[
https://issues.apache.org/jira/browse/ORC-743?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dongjoon Hyun closed ORC-743.
-----------------------------
> Conversion of SArg into Filters, to take advantage of LazyIO
> ------------------------------------------------------------
>
> Key: ORC-743
> URL: https://issues.apache.org/jira/browse/ORC-743
> Project: ORC
> Issue Type: Sub-task
> Components: Reader
> Reporter: Pavan Lanka
> Assignee: Pavan Lanka
> Priority: Major
> Fix For: 1.7.0
>
>
> ORC-742 introduces lazy evaluation of the non-filter columns in the presence
> of filters. This builds further on that to convert SArg into filters.
> h3. SArg to Filter
> SArg to Filter converts the passed SArg into a filter. This enables automatic
> compatibility with both Spark and Hive as they already push down Search
> Arguments down to ORC.
> The SArg is automatically converted into a Vector Filter. Which is applied
> during the read process.
> The builder for search argument should allow skipping normalization during
> the
> [build|https://github.com/apache/hive/blob/storage-branch-2.7/storage-api/src/java/org/apache/hadoop/hive/ql/io/sarg/SearchArgumentImpl.java#L491].
> This has already been proposed as part of HIVE-24458.
> Normalization is very poor in performance in the presence of multilevel
> predicates.
> ||Benchmark||(fSize)||(fType)||(normalize)||Mode||Cnt||Score||Error||Units||
> |ComplexFilterBench.filter|2|vector|true|avgt|20|74.321|± 0.156|us/op|
> |ComplexFilterBench.filter|2|vector|false|avgt|20|78.119|± 0.351|us/op|
> |ComplexFilterBench.filter|4|vector|true|avgt|20|267.405|± 1.202|us/op|
> |ComplexFilterBench.filter|4|vector|false|avgt|20|136.284|± 0.637|us/op|
> |ComplexFilterBench.filter|8|vector|true|avgt|20|9907.765|± 49.208|us/op|
> |ComplexFilterBench.filter|8|vector|false|avgt|20|247.714|± 0.651|us/op|
> Explanation:
> * *fSize* identifies the size of the OR clause that will be normalized.
> * *normalize* identifies whether normalize was carried out on the Search
> Argument.
> Observations:
> * Normalizing the search argument results in a significant performance
> penalty given the explosion of the operator tree
> ** In case where an AND includes 8 ORs, the unnormalized version is faster
> by *97.32%*
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
