[ 
https://issues.apache.org/jira/browse/TRAFODION-2949?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16416027#comment-16416027
 ] 

Sandhya Sundaresan commented on TRAFODION-2949:
-----------------------------------------------

Making a few fixes in this area.
 # LOB internal descriptor tables were treated as "special" tables and 
prevented caching of NATable information. So ever insert into a lob column that 
needed to update the LOB internal descriptor tables had to be reloaded each 
time. Removing that restriction. This signinifantyl improves insert/update 
operations. 
 #  Fixed an issue where only one type of internal table was detected as LOB 
table and caused query cache to not cache statements related to these tables. 
LOBDescChunks was not recognized. Added that.  
 # Made the units of the LOB Max size , LOB size consistent in  the compiletime 
and runtime structures.Now Describe returns the units  in bytes. 
 # Disabled doing GC during DML operations by default. There will be a separate 
utility for doing this or it can be enabled by setting CQD LOB_GC_LIMIT_SIZE to 
something other than -1. 
 # Increased the rowset size limit  checks  to allow for batch operations to  
support input rows as large as 16MB. 

> LOB: Address performance issues related to inserts in tables containing LOB 
> columns
> -----------------------------------------------------------------------------------
>
>                 Key: TRAFODION-2949
>                 URL: https://issues.apache.org/jira/browse/TRAFODION-2949
>             Project: Apache Trafodion
>          Issue Type: Improvement
>            Reporter: Sandhya Sundaresan
>            Assignee: Sandhya Sundaresan
>            Priority: Major
>
> # Investigate bottlenecks in the insert and update code paths.
> A couple of bottlenecks identified initially were hdfsWrite and the LOBCLi 
> call.
>  # Investigate if lib hdfs write operation is signifincantly slower than a 
> direct JNI call. If so  consider refactoring code being done for hive inserts 
> and reuse for LOB writes as well. Also investigate use of libhdfs3 , a 
> lightweight interface by Pivotal for Hawk ,that doesn’t use JNI at all.
>  # Investigate if LOB Cli calls that do prepare and execute of statements to 
> insert/update the LOB descriptor tables are being cached and reused when 
> multiple LOB inserts are being done in the same session.
>  # Investigate external LOBs as a solution for faster inserts/extracts for 
> usage compared to internal LOBs.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to