Hi Folks
This issue occurs on Hive 0.4 and 0.5. I wanted to wait on opening a JIRA
ticket until I ran it by the community first.
I'm testing Hive 0.5 running on Apache Hadoop 0.20.2 which is using IBM
Java 6 (32 bit x86 Java SR8 : which can be obtained here -
https://www.ibm.com/developerworks/java/jdk/linux/download.html)
To recreate this I'm using the pokes table loaded with data from the
examples directory, per the tutorial and I run the following in the Hive
CLI (bin/hive) : select count(1) from pokes;
This works just fine on Sun/Oracle Java 6, but when I change the
Hadoop-env to point to IBM Java 6 it fails in the Map with the following
exception :
Caused by: java.lang.ClassCastException:
org.apache.hadoop.hive.serde2.objectinspector.primitive.WritableIntObjectInspector
incompatible with
org.apache.hadoop.hive.serde2.objectinspector.primitive.LongObjectInspector
at
org.apache.hadoop.hive.ql.udf.generic.GenericUDAFCount$GenericUDAFCountEvaluator.merge(GenericUDAFCount.java:104)
at
org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator.aggregate(GenericUDAFEvaluator.java:113)
at
org.apache.hadoop.hive.ql.exec.GroupByOperator.updateAggregations(GroupByOperator.java:451)
at
org.apache.hadoop.hive.ql.exec.GroupByOperator.processHashAggr(GroupByOperator.java:591)
at
org.apache.hadoop.hive.ql.exec.GroupByOperator.processOp(GroupByOperator.java:500)
... 14 more
Note, the line number in GenericUDAFCount here is off by 4 based on a
couple of LOG.info calls I added for debugging purposes. The net of it is
that it is failing when it attempts to do the following cast in the merge
method:
(LongObjectInspector)inputOI
This is where it gets weird. In SUN Java, this method gets called in the
Reducer. In IBM Java, it gets called in the Mapper. If I use EXPLAIN in
the Hive CLI, the execution plans are identical regardless of which JRE is
being used in Hadoop. In SUN Java, the type for inputOI is a BigInt which
is being derived off of a single column schema called _col0_ in the
reducer (likely the output tuple of the count result) and casts to a Long
with no problem. In IBM Java, this call is happening in the Map and
inputOI is being derived off of what appears to be the first column of the
Spokes table schema, which is an int and is therefore failing when being
cats to a Long. It appears the cast is merely symptomatic of a difference
in the execution plans.
Debugging from this point, really requires someone who understands HIVE
execution plans better than I do. Is there anyone that can help with this
issue? This is really easy to replicate. Download the IBM JDK, mod your
hadoop env to point to the extracted dir of the IBM JDK and do a select
count from any table.
Regards
Steve Watt
