hadoop-hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ashish Thusoo (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HIVE-537) Hive TypeInfo/ObjectInspector to support union (besides struct, array, and map)
Date Wed, 03 Jun 2009 18:14:07 GMT

    [ https://issues.apache.org/jira/browse/HIVE-537?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12716015#action_12716015
] 

Ashish Thusoo commented on HIVE-537:
------------------------------------

One thing that you need to be careful about is the fact that you will be increasing the number
of rows between the map and the reduce boundaries which, if there are a lot of distincts can
lead to data explosion and a subsequent slowdown in the sort.

>From that I mean the following:

Suppose we have a query with m different distincts and the base table with N rows and p mappers
and r reducers
By doing multiple map/reduce jobs, the predominant term in our complexity is

O(mN/p) + O(m(N/p log (N/p))) + O(mN/r) + O(m)

ie.
map side scan + map side sort + Reduce side merge + fixed cost of starting the map/reduce
job.

how with the current approach the corresponding formula will be

O(mN/p) + O(mN/p log (mN/p)) + O(mN/r)
=
O(mN/p) + O(mN/p log (N/p)) + O(mN/p log m) + O(mN/r)

There may be situations where one is better than the other... Something to keep in mind.


> Hive TypeInfo/ObjectInspector to support union (besides struct, array, and map)
> -------------------------------------------------------------------------------
>
>                 Key: HIVE-537
>                 URL: https://issues.apache.org/jira/browse/HIVE-537
>             Project: Hadoop Hive
>          Issue Type: New Feature
>            Reporter: Zheng Shao
>
> There are already some cases inside the code that we use heterogeneous data: JoinOperator,
and UnionOperator (in the sense that different parents can pass in records with different
ObjectInspectors).
> We currently use Operator's parentID to distinguish that. However that approach does
not extend to more complex plans that might be needed in the future.
> We will support the union type like this:
> {code}
> TypeDefinition:
>   type: primitivetype | structtype | arraytype | maptype | uniontype
>   uniontype: "union" "<" tag ":" type ("," tag ":" type)* ">"
> Example:
>   union<0:int,1:double,2:array<string>,3:struct<a:int,b:string>>
> Example of serialized data format:
>   We will first store the tag byte before we serialize the object. On deserialization,
we will first read out the tag byte, then we know what is the current type of the following
object, so we can deserialize it successfully.
> Interface for ObjectInspector:
> interface UnionObjectInspector {
>   /** Returns the array of OIs that are for each of the tags
>    */
>   ObjectInspector[] getObjectInspectors();
>   /** Return the tag of the object.
>    */
>   byte getTag(Object o);
>   /** Return the field based on the tag value associated with the Object.
>    */
>   Object getField(Object o);
> };
> {code}

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message