[jira] [Commented] (SPARK-31276) Contrived working example that works with multiple URI file storages for Spark cluster mode

["Jim Huang (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-31276?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17074676#comment-17074676 ] 

Jim Huang commented on SPARK-31276:
-----------------------------------

The fault is my for not framing the scenario more succinctly.  


In the scenario when spark process is launched / deployed as `client` mode in YARN cluster:
 # What is the opposite of " {{sc.parallelize(data)}}"?
 ## Using "file:///" URI, writes the file "locally" on the executors nodes' file system, not the driver's local FS.  The challenge I have is that I am unable to explicitly specify the "file:///" URI to differentiate the driver vs executor.  

 

I am aware this is an possible corner case that can be exploited unintentionally when that dataset is too big and overflow the driver memory.  But it is also a valid use case if an user just want to store some summary statistics after some big data set processing on the driver local file system.  

 

> Contrived working example that works with multiple URI file storages for Spark cluster mode
> -------------------------------------------------------------------------------------------
>
>                 Key: SPARK-31276
>                 URL: https://issues.apache.org/jira/browse/SPARK-31276
>             Project: Spark
>          Issue Type: Wish
>          Components: Examples
>    Affects Versions: 2.4.5
>            Reporter: Jim Huang
>            Priority: Major
>
> This Spark SQL Guide --> Data sources --> Generic Load/Save Functions
> [https://spark.apache.org/docs/latest/sql-data-sources-load-save-functions.html]
> described a very simple "local file system load of an example file".  
>  
> I am looking for an example that demonstrates a workflow that exercises different file systems.  For example, 
>  # Driver loads an input file from local file system
>  # Add a simple column using lit() and stores that DataFrame in cluster mode to HDFS
>  # Write that a small limited subset of that DataFrame back to Driver's local file system.  (This is to avoid the anti-pattern of writing large file and out of the scope for this example.  The small limited DataFrame would be some basic statistics, not the actual complete dataset.)
>  
> The examples I found on the internet only uses simple paths without the explicit URI prefixes.
> Without the explicit URI prefixes, the "filepath" inherits how Spark (mode) was called, local stand alone vs YARN client mode.   So a "filepath" will be read/write locally (file system) vs cluster mode HDFS, without these explicit URIs.
> There are situations were a Spark program needs to deal with both local file system and YARN client mode (big data) in the same Spark application, like producing a summary table stored on the local file system of the driver at the end.  
> If there are any existing alternatives Spark documentation that provides examples that traverse through the different URIs in Spark YARN client mode or a better or smarter Spark pattern or API that is more suited for this, I am happy to accept that as well.  Thanks!



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org