[SQOOP-3014] Sqoop with HCatalog import loose precision for large numbers that does not fit into double

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.4.6
• Fix Version/s: 1.5.0
• Component/s: hive-integration
• Assignee: Zoltán Tóth
• Reporter: Pavel Benes
• Created: 23/Sep/16 06:18
• Updated: 12/Sep/17 15:16
• Resolved: 12/Sep/17 15:16

Description

When using sqoop with HCatalog to import data from JDBC (I have tried Oracle11) all numbers that does not fit into double are loosing its precision or are distorted.

Steps to reproduce:

1) Create test table in Oracle and fill it with test data

CREATE TABLE TEST_SQOOP_ERROR(ID VARCHAR(10), TYPE_NUMBER DECIMAL(22,5))
 
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-1', 454018528782.42006329)
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-2', 87658675864540185.123456789123456789) 
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-3', 87658675864540185.12345)
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-4', 87658675864540185.123)
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-5', 7658675864540185.12345) 
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-6', 7658675864540185.123456789) 
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-7', 658675864540185.12345)
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-8', 58675864540185.12345)
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-9', 8675864540185.12345) 
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-10', 675864540185.12345) 
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-11', 75864540185.12345) 
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-12', 35864540185.12345) 
INSERT INTO MMDINGEST.TEST_SQOOP_ERROR(ID, TYPE_NUMBER) VALUES ( 'row-13', 5864540185.12345) 

2) Create table in Hive database

CREATE TABLE pbe_test_sqoop_error(id string, type_number decimal(22,5)) STORED AS ORC;

3) Import data from Oracle to Hive using sqoop

export HADOOP_CLASSPATH=/opt/mantis/jdbc/oracle-11.2/ojdbc6.jar
sqoop import -connect jdbc:oracle:thin:@//1.1.1.1:1521/XE --username  XXX --password  XXX --hcatalog-database  default --hcatalog-table pbe_test_sqoop_error  --driver oracle.jdbc.OracleDriver  --query 'SELECT id, type_number FROM MMDINGEST.TEST_SQOOP_ERROR WHERE $CONDITIONS' -m 1

4) Display data from Hive table

hive> select * from pbe_test_sqoop_error;
OK
row-1   454018528782.42004
row-2   87658675864540192
row-3   87658675864540192
row-4   87658675864540192
row-5   7658675864540185
row-6   7658675864540185
row-7   658675864540185.125
row-8   58675864540185.125
row-9   8675864540185.12305
row-10  675864540185.12341
row-11  75864540185.12344
row-12  35864540185.12345
row-13  5864540185.12345
Time taken: 0.455 seconds, Fetched: 13 row(s)

Only the values at line 1, 12, 13 are correct. At the lines 2-4 even the part of the number before dot is wrong. All looks correctly in Oracle as can be seen on the attached screenshot.

The problem seems to be in the java class https://www.codatlas.com/github.com/apache/sqoop/branch-1.4.6/src/java/org/apache/sqoop/mapreduce/hcat/SqoopHCatImportHelper.java starting at line 437.

    } else if (hfsType == HCatFieldSchema.Type.DECIMAL) {
      BigDecimal bd = new BigDecimal(n.doubleValue(),
        MathContext.DECIMAL128);
      return HiveDecimal.create(bd);
    }

all numbers, even those that are stored in BigDecimal are squeezed through double which leads to the precision lost The same issue could be at some places when working with large numbers.

The following code fixes this issue:

} else if (hfsType == HCatFieldSchema.Type.DECIMAL) {
      BigDecimal bd = val instanceof BigDecimal ?  (BigDecimal) val : new BigDecimal(n.doubleValue(), MathContext.DECIMAL128);
      return HiveDecimal.create(bd);
    }

.

Issue Links

• links to: Review Board