Question about large windows and RocksDB state backend

[Gabriele Mencagli <ga...@gmail.com>]

Dear Flink Community,

I am working with an application developed using the DataStream API. The 
application has a window operator using with a large temporal window. 
Due to the processing done, we are using the ProcessWindowFunction 
implementation as we need to explicitly buffer all the tuples falling 
within the boundaries of each triggered window.

We are interested in running the application with the RocksDB embedded 
state backend to reduce the memory footprint. We would like to know if 
there is information available about how windows are stored and 
represented in RocksDB, whether windows represent separate objects in 
RocksDB, if tuples are grouped by key, replicated in case of overlapping 
windows, and so forth.

We understand that this question can be answered by manually inspecting 
the source code. However, we would be very grateful if someone could 
share their knowledge on this topic and suggest any relevant documents 
available online. Unfortunately, the documentation on Flink's website 
does not provide such low-level details.

Thanks a lot,

Gabriele