(); 44 try {
//获取oldPart对象信息 45 oldPart = msdb.getPartition(dbname, name, part_vals); 46 } catch (NoSuchObjectException e) { 47 // this means there is no existing partition 48 throw new InvalidObjectException( 49 "Unable to rename partition because old partition does not exist"); 50 } 51 Partition check_part = null; 52 try {
//组装newPart的partValues等Partition信息 53 check_part = msdb.getPartition(dbname, name, new_part.getValues()); 54 } catch(NoSuchObjectException e) { 55 // this means there is no existing partition 56 check_part = null; 57 }
//如果check_part组装成功,说明该part已经存在,则报already exists 58 if (check_part != null) { 59 throw new AlreadyExistsException("Partition already exists:" + dbname + "." + name + "." + 60 new_part.getValues()); 61 }
//table的信息校验 62 if (tbl == null) { 63 throw new InvalidObjectException( 64 "Unable to rename partition because table or database do not exist"); 65 } 66 67 //如果是外部表的分区变化了,那么不需要操作文件系统,直接更新meta信息即可 68 if (tbl.getTableType().equals(TableType.EXTERNAL_TABLE.toString())) { 69 new_part.getSd().setLocation(oldPart.getSd().getLocation()); 70 String oldPartName = Warehouse.makePartName(tbl.getPartitionKeys(), oldPart.getValues()); 71 try { 72 //existing partition column stats is no longer valid, remove 73 msdb.deletePartitionColumnStatistics(dbname, name, oldPartName, oldPart.getValues(), null); 74 } catch (NoSuchObjectException nsoe) { 75 //ignore 76 } catch (InvalidInputException iie) { 77 throw new InvalidOperationException("Unable to update partition stats in table rename." + iie); 78 } 79 msdb.alterPartition(dbname, name, part_vals, new_part); 80 } else { 81 try {
//获取Table的文件路径 82 destPath = new Path(wh.getTablePath(msdb.getDatabase(dbname), name), 83 Warehouse.makePartName(tbl.getPartitionKeys(), new_part.getValues()));
//拼接新的Partition的路径 84 destPath = constructRenamedPath(destPath, new Path(new_part.getSd().getLocation())); 85 } catch (NoSuchObjectException e) { 86 LOG.debug(e); 87 throw new InvalidOperationException( 88 "Unable to change partition or table. Database " + dbname + " does not exist" 89 + " Check metastore logs for detailed stack." + e.getMessage()); 90 }
//如果destPath不为空,说明改变了文件路径 91 if (destPath != null) { 92 newPartLoc = destPath.toString(); 93 oldPartLoc = oldPart.getSd().getLocation(); 94 //根据原有sd的路径获取老的part路径信息 95 srcPath = new Path(oldPartLoc); 96 97 LOG.info("srcPath:" + oldPartLoc); 98 LOG.info("descPath:" + newPartLoc); 99 srcFs = wh.getFs(srcPath); 100 destFs = wh.getFs(destPath); 101 //查看srcFS与destFs是否Wie同一个fileSystem 102 if (!FileUtils.equalsFileSystem(srcFs, destFs)) { 103 throw new InvalidOperationException("table new location " + destPath 104 + " is on a different file system than the old location " 105 + srcPath + ". This operation is not supported"); 106 } 107 try {
//校验老的partition路径与新的partition路径是否一致,同时新的partition路径是否已经存在 108 srcFs.exists(srcPath); // check that src exists and also checks 109 if (newPartLoc.compareTo(oldPartLoc) != 0 && destFs.exists(destPath)) { 110 throw new InvalidOperationException("New location for this table " 111 + tbl.getDbName() + "." + tbl.getTableName() 112 + " already exists : " + destPath); 113 } 114 } catch (IOException e) { 115 throw new InvalidOperationException("Unable to access new location " 116 + destPath + " for partition " + tbl.getDbName() + "." 117 + tbl.getTableName() + " " + new_part.getValues()); 118 } 119 new_part.getSd().setLocation(newPartLoc); 120 if (MetaStoreUtils.requireCalStats(hiveConf, oldPart, new_part, tbl)) { 121 MetaStoreUtils.updatePartitionStatsFast(new_part, wh, false, true); 122 }
//拼接oldPartName,并且删除原有oldPart的信息,写入新的partition信息 123 String oldPartName = Warehouse.makePartName(tbl.getPartitionKeys(), oldPart.getValues()); 124 try { 125 //existing partition column stats is no longer valid, remove 126 msdb.deletePartitionColumnStatistics(dbname, name, oldPartName, oldPart.getValues(), null); 127 } catch (NoSuchObjectException nsoe) { 128 //ignore 129 } catch (InvalidInputException iie) { 130 throw new InvalidOperationException("Unable to update partition stats in table rename." + iie); 131 } 132 msdb.alterPartition(dbname, name, part_vals, new_part); 133 } 134 } 135 136 success = msdb.commitTransaction(); 137 } finally { 138 if (!success) { 139 msdb.rollbackTransaction(); 140 } 141 if (success && newPartLoc != null && newPartLoc.compareTo(oldPartLoc) != 0) { 142 //rename the data directory 143 try{ 144 if (srcFs.exists(srcPath)) { 145 //如果根路径海微创建,需要重新进行创建,就好比计算引擎先调用了alterTable,又调用了alterPartition,这时partition的根路径或许还未创建 146 Path destParentPath = destPath.getParent(); 147 if (!wh.mkdirs(destParentPath, true)) { 148 throw new IOException("Unable to create path " + destParentPath); 149 }
//进行原路径与目标路径的rename 150 wh.renameDir(srcPath, destPath, true); 151 LOG.info("rename done!"); 152 } 153 } catch (IOException e) { 154 boolean revertMetaDataTransaction = false; 155 try { 156 msdb.openTransaction(); 157 msdb.alterPartition(dbname, name, new_part.getValues(), oldPart); 158 revertMetaDataTransaction = msdb.commitTransaction(); 159 } catch (Exception e1) { 160 LOG.error("Reverting metadata opeation failed During HDFS operation failed", e1); 161 if (!revertMetaDataTransaction) { 162 msdb.rollbackTransaction(); 163 } 164 } 165 throw new InvalidOperationException("Unable to access old location " 166 + srcPath + " for partition " + tbl.getDbName() + "." 167 + tbl.getTableName() + " " + part_vals); 168 } 169 } 170 } 171 return oldPart; 172 }
暂时到这里吧~后续咱们慢慢玩哈~