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：
private HiveConf conf = null;
private IMetaStoreClient metaStoreClient;
private UserGroupInformation owner; // metastore calls timing information
private final Map<String, Long> metaCallTimeMap = new HashMap<String, Long>(); private static ThreadLocal<Hive> hiveDB = new ThreadLocal<Hive>() {
@Override
protected synchronized Hive initialValue() {
return null;
} @Override
public synchronized void remove() {
if (this.get() != null) {
this.get().close();
}
super.remove();
}
};
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：
public static Hive get(HiveConf c, boolean needsRefresh) throws HiveException {
Hive db = hiveDB.get();
if (db == null || needsRefresh || !db.isCurrentUserOwner()) {
if (db != null) {
LOG.debug("Creating new db. db = " + db + ", needsRefresh = " + needsRefresh +
", db.isCurrentUserOwner = " + db.isCurrentUserOwner());
}
closeCurrent();
c.set("fs.scheme.class", "dfs");
Hive newdb = new Hive(c);
hiveDB.set(newdb);
return newdb;
}
db.conf = c;
return db;
}
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：
// register all permanent functions. need improvement
static {
try {
reloadFunctions();
} catch (Exception e) {
LOG.warn("Failed to access metastore. This class should not accessed in runtime.",e);
}
} public static void reloadFunctions() throws HiveException {
//获取 Hive对象，用于后续方法的调用
Hive db = Hive.get();
//通过遍历每一个dbName
for (String dbName : db.getAllDatabases()) {
//通过dbName查询挂在该db下的所有function的信息。
for (String functionName : db.getFunctions(dbName, "*")) {
Function function = db.getFunction(dbName, functionName);
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<String,FunctionInfo>中，以便计算引擎在调用时，不必再次查询数据库。
FunctionRegistry.registerPermanentFunction(
FunctionUtils.qualifyFunctionName(functionName, dbName), function.getClassName(),
false, FunctionTask.toFunctionResource(function.getResourceUris()));
} catch (Exception e) {
LOG.warn("Failed to register persistent function " +
functionName + ":" + function.getClassName() + ". Ignore and continue.");
}
}
}
}
调用getMSC()方法，进行metadataClient客户端的创建，代码如下：
1 private IMetaStoreClient createMetaStoreClient() throws MetaException {
2
//这里实现接口HiveMetaHookLoader
3 HiveMetaHookLoader hookLoader = new HiveMetaHookLoader() {
4 @Override
5 public HiveMetaHook getHook(
6 org.apache.hadoop.hive.metastore.api.Table tbl)
7 throws MetaException {
8
9 try {
10 if (tbl == null) {
11 return null;
12 }
//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
13 HiveStorageHandler storageHandler =
14 HiveUtils.getStorageHandler(conf,
15 tbl.getParameters().get(META_TABLE_STORAGE));
16 if (storageHandler == null) {
17 return null;
18 }
19 return storageHandler.getMetaHook();
20 } catch (HiveException ex) {
21 LOG.error(StringUtils.stringifyException(ex));
22 throw new MetaException(
23 "Failed to load storage handler: " + ex.getMessage());
24 }
25 }
26 };
27 return RetryingMetaStoreClient.getProxy(conf, hookLoader, metaCallTimeMap,
28 SessionHiveMetaStoreClient.class.getName());
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：
public HiveMetaStoreClient(HiveConf conf, HiveMetaHookLoader hookLoader)
throws MetaException { this.hookLoader = hookLoader;
if (conf == null) {
conf = new HiveConf(HiveMetaStoreClient.class);
}
this.conf = conf;
filterHook = loadFilterHooks();
//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
String msUri = conf.getVar(HiveConf.ConfVars.METASTOREURIS);
localMetaStore = HiveConfUtil.isEmbeddedMetaStore(msUri);
if (localMetaStore) {
//本地连接直接连接HiveMetaStore
client = HiveMetaStore.newRetryingHMSHandler("hive client", conf, true);
isConnected = true;
snapshotActiveConf();
return;
} //获取配置中的重试次数及timeout时间
retries = HiveConf.getIntVar(conf, HiveConf.ConfVars.METASTORETHRIFTCONNECTIONRETRIES);
retryDelaySeconds = conf.getTimeVar(
ConfVars.METASTORE_CLIENT_CONNECT_RETRY_DELAY, TimeUnit.SECONDS); //拼接metastore uri
if (conf.getVar(HiveConf.ConfVars.METASTOREURIS) != null) {
String metastoreUrisString[] = conf.getVar(
HiveConf.ConfVars.METASTOREURIS).split(",");
metastoreUris = new URI[metastoreUrisString.length];
try {
int i = 0;
for (String s : metastoreUrisString) {
URI tmpUri = new URI(s);
if (tmpUri.getScheme() == null) {
throw new IllegalArgumentException("URI: " + s
+ " does not have a scheme");
}
metastoreUris[i++] = tmpUri; }
} catch (IllegalArgumentException e) {
throw (e);
} catch (Exception e) {
MetaStoreUtils.logAndThrowMetaException(e);
}
} else {
LOG.error("NOT getting uris from conf");
throw new MetaException("MetaStoreURIs not found in conf file");
}
调用open方法创建连接
open();
}
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：
private void open() throws MetaException {
isConnected = false;
TTransportException tte = null;
//是否使用Sasl
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.
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
boolean useCompactProtocol = conf.getBoolVar(ConfVars.METASTORE_USE_THRIFT_COMPACT_PROTOCOL);
//获取socket timeout时间
int clientSocketTimeout = (int) conf.getTimeVar(
ConfVars.METASTORE_CLIENT_SOCKET_TIMEOUT, TimeUnit.MILLISECONDS);
for (int attempt = 0; !isConnected && attempt < retries; ++attempt) {
for (URI store : metastoreUris) {
LOG.info("Trying to connect to metastore with URI " + store);
try {
transport = new TSocket(store.getHost(), store.getPort(), clientSocketTimeout);
if (useSasl) {
// Wrap thrift connection with SASL for secure connection.
try {
//创建HadoopThriftAuthBridge client
HadoopThriftAuthBridge.Client authBridge =
ShimLoader.getHadoopThriftAuthBridge().createClient();
//权限认证相关
// check if we should use delegation tokens to authenticate
// the call below gets hold of the tokens if they are set up by hadoop
// this should happen on the map/reduce tasks if the client added the
// tokens into hadoop's credential store in the front end during job
// submission.
String tokenSig = conf.get("hive.metastore.token.signature");
// tokenSig could be null
tokenStrForm = Utils.getTokenStrForm(tokenSig);
if(tokenStrForm != null) {
// authenticate using delegation tokens via the "DIGEST" mechanism
transport = authBridge.createClientTransport(null, store.getHost(),
"DIGEST", tokenStrForm, transport,
MetaStoreUtils.getMetaStoreSaslProperties(conf));
} else {
String principalConfig =
conf.getVar(HiveConf.ConfVars.METASTORE_KERBEROS_PRINCIPAL);
transport = authBridge.createClientTransport(
principalConfig, store.getHost(), "KERBEROS", null,
transport, MetaStoreUtils.getMetaStoreSaslProperties(conf));
}
} catch (IOException ioe) {
LOG.error("Couldn't create client transport", ioe);
throw new MetaException(ioe.toString());
}
} else if (useFramedTransport) {
transport = new TFramedTransport(transport);
}
final TProtocol protocol;
//后续详细说明两者的区别（因为俺还没看，哈哈）
if (useCompactProtocol) {
protocol = new TCompactProtocol(transport);
} else {
protocol = new TBinaryProtocol(transport);
}
//创建ThriftHiveMetastore client
client = new ThriftHiveMetastore.Client(protocol);
try {
transport.open();
isConnected = true;
} catch (TTransportException e) {
tte = e;
if (LOG.isDebugEnabled()) {
LOG.warn("Failed to connect to the MetaStore Server...", e);
} else {
// Don't print full exception trace if DEBUG is not on.
LOG.warn("Failed to connect to the MetaStore Server...");
}
}
//用户组及用户的加载
if (isConnected && !useSasl && conf.getBoolVar(ConfVars.METASTORE_EXECUTE_SET_UGI)){
// Call set_ugi, only in unsecure mode.
try {
UserGroupInformation ugi = Utils.getUGI();
client.set_ugi(ugi.getUserName(), Arrays.asList(ugi.getGroupNames()));
} catch (LoginException e) {
LOG.warn("Failed to do login. set_ugi() is not successful, " +
"Continuing without it.", e);
} catch (IOException e) {
LOG.warn("Failed to find ugi of client set_ugi() is not successful, " +
"Continuing without it.", e);
} catch (TException e) {
LOG.warn("set_ugi() not successful, Likely cause: new client talking to old server. "
+ "Continuing without it.", e);
}
}
} catch (MetaException e) {
LOG.error("Unable to connect to metastore with URI " + store
+ " in attempt " + attempt, e);
}
if (isConnected) {
break;
}
}
// Wait before launching the next round of connection retries.
if (!isConnected && retryDelaySeconds > 0) {
try {
LOG.info("Waiting " + retryDelaySeconds + " seconds before next connection attempt.");
Thread.sleep(retryDelaySeconds * 1000);
} catch (InterruptedException ignore) {}
}
}
if (!isConnected) {
throw new MetaException("Could not connect to meta store using any of the URIs provided." +
" Most recent failure: " + StringUtils.stringifyException(tte));
}
snapshotActiveConf();
LOG.info("Connected to metastore.");
}
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：
public HMSHandler(String name, HiveConf conf, boolean init) throws MetaException {
super(name);
hiveConf = conf;
if (init) {
init();
}
}
下俩我们继续看它的init方法：
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。
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
initListeners = MetaStoreUtils.getMetaStoreListeners(
MetaStoreInitListener.class, hiveConf,
hiveConf.getVar(HiveConf.ConfVars.METASTORE_INIT_HOOKS));
for (MetaStoreInitListener singleInitListener: initListeners) {
MetaStoreInitContext context = new MetaStoreInitContext();
singleInitListener.onInit(context);
}
//初始化alter的实现类
String alterHandlerName = hiveConf.get("hive.metastore.alter.impl",
HiveAlterHandler.class.getName());
alterHandler = (AlterHandler) ReflectionUtils.newInstance(MetaStoreUtils.getClass(
alterHandlerName), hiveConf);
//初始化warehouse
wh = new Warehouse(hiveConf);
//创建默认db以及用户，同时加载currentUrl
synchronized (HMSHandler.class) {
if (currentUrl == null || !currentUrl.equals(MetaStoreInit.getConnectionURL(hiveConf))) {
createDefaultDB();
createDefaultRoles();
addAdminUsers();
currentUrl = MetaStoreInit.getConnectionURL(hiveConf);
}
}
//计数信息的初始化
if (hiveConf.getBoolean("hive.metastore.metrics.enabled", false)) {
try {
Metrics.init();
} catch (Exception e) {
// log exception, but ignore inability to start
LOG.error("error in Metrics init: " + e.getClass().getName() + " "
+ e.getMessage(), e);
}
}
//Listener的PreListener的初始化
preListeners = MetaStoreUtils.getMetaStoreListeners(MetaStorePreEventListener.class,
hiveConf,
hiveConf.getVar(HiveConf.ConfVars.METASTORE_PRE_EVENT_LISTENERS));
listeners = MetaStoreUtils.getMetaStoreListeners(MetaStoreEventListener.class, hiveConf,
hiveConf.getVar(HiveConf.ConfVars.METASTORE_EVENT_LISTENERS));
listeners.add(new SessionPropertiesListener(hiveConf));
endFunctionListeners = MetaStoreUtils.getMetaStoreListeners(
MetaStoreEndFunctionListener.class, hiveConf,
hiveConf.getVar(HiveConf.ConfVars.METASTORE_END_FUNCTION_LISTENERS));
//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
String partitionValidationRegex =
hiveConf.getVar(HiveConf.ConfVars.METASTORE_PARTITION_NAME_WHITELIST_PATTERN);
if (partitionValidationRegex != null && !partitionValidationRegex.isEmpty()) {
partitionValidationPattern = Pattern.compile(partitionValidationRegex);
} else {
partitionValidationPattern = null;
}
long cleanFreq = hiveConf.getTimeVar(ConfVars.METASTORE_EVENT_CLEAN_FREQ, TimeUnit.MILLISECONDS);
if (cleanFreq > 0) {
// In default config, there is no timer.
Timer cleaner = new Timer("Metastore Events Cleaner Thread", true);
cleaner.schedule(new EventCleanerTask(this), cleanFreq, cleanFreq);
}
}
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：
public void createTable(String tableName, List<String> columns, List<String> partCols,
Class<? extends InputFormat> fileInputFormat,
Class<?> fileOutputFormat, int bucketCount, List<String> bucketCols,
Map<String, String> parameters) throws HiveException {
if (columns == null) {
throw new HiveException("columns not specified for table " + tableName);
} 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
tbl.setInputFormatClass(fileInputFormat.getName());
tbl.setOutputFormatClass(fileOutputFormat.getName());
//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
for (String col : columns) {
FieldSchema field = new FieldSchema(col, STRING_TYPE_NAME, "default");
tbl.getCols().add(field);
}
//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
if (partCols != null) {
for (String partCol : partCols) {
FieldSchema part = new FieldSchema();
part.setName(partCol);
part.setType(STRING_TYPE_NAME); // default partition key
tbl.getPartCols().add(part);
}
}
//设置序列化的方式
tbl.setSerializationLib(LazySimpleSerDe.class.getName());
//设置分桶信息
tbl.setNumBuckets(bucketCount);
tbl.setBucketCols(bucketCols);
//设置table额外添加的kv信息
if (parameters != null) {
tbl.setParamters(parameters);
}
createTable(tbl);
}
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：
public void createTable(Table tbl, EnvironmentContext envContext) throws AlreadyExistsException,
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
HiveMetaHook hook = getHook(tbl);
if (hook != null) {
hook.preCreateTable(tbl);
}
boolean success = false;
try {
//随即调用HiveMetaStore进行服务端与数据库的创建交互
create_table_with_environment_context(tbl, envContext);
if (hook != null) {
hook.commitCreateTable(tbl);
}
success = true;
} finally {
如果创建失败的话，进行回滚操作
if (!success && (hook != null)) {
hook.rollbackCreateTable(tbl);
}
}
}
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：
public interface HiveMetaHook {
/**
* Called before a new table definition is added to the metastore
* during CREATE TABLE.
*
* @param table new table definition
*/
public void preCreateTable(Table table)
throws MetaException;
/**
* Called after failure adding a new table definition to the metastore
* during CREATE TABLE.
*
* @param table new table definition
*/
public void rollbackCreateTable(Table table)
throws MetaException;
public void preDropTale(Table table)
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)
throws AlreadyExistsException, MetaException,
InvalidObjectException, NoSuchObjectException {
//名称正则校验，校验是否含有非法字符
if (!MetaStoreUtils.validateName(tbl.getTableName())) {
throw new InvalidObjectException(tbl.getTableName()
+ " is not a valid object name");
}
//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
String validate = MetaStoreUtils.validateTblColumns(tbl.getSd().getCols());
if (validate != null) {
throw new InvalidObjectException("Invalid column " + validate);
}
if (tbl.getPartitionKeys() != null) {
validate = MetaStoreUtils.validateTblColumns(tbl.getPartitionKeys());
if (validate != null) {
throw new InvalidObjectException("Invalid partition column " + validate);
}
}
SkewedInfo skew = tbl.getSd().getSkewedInfo();
if (skew != null) {
validate = MetaStoreUtils.validateSkewedColNames(skew.getSkewedColNames());
if (validate != null) {
throw new InvalidObjectException("Invalid skew column " + validate);
}
validate = MetaStoreUtils.validateSkewedColNamesSubsetCol(
skew.getSkewedColNames(), tbl.getSd().getCols());
if (validate != null) {
throw new InvalidObjectException("Invalid skew column " + validate);
}
} Path tblPath = null;
boolean success = false, madeDir = false;
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。
firePreEvent(new PreCreateTableEvent(tbl, this));
//打开事务
ms.openTransaction();
//如果db不存在的情况下，则抛异常
Database db = ms.getDatabase(tbl.getDbName());
if (db == null) {
throw new NoSuchObjectException("The database " + tbl.getDbName() + " does not exist");
}
// 校验该db下，table是否存在
if (is_table_exists(ms, tbl.getDbName(), tbl.getTableName())) {
throw new AlreadyExistsException("Table " + tbl.getTableName()
+ " already exists");
}
// 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())
if (!TableType.VIRTUAL_VIEW.toString().equals(tbl.getTableType())) {
if (tbl.getSd().getLocation() == null
|| tbl.getSd().getLocation().isEmpty()) {
tblPath = wh.getTablePath(
ms.getDatabase(tbl.getDbName()), tbl.getTableName());
} 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
if (!isExternal(tbl) && !MetaStoreUtils.isNonNativeTable(tbl)) {
LOG.warn("Location: " + tbl.getSd().getLocation()
+ " specified for non-external table:" + tbl.getTableName());
}
tblPath = wh.getDnsPath(new Path(tbl.getSd().getLocation()));
}
//将拼接完的tblPath set到sd的location中
tbl.getSd().setLocation(tblPath.toString());
}
//创建table的路径
if (tblPath != null) {
if (!wh.isDir(tblPath)) {
if (!wh.mkdirs(tblPath, true)) {
throw new MetaException(tblPath
+ " is not a directory or unable to create one");
}
madeDir = true;
}
}
// hive.stats.autogather 配置判断
if (HiveConf.getBoolVar(hiveConf, HiveConf.ConfVars.HIVESTATSAUTOGATHER) &&
!MetaStoreUtils.isView(tbl)) {
if (tbl.getPartitionKeysSize() == 0) { // Unpartitioned table
MetaStoreUtils.updateUnpartitionedTableStatsFast(db, tbl, wh, madeDir);
} else { // Partitioned table with no partitions.
MetaStoreUtils.updateUnpartitionedTableStatsFast(db, tbl, wh, true);
}
}
// set create time
long time = System.currentTimeMillis() / 1000;
tbl.setCreateTime((int) time);
if (tbl.getParameters() == null ||
tbl.getParameters().get(hive_metastoreConstants.DDL_TIME) == null) {
tbl.putToParameters(hive_metastoreConstants.DDL_TIME, Long.toString(time));
}
执行createTable数据库操作
ms.createTable(tbl);
success = ms.commitTransaction();
} finally {
if (!success) {
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
if (madeDir) {
wh.deleteDir(tblPath, true);
}
}
//进行create完成时的listener类发送 比如 noftify通知
for (MetaStoreEventListener listener : listeners) {
CreateTableEvent createTableEvent =
new CreateTableEvent(tbl, success, this);
createTableEvent.setEnvironmentContext(envContext);
listener.onCreateTable(createTableEvent);
}
}
}
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：
public abstract Database getDatabase(String name)
throws NoSuchObjectException; public abstract boolean dropDatabase(String dbname) throws NoSuchObjectException, MetaException; public abstract boolean alterDatabase(String dbname, Database db) throws NoSuchObjectException, MetaException; public abstract List<String> getDatabases(String pattern) throws MetaException; public abstract List<String> getAllDatabases() throws MetaException; public abstract boolean createType(Type type); public abstract Type getType(String typeName); public abstract boolean dropType(String typeName); public abstract void createTable(Table tbl) throws InvalidObjectException,
MetaException; public abstract boolean dropTable(String dbName, String tableName)
throws MetaException, NoSuchObjectException, InvalidObjectException, InvalidInputException; public abstract Table getTable(String dbName, String tableName)
throws MetaException;
..................
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：
public RawStore getMS() throws MetaException {
//获取本地线程中已存在的RawStore
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
if (ms == null) {
ms = newRawStore();
ms.verifySchema();
threadLocalMS.set(ms);
ms = threadLocalMS.get();
}
return ms;
}
看到这里，是不是很想看看newRawStore它干嘛啦？那么我们继续：
public static RawStore getProxy(HiveConf hiveConf, Configuration conf, String rawStoreClassName,
int id) throws MetaException {
//通过反射，创建baseClass，随后再进行该实现对象的创建
Class<? extends RawStore> baseClass = (Class<? extends RawStore>) MetaStoreUtils.getClass(
rawStoreClassName); RawStoreProxy handler = new RawStoreProxy(hiveConf, conf, baseClass, id); // Look for interfaces on both the class and all base classes.
return (RawStore) Proxy.newProxyInstance(RawStoreProxy.class.getClassLoader(),
getAllInterfaces(baseClass), handler);
}
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：
@Override
public void createTable(Table tbl) throws InvalidObjectException, MetaException {
boolean commited = false;
try {
//创建事务
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
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
pm.makePersistent(mtbl);
//封装权限用户、角色、组对象并写入
PrincipalPrivilegeSet principalPrivs = tbl.getPrivileges();
List<Object> toPersistPrivObjs = new ArrayList<Object>();
if (principalPrivs != null) {
int now = (int)(System.currentTimeMillis()/1000); Map<String, List<PrivilegeGrantInfo>> userPrivs = principalPrivs.getUserPrivileges();
putPersistentPrivObjects(mtbl, toPersistPrivObjs, now, userPrivs, PrincipalType.USER); Map<String, List<PrivilegeGrantInfo>> groupPrivs = principalPrivs.getGroupPrivileges();
putPersistentPrivObjects(mtbl, toPersistPrivObjs, now, groupPrivs, PrincipalType.GROUP); Map<String, List<PrivilegeGrantInfo>> rolePrivs = principalPrivs.getRolePrivileges();
putPersistentPrivObjects(mtbl, toPersistPrivObjs, now, rolePrivs, PrincipalType.ROLE);
}
pm.makePersistentAll(toPersistPrivObjs);
commited = commitTransaction();
} finally {
//如果失败则回滚
if (!commited) {
rollbackTransaction();
}
}
}
4、dropTable
二话不说上从Hive类中上代码：
public void dropTable(String tableName, boolean ifPurge) throws HiveException {
//这里Hive 将dbName与TableName合并成一个数组
String[] names = Utilities.getDbTableName(tableName);
dropTable(names[0], names[1], true, true, ifPurge);
}
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：
public static String[] getDbTableName(String dbtable) throws SemanticException {
//获取当前Session中的DbName
return getDbTableName(SessionState.get().getCurrentDatabase(), dbtable);
}
public static String[] getDbTableName(String defaultDb, String dbtable) throws SemanticException {
if (dbtable == null) {
return new String[2];
}
String[] names = dbtable.split("\\.");
switch (names.length) {
case 2:
return names;
//如果长度为1，则重新组装
case 1:
return new String [] {defaultDb, dbtable};
default:
throw new SemanticException(ErrorMsg.INVALID_TABLE_NAME, dbtable);
}
}
随后通过getMSC()调用HiveMetaStoreClient中的dropTable,代码如下：
public void dropTable(String dbname, String name, boolean deleteData,
boolean ignoreUnknownTab, EnvironmentContext envContext) throws MetaException, TException,
NoSuchObjectException, UnsupportedOperationException {
Table tbl;
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
tbl = getTable(dbname, name);
} catch (NoSuchObjectException e) {
if (!ignoreUnknownTab) {
throw e;
}
return;
}
//根据table type来判断是否为IndexTable，如果为索引表则不允许删除
if (isIndexTable(tbl)) {
throw new UnsupportedOperationException("Cannot drop index tables");
}
//这里的getHook 与create时getHook一致，获取对应table存储的hook
HiveMetaHook hook = getHook(tbl);
if (hook != null) {
hook.preDropTable(tbl);
}
boolean success = false;
try {
调用HiveMetaStore服务端的dropTable方法
drop_table_with_environment_context(dbname, name, deleteData, envContext);
if (hook != null) {
hook.commitDropTable(tbl, deleteData);
}
success=true;
} catch (NoSuchObjectException e) {
if (!ignoreUnknownTab) {
throw e;
}
} finally {
if (!success && (hook != null)) {
hook.rollbackDropTable(tbl);
}
}
}
下面我们重点看下服务端HiveMetaStore干了些什么，代码如下：
private boolean drop_table_core(final RawStore ms, final String dbname, final String name,
final boolean deleteData, final EnvironmentContext envContext,
final String indexName) throws NoSuchObjectException,
MetaException, IOException, InvalidObjectException, InvalidInputException {
boolean success = false;
boolean isExternal = false;
Path tblPath = null;
List<Path> partPaths = null;
Table tbl = null;
boolean ifPurge = false;
try {
ms.openTransaction();
// 获取正个Table的对象属性
tbl = get_table_core(dbname, name);
if (tbl == null) {
throw new NoSuchObjectException(name + " doesn't exist");
}
//如果sd数据为空，则认为该表数据损坏
if (tbl.getSd() == null) {
throw new MetaException("Table metadata is corrupted");
}
ifPurge = isMustPurge(envContext, tbl); 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
boolean isIndexTable = isIndexTable(tbl);
if (indexName == null && isIndexTable) {
throw new RuntimeException(
"The table " + name + " is an index table. Please do drop index instead.");
}
//如果不是索引表，则删除索引元数据
if (!isIndexTable) {
try {
List<Index> indexes = ms.getIndexes(dbname, name, Short.MAX_VALUE);
while (indexes != null && indexes.size() > 0) {
for (Index idx : indexes) {
this.drop_index_by_name(dbname, name, idx.getIndexName(), true);
}
indexes = ms.getIndexes(dbname, name, Short.MAX_VALUE);
}
} catch (TException e) {
throw new MetaException(e.getMessage());
}
}
//判断是否为外部表
isExternal = isExternal(tbl);
if (tbl.getSd().getLocation() != null) {
tblPath = new Path(tbl.getSd().getLocation());
if (!wh.isWritable(tblPath.getParent())) {
String target = indexName == null ? "Table" : "Index table";
throw new MetaException(target + " metadata not deleted since " +
tblPath.getParent() + " is not writable by " +
hiveConf.getUser());
}
} checkTrashPurgeCombination(tblPath, dbname + "." + name, ifPurge);
//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
partPaths = dropPartitionsAndGetLocations(ms, dbname, name, tblPath,
tbl.getPartitionKeys(), deleteData && !isExternal);
//调用ObjectStore进行meta数据的删除
if (!ms.dropTable(dbname, name)) {
String tableName = dbname + "." + name;
throw new MetaException(indexName == null ? "Unable to drop table " + tableName:
"Unable to drop index table " + tableName + " for index " + indexName);
}
success = ms.commitTransaction();
} finally {
if (!success) {
ms.rollbackTransaction();
} else if (deleteData && !isExternal) {
//删除物理partition
deletePartitionData(partPaths, ifPurge);
//删除Table路径
deleteTableData(tblPath, ifPurge);
// ok even if the data is not deleted
//Listener 处理
for (MetaStoreEventListener listener : listeners) {
DropTableEvent dropTableEvent = new DropTableEvent(tbl, success, deleteData, this);
dropTableEvent.setEnvironmentContext(envContext);
listener.onDropTable(dropTableEvent);
}
}
return success;
}
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：
public boolean dropTable(String dbName, String tableName) throws MetaException,
NoSuchObjectException, InvalidObjectException, InvalidInputException {
boolean success = false;
try {
openTransaction();
//重新获取Table对象
MTable tbl = getMTable(dbName, tableName);
pm.retrieve(tbl);
if (tbl != null) {
//下列代码查询并删除所有的权限
List<MTablePrivilege> tabGrants = listAllTableGrants(dbName, tableName);
if (tabGrants != null && tabGrants.size() > 0) {
pm.deletePersistentAll(tabGrants);
}
List<MTableColumnPrivilege> tblColGrants = listTableAllColumnGrants(dbName,
tableName);
if (tblColGrants != null && tblColGrants.size() > 0) {
pm.deletePersistentAll(tblColGrants);
} List<MPartitionPrivilege> partGrants = this.listTableAllPartitionGrants(dbName, tableName);
if (partGrants != null && partGrants.size() > 0) {
pm.deletePersistentAll(partGrants);
} List<MPartitionColumnPrivilege> partColGrants = listTableAllPartitionColumnGrants(dbName,
tableName);
if (partColGrants != null && partColGrants.size() > 0) {
pm.deletePersistentAll(partColGrants);
}
// delete column statistics if present
try {
//删除column统计表数据
deleteTableColumnStatistics(dbName, tableName, null);
} catch (NoSuchObjectException e) {
LOG.info("Found no table level column statistics associated with db " + dbName +
" table " + tableName + " record to delete");
}
//删除mcd表数据
preDropStorageDescriptor(tbl.getSd());
//删除整个Table对象相关表数据
pm.deletePersistentAll(tbl);
}
success = commitTransaction();
} finally {
if (!success) {
rollbackTransaction();
}
}
return success;
}
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：
public void alterTable(String tblName, Table newTbl, boolean cascade)
throws InvalidOperationException, HiveException {
String[] names = Utilities.getDbTableName(tblName);
try {
//删除table kv中的DDL_TIME 因为要alterTable所以，该事件会被改变
if (newTbl.getParameters() != null) {
newTbl.getParameters().remove(hive_metastoreConstants.DDL_TIME);
}
//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
newTbl.checkValidity();
//调用alterTable
getMSC().alter_table(names[0], names[1], newTbl.getTTable(), cascade);
} catch (MetaException e) {
throw new HiveException("Unable to alter table. " + e.getMessage(), e);
} catch (TException e) {
throw new HiveException("Unable to alter table. " + e.getMessage(), e);
}
}
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？
private void alter_table_core(final String dbname, final String name, final Table newTable,
final EnvironmentContext envContext, final boolean cascade)
throws InvalidOperationException, MetaException {
startFunction("alter_table", ": db=" + dbname + " tbl=" + name
+ " newtbl=" + newTable.getTableName()); //更新DDL_Time
if (newTable.getParameters() == null ||
newTable.getParameters().get(hive_metastoreConstants.DDL_TIME) == null) {
newTable.putToParameters(hive_metastoreConstants.DDL_TIME, Long.toString(System
.currentTimeMillis() / 1000));
}
boolean success = false;
Exception ex = null;
try {
//获取已有Table的整个对象
Table oldt = get_table_core(dbname, name);
//进行Event处理
firePreEvent(new PreAlterTableEvent(oldt, newTable, this));
//进行alterTable处理，后面详细说明
alterHandler.alterTable(getMS(), wh, dbname, name, newTable, cascade);
success = true;
//进行Listener处理
for (MetaStoreEventListener listener : listeners) { AlterTableEvent alterTableEvent =
new AlterTableEvent(oldt, newTable, success, this);
alterTableEvent.setEnvironmentContext(envContext);
listener.onAlterTable(alterTableEvent);
}
} catch (NoSuchObjectException e) {
// thrown when the table to be altered does not exist
ex = e;
throw new InvalidOperationException(e.getMessage());
} catch (Exception e) {
ex = e;
if (e instanceof MetaException) {
throw (MetaException) e;
} else if (e instanceof InvalidOperationException) {
throw (InvalidOperationException) e;
} else {
throw newMetaException(e);
}
} finally {
endFunction("alter_table", success, ex, name);
}
}
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：）
public void alterTable(RawStore msdb, Warehouse wh, String dbname,
String name, Table newt, boolean cascade) throws InvalidOperationException, MetaException {
if (newt == null) {
throw new InvalidOperationException("New table is invalid: " + newt);
}
//校验新的tableName是否合法
if (!MetaStoreUtils.validateName(newt.getTableName())) {
throw new InvalidOperationException(newt.getTableName()
+ " is not a valid object name");
}
//校验新的column Name type是否合法
String validate = MetaStoreUtils.validateTblColumns(newt.getSd().getCols());
if (validate != null) {
throw new InvalidOperationException("Invalid column " + validate);
} Path srcPath = null;
FileSystem srcFs = null;
Path destPath = null;
FileSystem destFs = null; boolean success = false;
boolean moveData = false;
boolean rename = false;
Table oldt = null;
List<ObjectPair<Partition, String>> altps = new ArrayList<ObjectPair<Partition, String>>(); try {
msdb.openTransaction();
//这里直接转换小写，可以看出 代码不是一个人写的
name = name.toLowerCase();
dbname = dbname.toLowerCase(); //校验新的tableName是否存在
if (!newt.getTableName().equalsIgnoreCase(name)
|| !newt.getDbName().equalsIgnoreCase(dbname)) {
if (msdb.getTable(newt.getDbName(), newt.getTableName()) != null) {
throw new InvalidOperationException("new table " + newt.getDbName()
+ "." + newt.getTableName() + " already exists");
}
rename = true;
} //获取老的table对象
oldt = msdb.getTable(dbname, name);
if (oldt == null) {
throw new InvalidOperationException("table " + newt.getDbName() + "."
+ newt.getTableName() + " doesn't exist");
}
//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
if (HiveConf.getBoolVar(hiveConf,
HiveConf.ConfVars.METASTORE_DISALLOW_INCOMPATIBLE_COL_TYPE_CHANGES,
false)) {
// Throws InvalidOperationException if the new column types are not
// compatible with the current column types.
MetaStoreUtils.throwExceptionIfIncompatibleColTypeChange(
oldt.getSd().getCols(), newt.getSd().getCols());
}
//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
if (cascade) {
//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
if(MetaStoreUtils.isCascadeNeededInAlterTable(oldt, newt)) {
//根据dbName与tableName获取整个partition的信息
List<Partition> parts = msdb.getPartitions(dbname, name, -1);
for (Partition part : parts) {
List<FieldSchema> oldCols = part.getSd().getCols();
part.getSd().setCols(newt.getSd().getCols());
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
updatePartColumnStatsForAlterColumns(msdb, part, oldPartName, part.getValues(), oldCols, part);
//更新整个Partition的信息
msdb.alterPartition(dbname, name, part.getValues(), part);
}
} else {
LOG.warn("Alter table does not cascade changes to its partitions.");
}
} //判断parititonkey是否改变，也就是dt 或 hour等partName是否改变
boolean partKeysPartiallyEqual = checkPartialPartKeysEqual(oldt.getPartitionKeys(),
newt.getPartitionKeys());
//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
if(!oldt.getTableType().equals(TableType.VIRTUAL_VIEW.toString())){
if (oldt.getPartitionKeys().size() != newt.getPartitionKeys().size()
|| !partKeysPartiallyEqual) {
throw new InvalidOperationException(
"partition keys can not be changed.");
}
} //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
if (rename
&& !oldt.getTableType().equals(TableType.VIRTUAL_VIEW.toString())
&& (oldt.getSd().getLocation().compareTo(newt.getSd().getLocation()) == 0
|| StringUtils.isEmpty(newt.getSd().getLocation()))
&& !MetaStoreUtils.isExternalTable(oldt)) {
//获取新的location信息
srcPath = new Path(oldt.getSd().getLocation());
srcFs = wh.getFs(srcPath); // that means user is asking metastore to move data to new location
// corresponding to the new name
// get new location
Database db = msdb.getDatabase(newt.getDbName());
Path databasePath = constructRenamedPath(wh.getDatabasePath(db), srcPath);
destPath = new Path(databasePath, newt.getTableName());
destFs = wh.getFs(destPath);
//设置新的table location信息 用于后续更新动作
newt.getSd().setLocation(destPath.toString());
moveData = true; //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。
if (!FileUtils.equalsFileSystem(srcFs, destFs)) {
throw new InvalidOperationException("table new location " + destPath
+ " is on a different file system than the old location "
+ srcPath + ". This operation is not supported");
}
try {
srcFs.exists(srcPath); // check that src exists and also checks
// permissions necessary
if (destFs.exists(destPath)) {
throw new InvalidOperationException("New location for this table "
+ newt.getDbName() + "." + newt.getTableName()
+ " already exists : " + destPath);
}
} catch (IOException e) {
throw new InvalidOperationException("Unable to access new location "
+ destPath + " for table " + newt.getDbName() + "."
+ newt.getTableName());
}
String oldTblLocPath = srcPath.toUri().getPath();
String newTblLocPath = destPath.toUri().getPath();
//获取old table中的所有partition信息
List<Partition> parts = msdb.getPartitions(dbname, name, -1);
for (Partition part : parts) {
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
if (oldPartLoc.contains(oldTblLocPath)) {
URI oldUri = new Path(oldPartLoc).toUri();
String newPath = oldUri.getPath().replace(oldTblLocPath, newTblLocPath);
Path newPartLocPath = new Path(oldUri.getScheme(), oldUri.getAuthority(), newPath);
altps.add(ObjectPair.create(part, part.getSd().getLocation()));
part.getSd().setLocation(newPartLocPath.toString());
String oldPartName = Warehouse.makePartName(oldt.getPartitionKeys(), part.getValues());
try {
//existing partition column stats is no longer valid, remove them
msdb.deletePartitionColumnStatistics(dbname, name, oldPartName, part.getValues(), null);
} catch (InvalidInputException iie) {
throw new InvalidOperationException("Unable to update partition stats in table rename." + iie);
}
msdb.alterPartition(dbname, name, part.getValues(), part);
}
}
//更新stats相关信息
} else if (MetaStoreUtils.requireCalStats(hiveConf, null, null, newt) &&
(newt.getPartitionKeysSize() == 0)) {
Database db = msdb.getDatabase(newt.getDbName());
// Update table stats. For partitioned table, we update stats in
// alterPartition()
MetaStoreUtils.updateUnpartitionedTableStatsFast(db, newt, wh, false, true);
}
updateTableColumnStatsForAlterTable(msdb, oldt, newt);
// now finally call alter table
msdb.alterTable(dbname, name, newt);
// commit the changes
success = msdb.commitTransaction();
} catch (InvalidObjectException e) {
LOG.debug(e);
throw new InvalidOperationException(
"Unable to change partition or table."
+ " Check metastore logs for detailed stack." + e.getMessage());
} catch (NoSuchObjectException e) {
LOG.debug(e);
throw new InvalidOperationException(
"Unable to change partition or table. Database " + dbname + " does not exist"
+ " Check metastore logs for detailed stack." + e.getMessage());
} finally {
if (!success) {
msdb.rollbackTransaction();
}
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
try {
if (srcFs.exists(srcPath) && !srcFs.rename(srcPath, destPath)) {
throw new IOException("Renaming " + srcPath + " to " + destPath + " failed");
}
} catch (IOException e) {
LOG.error("Alter Table operation for " + dbname + "." + name + " failed.", e);
boolean revertMetaDataTransaction = false;
try {
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？
msdb.alterTable(newt.getDbName(), newt.getTableName(), oldt);
for (ObjectPair<Partition, String> pair : altps) {
Partition part = pair.getFirst();
part.getSd().setLocation(pair.getSecond());
msdb.alterPartition(newt.getDbName(), name, part.getValues(), part);
}
revertMetaDataTransaction = msdb.commitTransaction();
} catch (Exception e1) {
// we should log this for manual rollback by administrator
LOG.error("Reverting metadata by HDFS operation failure failed During HDFS operation failed", e1);
LOG.error("Table " + Warehouse.getQualifiedName(newt) +
" should be renamed to " + Warehouse.getQualifiedName(oldt));
LOG.error("Table " + Warehouse.getQualifiedName(newt) +
" should have path " + srcPath);
for (ObjectPair<Partition, String> pair : altps) {
LOG.error("Partition " + Warehouse.getQualifiedName(pair.getFirst()) +
" should have path " + pair.getSecond());
}
if (!revertMetaDataTransaction) {
msdb.rollbackTransaction();
}
}
throw new InvalidOperationException("Alter Table operation for " + dbname + "." + name +
" failed to move data due to: '" + getSimpleMessage(e) + "' See hive log file for details.");
}
}
}
if (!success) {
throw new MetaException("Committing the alter table transaction was not successful.");
}
}
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：
public Partition createPartition(Table tbl, Map<String, String> partSpec) throws HiveException {
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
return new Partition(tbl, getMSC().add_partition(
Partition.createMetaPartitionObject(tbl, partSpec, null)));
} catch (Exception e) {
LOG.error(StringUtils.stringifyException(e));
throw new HiveException(e);
}
}
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：
public static org.apache.hadoop.hive.metastore.api.Partition createMetaPartitionObject(
Table tbl, Map<String, String> partSpec, Path location) throws HiveException {
List<String> pvals = new ArrayList<String>();
//遍历整个PartCols，并且校验partMap中是否一一对应
for (FieldSchema field : tbl.getPartCols()) {
String val = partSpec.get(field.getName());
if (val == null || val.isEmpty()) {
throw new HiveException("partition spec is invalid; field "
+ field.getName() + " does not exist or is empty");
}
pvals.add(val);
}
//set相关的属性信息，包括DbName、TableName、PartValues、以及sd信息
org.apache.hadoop.hive.metastore.api.Partition tpart =
new org.apache.hadoop.hive.metastore.api.Partition();
tpart.setDbName(tbl.getDbName());
tpart.setTableName(tbl.getTableName());
tpart.setValues(pvals); if (!tbl.isView()) {
tpart.setSd(cloneS d(tbl));
tpart.getSd().setLocation((location != null) ? location.toString() : null);
}
return tpart;
}
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：
private Partition add_partition_core(final RawStore ms,
final Partition part, final EnvironmentContext envContext)
throws InvalidObjectException, AlreadyExistsException, MetaException, TException {
boolean success = false;
Table tbl = null;
try {
ms.openTransaction();
//根据DbName、TableName获取整个Table对象信息
tbl = ms.getTable(part.getDbName(), part.getTableName());
if (tbl == null) {
throw new InvalidObjectException(
"Unable to add partition because table or database do not exist");
}
//事件处理
firePreEvent(new PreAddPartitionEvent(tbl, part, this));
//在创建Partition之前，首先会校验元数据中该partition是否存在
boolean shouldAdd = startAddPartition(ms, part, false);
assert shouldAdd; // start would throw if it already existed here
//创建Partition路径
boolean madeDir = createLocationForAddedPartition(tbl, part);
try {
//加载一些kv信息
initializeAddedPartition(tbl, part, madeDir);
//写入元数据
success = ms.addPartition(part);
} finally {
if (!success && madeDir) {
//如果没有成功，便删除物理路径
wh.deleteDir(new Path(part.getSd().getLocation()), true);
}
}
// we proceed only if we'd actually succeeded anyway, otherwise,
// we'd have thrown an exception
success = success && ms.commitTransaction();
} finally {
if (!success) {
ms.rollbackTransaction();
}
fireMetaStoreAddPartitionEvent(tbl, Arrays.asList(part), envContext, success);
}
return part;
}
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：
private boolean createLocationForAddedPartition(
final Table tbl, final Partition part) throws MetaException {
Path partLocation = null;
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
if (part.getSd() != null) {
partLocationStr = part.getSd().getLocation();
}
//如果为null,则重新拼接part Location
if (partLocationStr == null || partLocationStr.isEmpty()) {
// set default location if not specified and this is
// a physical table partition (not a view)
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
partLocation = new Path(tbl.getSd().getLocation(), Warehouse
.makePartName(tbl.getPartitionKeys(), part.getValues()));
}
} else {
if (tbl.getSd().getLocation() == null) {
throw new MetaException("Cannot specify location for a view partition");
}
partLocation = wh.getDnsPath(new Path(partLocationStr));
} boolean result = false;
//将location信息写入sd表
if (partLocation != null) {
part.getSd().setLocation(partLocation.toString()); // Check to see if the directory already exists before calling
// mkdirs() because if the file system is read-only, mkdirs will
// throw an exception even if the directory already exists.
if (!wh.isDir(partLocation)) {
if (!wh.mkdirs(partLocation, true)) {
throw new MetaException(partLocation
+ " is not a directory or unable to create one");
}
result = true;
}
}
return result;
}
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：
private boolean drop_partition_common(RawStore ms, String db_name, String tbl_name,
List<String> part_vals, final boolean deleteData, final EnvironmentContext envContext)
throws MetaException, NoSuchObjectException, IOException, InvalidObjectException,
InvalidInputException {
boolean success = false;
Path partPath = null;
Table tbl = null;
Partition part = null;
boolean isArchived = false;
Path archiveParentDir = null;
boolean mustPurge = false; try {
ms.openTransaction();
//根据dbName、tableName、part_values获取整个part信息
part = ms.getPartition(db_name, tbl_name, part_vals);
//获取所有Table对象
tbl = get_table_core(db_name, tbl_name);
firePreEvent(new PreDropPartitionEvent(tbl, part, deleteData, this));
mustPurge = isMustPurge(envContext, tbl); if (part == null) {
throw new NoSuchObjectException("Partition doesn't exist. "
+ part_vals);
}
//这一片还没有深入看Arrchived partition
isArchived = MetaStoreUtils.isArchived(part);
if (isArchived) {
archiveParentDir = MetaStoreUtils.getOriginalLocation(part);
verifyIsWritablePath(archiveParentDir);
checkTrashPurgeCombination(archiveParentDir, db_name + "." + tbl_name + "." + part_vals, mustPurge);
}
if (!ms.dropPartition(db_name, tbl_name, part_vals)) {
throw new MetaException("Unable to drop partition");
}
success = ms.commitTransaction();
if ((part.getSd() != null) && (part.getSd().getLocation() != null)) {
partPath = new Path(part.getSd().getLocation());
verifyIsWritablePath(partPath);
checkTrashPurgeCombination(partPath, db_name + "." + tbl_name + "." + part_vals, mustPurge);
}
} finally {
if (!success) {
ms.rollbackTransaction();
} else if (deleteData && ((partPath != null) || (archiveParentDir != null))) {
if (tbl != null && !isExternal(tbl)) {
if (mustPurge) {
LOG.info("dropPartition() will purge " + partPath + " directly, skipping trash.");
}
else {
LOG.info("dropPartition() will move " + partPath + " to trash-directory.");
}
//删除partition
// Archived partitions have har:/to_har_file as their location.
// The original directory was saved in params
if (isArchived) {
assert (archiveParentDir != null);
wh.deleteDir(archiveParentDir, true, mustPurge);
} else {
assert (partPath != null);
wh.deleteDir(partPath, true, mustPurge);
deleteParentRecursive(partPath.getParent(), part_vals.size() - 1, mustPurge);
}
// ok even if the data is not deleted
}
}
for (MetaStoreEventListener listener : listeners) {
DropPartitionEvent dropPartitionEvent =
new DropPartitionEvent(tbl, part, success, deleteData, this);
dropPartitionEvent.setEnvironmentContext(envContext);
listener.onDropPartition(dropPartitionEvent);
}
}
return true;
}
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：
private void rename_partition(final String db_name, final String tbl_name,
final List<String> part_vals, final Partition new_part,
final EnvironmentContext envContext)
throws InvalidOperationException, MetaException,
TException {
//日志记录
startTableFunction("alter_partition", db_name, tbl_name); if (LOG.isInfoEnabled()) {
LOG.info("New partition values:" + new_part.getValues());
if (part_vals != null && part_vals.size() > 0) {
LOG.info("Old Partition values:" + part_vals);
}
} Partition oldPart = null;
Exception ex = null;
try {
firePreEvent(new PreAlterPartitionEvent(db_name, tbl_name, part_vals, new_part, this));
//校验PartName的规范性
if (part_vals != null && !part_vals.isEmpty()) {
MetaStoreUtils.validatePartitionNameCharacters(new_part.getValues(),
partitionValidationPattern);
}
调用alterHandler的alterPartition进行partition物理上的rename，以及元数据修改
oldPart = alterHandler.alterPartition(getMS(), wh, db_name, tbl_name, part_vals, new_part); // Only fetch the table if we actually have a listener
Table table = null;
for (MetaStoreEventListener listener : listeners) {
if (table == null) {
table = getMS().getTable(db_name, tbl_name);
}
AlterPartitionEvent alterPartitionEvent =
new AlterPartitionEvent(oldPart, new_part, table, true, this);
alterPartitionEvent.setEnvironmentContext(envContext);
listener.onAlterPartition(alterPartitionEvent);
}
} catch (InvalidObjectException e) {
ex = e;
throw new InvalidOperationException(e.getMessage());
} catch (AlreadyExistsException e) {
ex = e;
throw new InvalidOperationException(e.getMessage());
} catch (Exception e) {
ex = e;
if (e instanceof MetaException) {
throw (MetaException) e;
} else if (e instanceof InvalidOperationException) {
throw (InvalidOperationException) e;
} else if (e instanceof TException) {
throw (TException) e;
} else {
throw newMetaException(e);
}
} finally {
endFunction("alter_partition", oldPart != null, ex, tbl_name);
}
return;
}
这里我们着重看一下，alterHandler.alterPartition方法，前方高能：
public Partition alterPartition(final RawStore msdb, Warehouse wh, final String dbname,
final String name, final List<String> part_vals, final Partition new_part)
throws InvalidOperationException, InvalidObjectException, AlreadyExistsException,
MetaException {
boolean success = false; Path srcPath = null;
Path destPath = null;
FileSystem srcFs = null;
FileSystem destFs = null;
Partition oldPart = null;
String oldPartLoc = null;
String newPartLoc = null; //修改新的partition的DDL时间
if (new_part.getParameters() == null ||
new_part.getParameters().get(hive_metastoreConstants.DDL_TIME) == null ||
Integer.parseInt(new_part.getParameters().get(hive_metastoreConstants.DDL_TIME)) == 0) {
new_part.putToParameters(hive_metastoreConstants.DDL_TIME, Long.toString(System
.currentTimeMillis() / 1000));
}
//根据dbName、tableName获取整个Table对象
Table tbl = msdb.getTable(dbname, name);
//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
if (part_vals == null || part_vals.size() == 0) {
try {
oldPart = msdb.getPartition(dbname, name, new_part.getValues());
if (MetaStoreUtils.requireCalStats(hiveConf, oldPart, new_part, tbl)) {
MetaStoreUtils.updatePartitionStatsFast(new_part, wh, false, true);
}
updatePartColumnStats(msdb, dbname, name, new_part.getValues(), new_part);
msdb.alterPartition(dbname, name, new_part.getValues(), new_part);
} catch (InvalidObjectException e) {
throw new InvalidOperationException("alter is not possible");
} catch (NoSuchObjectException e){
//old partition does not exist
throw new InvalidOperationException("alter is not possible");
}
return oldPart;
}
//rename partition
try {
msdb.openTransaction();
try {
//获取oldPart对象信息
oldPart = msdb.getPartition(dbname, name, part_vals);
} catch (NoSuchObjectException e) {
// this means there is no existing partition
throw new InvalidObjectException(
"Unable to rename partition because old partition does not exist");
}
Partition check_part = null;
try {
//组装newPart的partValues等Partition信息
check_part = msdb.getPartition(dbname, name, new_part.getValues());
} catch(NoSuchObjectException e) {
// this means there is no existing partition
check_part = null;
}
//如果check_part组装成功，说明该part已经存在，则报already exists
if (check_part != null) {
throw new AlreadyExistsException("Partition already exists:" + dbname + "." + name + "." +
new_part.getValues());
}
//table的信息校验
if (tbl == null) {
throw new InvalidObjectException(
"Unable to rename partition because table or database do not exist");
} //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
if (tbl.getTableType().equals(TableType.EXTERNAL_TABLE.toString())) {
new_part.getSd().setLocation(oldPart.getSd().getLocation());
String oldPartName = Warehouse.makePartName(tbl.getPartitionKeys(), oldPart.getValues());
try {
//existing partition column stats is no longer valid, remove
msdb.deletePartitionColumnStatistics(dbname, name, oldPartName, oldPart.getValues(), null);
} catch (NoSuchObjectException nsoe) {
//ignore
} catch (InvalidInputException iie) {
throw new InvalidOperationException("Unable to update partition stats in table rename." + iie);
}
msdb.alterPartition(dbname, name, part_vals, new_part);
} else {
try {
//获取Table的文件路径
destPath = new Path(wh.getTablePath(msdb.getDatabase(dbname), name),
Warehouse.makePartName(tbl.getPartitionKeys(), new_part.getValues()));
//拼接新的Partition的路径
destPath = constructRenamedPath(destPath, new Path(new_part.getSd().getLocation()));
} catch (NoSuchObjectException e) {
LOG.debug(e);
throw new InvalidOperationException(
"Unable to change partition or table. Database " + dbname + " does not exist"
+ " Check metastore logs for detailed stack." + e.getMessage());
}
//如果destPath不为空，说明改变了文件路径
if (destPath != null) {
newPartLoc = destPath.toString();
oldPartLoc = oldPart.getSd().getLocation();
//根据原有sd的路径获取老的part路径信息
srcPath = new Path(oldPartLoc); LOG.info("srcPath:" + oldPartLoc);
LOG.info("descPath:" + newPartLoc);
srcFs = wh.getFs(srcPath);
destFs = wh.getFs(destPath);
//查看srcFS与destFs是否Wie同一个fileSystem
if (!FileUtils.equalsFileSystem(srcFs, destFs)) {
throw new InvalidOperationException("table new location " + destPath
+ " is on a different file system than the old location "
+ srcPath + ". This operation is not supported");
}
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
srcFs.exists(srcPath); // check that src exists and also checks
if (newPartLoc.compareTo(oldPartLoc) != 0 && destFs.exists(destPath)) {
throw new InvalidOperationException("New location for this table "
+ tbl.getDbName() + "." + tbl.getTableName()
+ " already exists : " + destPath);
}
} catch (IOException e) {
throw new InvalidOperationException("Unable to access new location "
+ destPath + " for partition " + tbl.getDbName() + "."
+ tbl.getTableName() + " " + new_part.getValues());
}
new_part.getSd().setLocation(newPartLoc);
if (MetaStoreUtils.requireCalStats(hiveConf, oldPart, new_part, tbl)) {
MetaStoreUtils.updatePartitionStatsFast(new_part, wh, false, true);
}
//拼接oldPartName，并且删除原有oldPart的信息，写入新的partition信息
String oldPartName = Warehouse.makePartName(tbl.getPartitionKeys(), oldPart.getValues());
try {
//existing partition column stats is no longer valid, remove
msdb.deletePartitionColumnStatistics(dbname, name, oldPartName, oldPart.getValues(), null);
} catch (NoSuchObjectException nsoe) {
//ignore
} catch (InvalidInputException iie) {
throw new InvalidOperationException("Unable to update partition stats in table rename." + iie);
}
msdb.alterPartition(dbname, name, part_vals, new_part);
}
} success = msdb.commitTransaction();
} finally {
if (!success) {
msdb.rollbackTransaction();
}
if (success && newPartLoc != null && newPartLoc.compareTo(oldPartLoc) != 0) {
//rename the data directory
try{
if (srcFs.exists(srcPath)) {
//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
Path destParentPath = destPath.getParent();
if (!wh.mkdirs(destParentPath, true)) {
throw new IOException("Unable to create path " + destParentPath);
}
//进行原路径与目标路径的rename
wh.renameDir(srcPath, destPath, true);
LOG.info("rename done!");
}
} catch (IOException e) {
boolean revertMetaDataTransaction = false;
try {
msdb.openTransaction();
msdb.alterPartition(dbname, name, new_part.getValues(), oldPart);
revertMetaDataTransaction = msdb.commitTransaction();
} catch (Exception e1) {
LOG.error("Reverting metadata opeation failed During HDFS operation failed", e1);
if (!revertMetaDataTransaction) {
msdb.rollbackTransaction();
}
}
throw new InvalidOperationException("Unable to access old location "
+ srcPath + " for partition " + tbl.getDbName() + "."
+ tbl.getTableName() + " " + part_vals);
}
}
}
return oldPart;
}
暂时到这里吧~后续咱们慢慢玩哈~