David Mollitor created HIVE-21466:
-------------------------------------
Summary: Increase Default Size of SPLIT_MAXSIZE
Key: HIVE-21466
URL: https://issues.apache.org/jira/browse/HIVE-21466
Project: Hive
Issue Type: Improvement
Components: Configuration
Affects Versions: 4.0.0, 3.2.0
Reporter: David Mollitor
Assignee: David Mollitor
Attachments: HIVE-21466.1.patch
{code:java}
MAPREDMAXSPLITSIZE(FileInputFormat.SPLIT_MAXSIZE, 256000000L, "", true),
{code}
[https://github.com/apache/hive/blob/8d4300a02691777fc96f33861ed27e64fed72f2c/common/src/java/org/apache/hadoop/hive/conf/HiveConf.java#L682]
This field specifies a maximum size for each MR (maybe other?) splits.
This number should be a multiple of the HDFS Block size. The way that this
maximum is implemented, is that each block is added to the split, and if the
split grows to be larger than the maximum allowed, the split is submitted to
the cluster and a new split is opened.
So, imagine the following scenario:
* HDFS block size of 16 bytes
* Maximum size of 40 bytes
This will produce a split with 3 blocks. (2x16) = 32; another block will be
inserted, (3x16) = 48 bytes in the split. So, while many operators would assume
a split of 2 blocks, the actual is 3 blocks. Setting the maximum split size to
a multiple of the HDFS block size will make this behavior less confusing.
The current setting is ~256MB and when this was introduced, the default HDFS
block size was 64MB. That is a factor of 4x. However, now HDFS block sizes are
128MB by default, so I propose setting this to 4x128MB. The larger splits
(fewer tasks) should give a nice performance boost for modern hardware.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
