Hive MetadataOnlyOptimizer NPE

昨天同事在Web查询UI发现一个hivequery不能跑,SQL类似如下：

select c
from
(select * from db1.t1
where partkey1 = 'xxx') a
left  outer join db2.t2 b on (a.c1 = b.c1)
left  outer join db2.t2 c on (a.c2 = c.c2)
group by c

报NPE，尝试下发现连explain都不行，说明在解析阶段就出问题了，查看日志：

atorg.apache.hadoop.hive.ql.optimizer.physical.MetadataOnlyOptimizer$MetadataOnlyTaskDispatcher.dispatch(MetadataOnlyOptimizer.java:308)
        atorg.apache.hadoop.hive.ql.lib.DefaultGraphWalker.dispatch(DefaultGraphWalker.java:87)
        atorg.apache.hadoop.hive.ql.lib.DefaultGraphWalker.walk(DefaultGraphWalker.java:124)
        atorg.apache.hadoop.hive.ql.lib.DefaultGraphWalker.startWalking(DefaultGraphWalker.java:101)
        atorg.apache.hadoop.hive.ql.optimizer.physical.MetadataOnlyOptimizer.resolve(MetadataOnlyOptimizer.java:175)
        atorg.apache.hadoop.hive.ql.optimizer.physical.PhysicalOptimizer.optimize(PhysicalOptimizer.java:79)
        at org.apache.hadoop.hive.ql.parse.SemanticAnalyzer.genMapRedTasks(SemanticAnalyzer.java:8399)
        atorg.apache.hadoop.hive.ql.parse.SemanticAnalyzer.analyzeInternal(SemanticAnalyzer.java:8741)
        atorg.apache.hadoop.hive.ql.parse.BaseSemanticAnalyzer.analyze(BaseSemanticAnalyzer.java:278)
        atorg.apache.hadoop.hive.ql.Driver.compile(Driver.java:433)
        atorg.apache.hadoop.hive.ql.Driver.compile(Driver.java:337)
        atorg.apache.hadoop.hive.ql.Driver.run(Driver.java:902)
        at org.apache.hadoop.hive.cli.CliDriver.processLocalCmd(CliDriver.java:259)
        atorg.apache.hadoop.hive.cli.CliDriver.processCmd(CliDriver.java:216)
        atorg.apache.hadoop.hive.cli.CliDriver.processLine(CliDriver.java:413)
        at org.apache.hadoop.hive.cli.CliDriver.run(CliDriver.java:756)
        atorg.apache.hadoop.hive.cli.CliDriver.main(CliDriver.java:614)
        atsun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
        atsun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:39)
        atsun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:25)
        atjava.lang.reflect.Method.invoke(Method.java:597)
        atorg.apache.hadoop.util.RunJar.main(RunJar.java:208)

是在物理优化阶段MetadataOnlyOptimizer出了问题，具体到NPE的部分是：

while(iterator.hasNext()) {
        TableScanOperator tso =iterator.next();
LOG.info("Metadata only table scan for "+ tso.getConf().getAlias());
        convertToMetadataOnlyQuery((MapredWork)task.getWork(), tso);
      }

该处是有一个Bug（HIVE-4935）的，按道理直接关闭hive.optimize.metadataonly就可以简单workaround，不过要弄清楚为什么0.11之前的版本不会触发，因为metadataonly 0.10之前就引入了.

查看Pla才知道是和0.11引入的mapjoin merge的优化有关系：

TS1->MapJoin1->TS2->MapJoin2->

由于默认开启了mapjoinmerge，并且都是noconditional的task，TSO是不需要任何column因此conf为空，进而触发了前面的Bug，这个触发本身也作为一个Bug（HIVE-4927）在0.12得到fix，而work around这个Bug也简单：

关闭hive.auto.convert.join.noconditionaltask，因此解决这个问题可以有两种方式来选择.

PS：这个hive.auto.convert.join.noconditionaltask开启后，启动的TASK是被标记为MAPJOIN_ONLY_NOBACKUP：The join task is converted to a mapjoin task. This can only happen if hive.auto.convert.join.noconditionaltaskis set to true.No conditional task was created in case the mapjoin failed.

也就是说这个这个特性启动的task是不会有backup task的，如果某个mapredlocaltask fail的话整个Job也就Failed了.