【原创】大叔问题定位分享(22)hive同时执行多个insert overwrite table只有1个可以执行

时间:2023-12-09 23:39:37

hive 2.1

一 问题

最近有一个场景,要向一个表的多个分区写数据,为了缩短执行时间,采用并发的方式,多个sql同时执行,分别写不同的分区,同时开启动态分区:

set hive.exec.dynamic.partition=true

insert overwrite table test_table partition(dt) select * from test_table_another where dt = 1;

结果发现只有1个sql运行,其他sql都会卡住;
查看hive thrift server线程堆栈发现请求都卡在DbTxnManager上,hive关键配置如下:

hive.support.concurrency=true
hive.txn.manager=org.apache.hadoop.hive.ql.lockmgr.DbTxnManager

配置对应的默认值及注释:

org.apache.hadoop.hive.conf.HiveConf

    HIVE_SUPPORT_CONCURRENCY("hive.support.concurrency", false,
"Whether Hive supports concurrency control or not. \n" +
"A ZooKeeper instance must be up and running when using zookeeper Hive lock manager "), HIVE_TXN_MANAGER("hive.txn.manager",
"org.apache.hadoop.hive.ql.lockmgr.DummyTxnManager",
"Set to org.apache.hadoop.hive.ql.lockmgr.DbTxnManager as part of turning on Hive\n" +
"transactions, which also requires appropriate settings for hive.compactor.initiator.on,\n" +
"hive.compactor.worker.threads, hive.support.concurrency (true), hive.enforce.bucketing\n" +
"(true), and hive.exec.dynamic.partition.mode (nonstrict).\n" +
"The default DummyTxnManager replicates pre-Hive-0.13 behavior and provides\n" +
"no transactions."),

二 代码分析

hive执行sql的详细过程详见:https://www.cnblogs.com/barneywill/p/10185168.html

hive中执行sql最终都会调用到Driver.run,run会调用runInternal,下面直接看runInternal代码:

org.apache.hadoop.hive.ql.Driver

  private CommandProcessorResponse runInternal(String command, boolean alreadyCompiled)
throws CommandNeedRetryException {
...
if (requiresLock()) {
// a checkpoint to see if the thread is interrupted or not before an expensive operation
if (isInterrupted()) {
ret = handleInterruption("at acquiring the lock.");
} else {
ret = acquireLocksAndOpenTxn(startTxnImplicitly);
}
... private boolean requiresLock() {
if (!checkConcurrency()) {
return false;
}
// Lock operations themselves don't require the lock.
if (isExplicitLockOperation()){
return false;
}
if (!HiveConf.getBoolVar(conf, ConfVars.HIVE_LOCK_MAPRED_ONLY)) {
return true;
}
Queue<Task<? extends Serializable>> taskQueue = new LinkedList<Task<? extends Serializable>>();
taskQueue.addAll(plan.getRootTasks());
while (taskQueue.peek() != null) {
Task<? extends Serializable> tsk = taskQueue.remove();
if (tsk.requireLock()) {
return true;
}
... private boolean checkConcurrency() {
boolean supportConcurrency = conf.getBoolVar(HiveConf.ConfVars.HIVE_SUPPORT_CONCURRENCY);
if (!supportConcurrency) {
LOG.info("Concurrency mode is disabled, not creating a lock manager");
return false;
}
return true;
} private int acquireLocksAndOpenTxn(boolean startTxnImplicitly) {
...
txnMgr.acquireLocks(plan, ctx, userFromUGI);
...

runInternal会调用requiresLock判断是否需要lock,requiresLock有两个判断:

  • 调用checkConcurrency,checkConcurrency会检查hive.support.concurrency=true才需要lock;
  • 调用Task.requireLock,只有部分task才需要lock;

如果判断需要lock,会调用acquireLocksAndOpenTxn,acquireLocksAndOpenTxn会调用HiveTxnManager.acquireLocks来获取lock;

1)先看那些task需要lock:

org.apache.hadoop.hive.ql.parse.DDLSemanticAnalyzer

  private void analyzeAlterTablePartMergeFiles(ASTNode ast,
String tableName, HashMap<String, String> partSpec)
throws SemanticException {
...
DDLWork ddlWork = new DDLWork(getInputs(), getOutputs(), mergeDesc);
ddlWork.setNeedLock(true);
...

可见DDL操作需要;

2)再看怎样获取lock:

org.apache.hadoop.hive.ql.lockmgr.DbTxnManager

  public void acquireLocks(QueryPlan plan, Context ctx, String username) throws LockException {
try {
acquireLocksWithHeartbeatDelay(plan, ctx, username, 0);
... void acquireLocksWithHeartbeatDelay(QueryPlan plan, Context ctx, String username, long delay) throws LockException {
LockState ls = acquireLocks(plan, ctx, username, true);
... LockState acquireLocks(QueryPlan plan, Context ctx, String username, boolean isBlocking) throws LockException {
...
switch (output.getType()) {
case DATABASE:
compBuilder.setDbName(output.getDatabase().getName());
break; case TABLE:
case DUMMYPARTITION: // in case of dynamic partitioning lock the table
t = output.getTable();
compBuilder.setDbName(t.getDbName());
compBuilder.setTableName(t.getTableName());
break; case PARTITION:
compBuilder.setPartitionName(output.getPartition().getName());
t = output.getPartition().getTable();
compBuilder.setDbName(t.getDbName());
compBuilder.setTableName(t.getTableName());
break; default:
// This is a file or something we don't hold locks for.
continue;
}
...
LockState lockState = lockMgr.lock(rqstBuilder.build(), queryId, isBlocking, locks);
ctx.setHiveLocks(locks);
return lockState;
}

可见当开启动态分区时,锁的粒度是DbName+TableName,这样就会导致多个sql只有1个sql可以拿到lock,其他sql只能等待;

三 总结

解决问题的方式有几种:

  1. 关闭动态分区:set hive.exec.dynamic.partition=false
  2. 关闭并发:set hive.support.concurrency=false
  3. 关闭事务:set hive.txn.manager=org.apache.hadoop.hive.ql.lockmgr.DummyTxnManager

三者任选其一,推荐第1种,因为在刚才的场景下,不需要动态分区;