I would say you might be better off using a connection pooler if you need this many connections.
Yeah that's a lot. pgbouncer might be a good option in your case.
work_mem = 1024MB
work_mem is pretty high. It would make sense in a data warehouse-type environment, but with a max of 900 connections, that can get used up in a hurry. Do you find your queries regularly spilling sorts to disk (something like "External merge Disk" in your EXPLAIN ANALYZE plans)?
work_mem is a per-operation setting for sort/hash operations. So in your case you might finish with a maximum of 900GB of memory allocated based on the maximum number of sessions that can run in parallel on your server. Simply reduce the value of work_mem to something your server can manage and you should be able to solve your problems of OOM. --
