Data loss after writing a transformed pyspark dataframe to delta table in unity catalog

[datastones]

Hey guys, after some successful data preprocessing without any errors, i have a final dataframe shape with the shape of ~ (200M, 150). the cluster i am using has sufficient ram + cpus + autoscaling, all metrics look fine after the job was done.

The problem that i am facing is that only approx half of my data is written to the delta table. is there any thoughts on as to why this is happening? I'd really appreciate some guidance here! here is my code snippet:

repartition_df = df.repartition(<num_cores*2; or 1>).persist() # i also tried without using df persist but no luck either, none of them works
repartition_df.write.format("delta").mode("overwrite").option("overwriteSchema", "true").saveAsTable(output_table)
chunked_df.unpersist()

[szymon_dybczak]

Hi @datastones ,

Are you 100% sure that your final df contains all the rows that you're expecting? Maybe some join condition at earlier stage causes this problem? 

[datastones]

Hey @szymon_dybczak, i appreciate your prompt response! yup, I check the dataframe shape and show some output along the way prior to writing it out. the shape looks checked out right before i perform the write... what a mystery this is!

[datastones]

1 interesting thing that I observed when playing around with different repartition size: 2,4,8,16,32,64. i have 32 cores per node. The total written result increased up to 65% of the final df at repartition = 8 (but this is very inconsistent and hard to reproduce...)

[szymon_dybczak]

Definitely weird, basically since you're writing in delta format you shouldn't lose any data. Delta tables support ACID transaction, so either it writes to your table all data or it leaves table in consistent state (before write action).
In your situation, I would run df.count() before wrtiting to delta table to double check that dataframe contains expected number of rows.
You can also get rid of persist from above code, it's useful if you're going to use the same dataframe in different set of calculation. No use for that just before writing to the table.