Error message when using a data with 100s of cols

[ronald-smith-angel]

When a DS is large and has a big number of columns, the scan function scan.execute(scan_definition, df) fails with a spark OOM issue in the master due to the collection part of the metrics. A more meaningful message here would help to avoid miss leading the developer and let them know that the final result is too large and should be either filtered or split.

[JCZuurmond]

Hi @ronald-smith-angel , thanks for opening this issue. What I do not know yet is: how do we catch the OOM java error and how do add a test for this?

Also, we maybe should implement this in the soda-sql-spark instead of here. @vijaykiran : what do you think?

Anyway, what I expect is a try except around the scan.execute. Preferably we do this for a specific error and add a more meaningful message with a raise RuntimeError("Dataset too large, try row filtering or column selection") from e - or something alike.

@ronald-smith-angel : Would you be interested in contributing?

[stiebels]

Depending on where/how the scan is run it could also make sense to give the user the ability to optionally pass a spark.conf object to the scan.execute function, which would trigger the creation of a dedicated (potentially additional) SparkSession via the newSession method that is then used for the execution of the scan (now it's getOrCreate [here], fetching the global session or creating a vanilla new one).

This would allow the user to configure the cluster configuration (such as upping driver memory or heap space) in a way that might be more suitable to run the specific tests efficiently. I think this is one of the few common use cases for actually having two SparkSessions existing in parallel.

If this is something you'd consider making sense, I'm happy to make a PR for it.

[vijaykiran]

I have changed the logic in spark dialect when there are many columns. We are firing one query for every column to get the column metadata - now it should be just one. Please try reinstalling and/or pulling latest soda-sql-spark (2.1.5) and see if it is still a problem with too many columns.

cc @stiebels @ronald-smith-angel

@stiebels regardless of this issue, I think your suggestion makes a lot of sense. Please do open a PR when you have time 🙏🏽

[stiebels]

Thanks for the info! Great, I'll open a PR soon.