[
https://issues.apache.org/jira/browse/PIG-955?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12754319#action_12754319
]
Ying He commented on PIG-955:
-----------------------------
the sampling process generated a file which contains skewed keys and their
pre-allocated reducer indexes. Each (key, beginning index, ending index) is
stored as a tuple.
during join process, this file is loaded by SkewedPartitioner as lookup table.
For tuples from partition table, its key is matched against this lookup table,
if match is found, it returns a value in range of [beginning index, ending
index] in round robin fashion. If no match found, it then use hash() to
calculate index.
the problem is in SkewedPartitioner, when looking up the table, the
PigNullableWritable format of input tuple is used, while the lookup table uses
Pig type Tuple as keys. Therefore, no match is found. The indexes are
calculated using hash() even for skewed keys. This causes the data for this
key all goes to the same reducer.
But for streaming table, if key is skewed key, each tuple is replicated to
each reducer that are pre-allocated during sampling process.
Because the reducer indexes are calculated wrong for skewed keys in partition
table, tuples from first table are sent to wrong reducers, if it doesn't fall
into its pre-calculated index range, the join with second table ends up with
empty data set for that key. The query still appears successfully, but it has
data loss.
The fix is to change SkewedPartitioner to use correct object type to lookup
skewed key tables
> Skewed join generates incorrect results
> -----------------------------------------
>
> Key: PIG-955
> URL: https://issues.apache.org/jira/browse/PIG-955
> Project: Pig
> Issue Type: Improvement
> Reporter: Ying He
> Attachments: PIG-955.patch
>
>
> Fragmented replicated join has a few limitations:
> - One of the tables needs to be loaded into memory
> - Join is limited to two tables
> Skewed join partitions the table and joins the records in the reduce phase.
> It computes a histogram of the key space to account for skewing in the input
> records. Further, it adjusts the number of reducers depending on the key
> distribution.
> We need to implement the skewed join in pig.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
