[
https://issues.apache.org/jira/browse/SPARK-18800?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Liang-Chi Hsieh updated SPARK-18800:
------------------------------------
Description:
UnsafeKVExternalSorter uses UnsafeInMemorySorter to sort the records of
BytesToBytesMap if it is given a map.
Currently we use the number of keys in BytesToBytesMap to determine if the
array used for sort is enough or not. We has an assert that ensures the size of
the array is enough: map.numKeys() <= map.getArray().size() / 2.
However, each record in the map takes two entries in the array, one is record
pointer, another is key prefix. So the correct assert should be map.numKeys() *
2 <= map.getArray().size() / 2.
was:
UnsafeKVExternalSorter uses UnsafeInMemorySorter to sort the records of
BytesToBytesMap if it is given a map.
Currently we use the number of keys in BytesToBytesMap to determine if the
array used for sort is enough or not. It should be wrong. Because we can have
multiple values of the same key. Extremely said, you can have
BytesToBytesMap.numKeys() == 1, but BytesToBytesMap.numValues() is a big number.
In this case, we cannot just use BytesToBytesMap's array to do sorting.
Otherwise, a exception will be thrown like this:
{code}
[info] - SPARK-18800: pass BytesToBytesMap which contains numValues is more
than numKeys *** FAILED *** (61 milliseconds)
[info] java.lang.IllegalStateException: There is no space for new record
[info] at
org.apache.spark.util.collection.unsafe.sort.UnsafeInMemorySorter.insertRecord(UnsafeInMemorySorter.jav
a:225)
[info] at
org.apache.spark.sql.execution.UnsafeKVExternalSorter.<init>(UnsafeKVExternalSorter.java:147)
{code}
> Correct the assert in UnsafeKVExternalSorter which ensures array size
> ---------------------------------------------------------------------
>
> Key: SPARK-18800
> URL: https://issues.apache.org/jira/browse/SPARK-18800
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Reporter: Liang-Chi Hsieh
>
> UnsafeKVExternalSorter uses UnsafeInMemorySorter to sort the records of
> BytesToBytesMap if it is given a map.
> Currently we use the number of keys in BytesToBytesMap to determine if the
> array used for sort is enough or not. We has an assert that ensures the size
> of the array is enough: map.numKeys() <= map.getArray().size() / 2.
> However, each record in the map takes two entries in the array, one is record
> pointer, another is key prefix. So the correct assert should be map.numKeys()
> * 2 <= map.getArray().size() / 2.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
