You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@parquet.apache.org by "Wang, Gang (JIRA)" <ji...@apache.org> on 2019/02/27 07:09:00 UTC
[jira] [Updated] (PARQUET-1542) Merge multiple IO to one time IO in
method readFooter
[ https://issues.apache.org/jira/browse/PARQUET-1542?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Wang, Gang updated PARQUET-1542:
--------------------------------
Description:
Currently in method readFooter of class ParquetFileReader, it will invoke method *converter.readParquetMetadata.* To decode file metadata, there may be quite a number of times of IO processing underlying, which may heavily slow down parquet read performance.
A better way is to read all the footer bytes in a single IO, since we have already got footer length before, and pass down all the bytes to underlying*.*
And we have a test in our environment, this change may bring 50 percent improvement for some cases.
was:
Currently in method readFooter of class ParquetFileReader, it will invoke method *converter.readParquetMetadata.* To decode file metadata, there may be ** quite a number of times of IO processing underlying, which may heavily slow down parquet read performance.
A better way is to read all the footer bytes in a single IO, since we have already got footer length before, and pass down all the bytes to underlying*.*
And we have a test in our environment, this change may bring 50 percent improvement for some cases.
> Merge multiple IO to one time IO in method readFooter
> -----------------------------------------------------
>
> Key: PARQUET-1542
> URL: https://issues.apache.org/jira/browse/PARQUET-1542
> Project: Parquet
> Issue Type: Improvement
> Components: parquet-mr
> Reporter: Wang, Gang
> Priority: Major
>
> Currently in method readFooter of class ParquetFileReader, it will invoke method *converter.readParquetMetadata.* To decode file metadata, there may be quite a number of times of IO processing underlying, which may heavily slow down parquet read performance.
> A better way is to read all the footer bytes in a single IO, since we have already got footer length before, and pass down all the bytes to underlying*.*
>
> And we have a test in our environment, this change may bring 50 percent improvement for some cases.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)