[
https://issues.apache.org/jira/browse/DRILL-2743?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14707029#comment-14707029
]
Aman Sinha commented on DRILL-2743:
-----------------------------------
I added specific comments in the pull request. A few broader review comments
below:
1. The first query that does not see the metadata file will gather the
metadata, so
the elapsed time of the first query will be very different from a
subsequent query
. We should probably document that.
2. How do we prevent concurrent REFRESH METADATA operations on the
same table or subdirectory since they will be updating the same file ?
3. Agree with Rahul's comment for log message and Jacques's comment on
versioning..these will be very useful. I also foresee that people might
want to
point to a specific metadata file and not be restricted to the hardcoded
name.
This could be an enhancement.
> Parquet file metadata caching
> -----------------------------
>
> Key: DRILL-2743
> URL: https://issues.apache.org/jira/browse/DRILL-2743
> Project: Apache Drill
> Issue Type: New Feature
> Components: Storage - Parquet
> Reporter: Steven Phillips
> Assignee: Aman Sinha
> Fix For: 1.2.0
>
> Attachments: DRILL-2743.patch, drill.parquet_metadata
>
>
> To run a query against parquet files, we have to first recursively search the
> directory tree for all of the files, get the block locations for each file,
> and read the footer from each file, and this is done during the planning
> phase. When there are many files, this can result in a very large delay in
> running the query, and it does not scale.
> However, there isn't really any need to read the footers during planning, if
> we instead treat each parquet file as a single work unit, all we need to know
> are the block locations for the file, the number of rows, and the columns. We
> should store only the information which we need for planning in a file
> located in the top directory for a given parquet table, and then we can delay
> reading of the footers until execution time, which can be done in parallel.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
