[HIVE-7803] Enable Hadoop speculative execution may cause corrupt output directory (dynamic partition)

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 0.13.1
• Component/s: HCatalog
• Assignee: Selina Zhang
• Reporter: Selina Zhang
• Created: 20/Aug/14 19:11
• Updated: 01/Nov/14 21:32
• Resolved: 01/Nov/14 21:32

Description

One of our users reports they see intermittent failures due to attempt directories in the input paths. We found with speculative execution turned on, two mappers tried to commit task at the same time using the same committed task path, which cause the corrupt output directory.

The original Pig script:

STORE AdvertiserDataParsedClean INTO '$DB_NAME.$ADVERTISER_META_TABLE_NAME'
USING org.apache.hcatalog.pig.HCatStorer();

Two mappers
attempt_1405021984947_5394024_m_000523_0: KILLED
attempt_1405021984947_5394024_m_000523_1: SUCCEEDED

attempt_1405021984947_5394024_m_000523_0 was killed right after the commit.

As a result, it created corrupt directory as

/projects/.../tablename/_DYN0.7192688458252056/load_time=201408050000/type=complete/_temporary/1/task_1405021984947_5394024_m_000523/
containing
part-m-00523 (from attempt_1405021984947_5394024_m_000523_0)
and
attempt_1405021984947_5394024_m_000523_1/part-m-00523

Namenode Audit log
==========================
1. 2014-08-05 05:04:36,811 INFO FSNamesystem.audit: ugi=* ip=ipaddress1 cmd=create
src=/projects/.../tablename/_DYN0.7192688458252056/load_time=201408050000/type=complete/_temporary/1/_temporary/attempt_1405021984947_5394024_m_000523_0/part-m-00523
dst=null perm=user:group:rw-r-----

2. 2014-08-05 05:04:53,112 INFO FSNamesystem.audit: ugi=* ip=ipaddress2 cmd=create
src=/projects/.../tablename/_DYN0.7192688458252056/load_time=201408050000/type=complete/_temporary/1/_temporary/attempt_1405021984947_5394024_m_000523_1/part-m-00523
dst=null perm=user:group:rw-r-----

3. 2014-08-05 05:05:13,001 INFO FSNamesystem.audit: ugi=* ip=ipaddress1 cmd=rename
src=/projects/.../tablename/_DYN0.7192688458252056/load_time=201408050000/type=complete/_temporary/1/_temporary/attempt_1405021984947_5394024_m_000523_0
dst=/projects/.../tablename/_DYN0.7192688458252056/load_time=201408050000/type=complete/_temporary/1/task_1405021984947_5394024_m_000523
perm=user:group:rwxr-x---

4. 2014-08-05 05:05:13,004 INFO FSNamesystem.audit: ugi=* ip=ipaddress2 cmd=rename
src=/projects/.../tablename/_DYN0.7192688458252056/load_time=201408050000/type=complete/_temporary/1/_temporary/attempt_1405021984947_5394024_m_000523_1
dst=/projects/.../tablename/_DYN0.7192688458252056/load_time=201408050000/type=complete/_temporary/1/task_1405021984947_5394024_m_000523
perm=user:group:rwxr-x---

After consulting our Hadoop core team, we was pointed out some HCat code does not participating in the two-phase commit protocol, for example in FileRecordWriterContainer.close():

            for (Map.Entry<String, org.apache.hadoop.mapred.OutputCommitter> entry : baseDynamicCommitters.entrySet()) {
                org.apache.hadoop.mapred.TaskAttemptContext currContext = dynamicContexts.get(entry.getKey());
                OutputCommitter baseOutputCommitter = entry.getValue();
                if (baseOutputCommitter.needsTaskCommit(currContext)) {
                    baseOutputCommitter.commitTask(currContext);
                }
            }

Issue Links

• is duplicated by: HIVE-8394 HIVE-7803 doesn't handle Pig MultiQuery, can cause data-loss.