In pyspark, why is writing to elastic search longer and longer for the same number of documents in the same job?

Question

Context

I have a python spark job transforming a dataframe into json and writing it to elastic search. The time to write the (almost) identical numbe