Error in hadoop jobs due to hive query error

The solution from AWS experts is to

  1. git clone open source emr-dynamodb-connector
  2. modify the code
  3. prepare your own Jar
  4. Using bootstrapper upload it to EMR
  5. In run_job_flow , send configurations for hadoop env appending your own location of jars in HADOOP_CLASSPATH.

Being not so much into Java, modifying emr-dynamodb-connector was not possible for me, but this is the solution. Also two things can be done... if you don't use Strings in Dynamodb, map string of hive to number of Dynamodb, else add mapping and support for decimal from hive to Dynamodb number


Is this number too big to be converted to int?

Yes, this number is too big to convert to an integral type. According to the Apache Hive documentation on Numeric Types, the maximum value for a BIGINT is 9223372036854775807. Your input, 17664956244983174066, is larger than that.

The following is a vanilla Hive query (no DynamoDB integration) demonstrating the effects of attempting to convert various inputs to a BIGINT.

SELECT
    "9223372036854775807" AS str,
    cast("9223372036854775807" AS BIGINT) AS numbigint,
    cast("9223372036854775807" AS DOUBLE) AS numdouble
UNION ALL
SELECT
    "9223372036854775808" AS str,
    cast("9223372036854775808" AS BIGINT) AS numbigint,
    cast("9223372036854775808" AS DOUBLE) AS numdouble
UNION ALL
SELECT
    "17664956244983174066" AS str,
    cast("17664956244983174066" AS BIGINT) AS numbigint,
    cast("17664956244983174066" AS DOUBLE) AS numdouble
;

    str numbigint   numdouble
0   9223372036854775807 9223372036854775807 9.2233720368547758e+18
1   9223372036854775808 NULL    9.2233720368547758e+18
2   17664956244983174066    NULL    1.7664956244983173e+19

At the documented maximum value of BIGINT, the value converts correctly. At just 1 higher, the conversion fails, resulting in NULL. The same thing happens for your input.

The query also demonstrates that conversion to DOUBLE is successful. Perhaps that's a solution, depending on your use case. Compared to an integral data type, this would open a risk of encountering floating point precision issues.

From your stack trace, it appears that the DynamoDB integration results in a NumberFormatException for this case rather than NULL. This is arguably a bug in the DynamoDB connector, but even if it were changed to map to NULL, you still wouldn't have a successful conversion.