利用SemanticAnalyzerHook回过滤不加分区条件的Hive查询

我们Hadoop集群中将近百分之80的作业是通过Hive来提交的，由于Hive写起来简单便捷，而且我们又提供了Hive Web Client，所以使用范围很广，包括ba，pm，po，sales都在使用hive进行ad-hoc查询，但是hive在降低用户使用门槛的同时，也使得用户经常写不合理开销很大的语句，生成了很多的mapreduce job，占用了大量slot数，其中最典型的例子就是分区表查询，不指定分区条件，导致hive没有做partition pruner优化，进而读入了所有的表数据，占用大量IO和计算资源。

为了尽可能规避这种情况，我们可以利用了hive的hook机制，在hook中实现一些方法来对语句做预判，第一期先不会直接block住语句，而是记录有问题的语句来公告警示.

具体做法是实现HiveSemanticAnalyzerHook接口，preAnalyze方法和postAnalyze方法会分别在compile函数之前和之后执行，我们只要实现preAnalyze方法，遍历传进来的ASTNode抽象语法树，获取左子树的From表名和右子树的where判断条件key值，如果该From表是分区表的话，会通过metastore client获取它的所有分区key名字，用户指定的where条件中只要出现任何一个分区key，则此语句通过检测，否则会在标准错误中输出一条warning，并且在后台log中记录用户名和执行语句，每隔一段时间会将这些bad case在hive-user组邮箱进行公示，希望能通过这种方式来起到相互警示和学习的效果.

compile函数中根据hiveconf中指定的hive.semantic.analyzer.hook来反射实例化hook类，此处为实现AbstractSemanticAnalyzerHook的DPSemanticAnalyzerHook
package org.apache.hadoop.hive.ql.parse;

import java.io.Serializable;
import java.util.ArrayList;
import java.util.List;

import org.apache.commons.lang.StringUtils;
import org.apache.hadoop.hive.metastore.api.FieldSchema;
import org.apache.hadoop.hive.ql.exec.Task;
import org.apache.hadoop.hive.ql.metadata.Hive;
import org.apache.hadoop.hive.ql.metadata.HiveException;
import org.apache.hadoop.hive.ql.metadata.Table;
import org.apache.hadoop.hive.ql.session.SessionState;
import org.apache.hadoop.hive.ql.session.SessionState.LogHelper;
import org.apache.hadoop.hive.shims.ShimLoader;

public class DPSemanticAnalyzerHook extends AbstractSemanticAnalyzerHook {
private final static String NO_PARTITION_WARNING = "WARNING: HQL is not efficient, Please specify partition condition! HQL:%s ;USERNAME:%s";

private final SessionState ss = SessionState.get();
private final LogHelper console = SessionState.getConsole();
private Hive hive = null;
private String username;
private String currentDatabase = "default";
private String hql;
private String whereHql;
private String tableAlias;
private String tableName;
private String tableDatabaseName;
private Boolean needCheckPartition = false;

@Override
public ASTNode preAnalyze(HiveSemanticAnalyzerHookContext context, ASTNode ast)
throws SemanticException {
try {
hql = ss.getCmd().toLowerCase();
hql = StringUtils.replaceChars(hql, '\n', ' ');
if (hql.contains("where")) {
whereHql = hql.substring(hql.indexOf("where"));
}
username = ShimLoader.getHadoopShims().getUserName(context.getConf());

if (ast.getToken().getType() == HiveParser.TOK_QUERY) {
try {
hive = context.getHive();
currentDatabase = hive.getCurrentDatabase();
} catch (HiveException e) {
throw new SemanticException(e);
}

extractFromClause((ASTNode) ast.getChild(0));

if (needCheckPartition && !StringUtils.isBlank(tableName)) {
String dbname = StringUtils.isEmpty(tableDatabaseName) ? currentDatabase
: tableDatabaseName;
String tbname = tableName;
String[] parts = tableName.split(".");
if (parts.length == 2) {
dbname = parts[0];
tbname = parts[1];
}
Table t = hive.getTable(dbname, tbname);
if (t.isPartitioned()) {
if (StringUtils.isBlank(whereHql)) {
console.printError(String.format(NO_PARTITION_WARNING, hql, username));
} else {
List<FieldSchema> partitionKeys = t.getPartitionKeys();
List<String> partitionNames = new ArrayList<String>();
for (int i = 0; i < partitionKeys.size(); i++) {
partitionNames.add(partitionKeys.get(i).getName().toLowerCase());
}

if (!containsPartCond(partitionNames, whereHql, tableAlias)) {
console.printError(String.format(NO_PARTITION_WARNING, hql, username));
}
}
}
}

}
} catch (Exception ex) {
ex.printStackTrace();
}
return ast;
}

private boolean containsPartCond(List<String> partitionKeys, String sql, String alias) {
for (String pk : partitionKeys) {
if (sql.contains(pk)) {
return true;
}
if (!StringUtils.isEmpty(alias) && sql.contains(alias + "." + pk)) {
return true;
}
}
return false;
}

private void extractFromClause(ASTNode ast) {
if (HiveParser.TOK_FROM == ast.getToken().getType()) {
ASTNode refNode = (ASTNode) ast.getChild(0);
if (refNode.getToken().getType() == HiveParser.TOK_TABREF && ast.getChildCount() == 1) {
ASTNode tabNameNode = (ASTNode) (refNode.getChild(0));
int refNodeChildCount = refNode.getChildCount();
if (tabNameNode.getToken().getType() == HiveParser.TOK_TABNAME) {
if (tabNameNode.getChildCount() == 2) {
tableDatabaseName = tabNameNode.getChild(0).getText().toLowerCase();
tableName = BaseSemanticAnalyzer.getUnescapedName((ASTNode) tabNameNode.getChild(1))
.toLowerCase();
} else if (tabNameNode.getChildCount() == 1) {
tableName = BaseSemanticAnalyzer.getUnescapedName((ASTNode) tabNameNode.getChild(0))
.toLowerCase();
} else {
return;
}

if (refNodeChildCount == 2) {
tableAlias = BaseSemanticAnalyzer.unescapeIdentifier(refNode.getChild(1).getText())
.toLowerCase();
}
needCheckPartition = true;
}
}
}
}

@Override
public void postAnalyze(HiveSemanticAnalyzerHookContext context,
List<Task<? extends Serializable>> rootTasks) throws SemanticException {
// LogHelper console = SessionState.getConsole();
// Set<ReadEntity> readEntitys = context.getInputs();
// console.printInfo("Total Read Entity Size:" + readEntitys.size());
// for (ReadEntity readEntity : readEntitys) {
// Partition p = readEntity.getPartition();
// Table t = readEntity.getTable();
// }
}
}