On highly concurrent PG systems by when using parallel queries you are sacrificing throughput for better latency. You really don't want to use more than a fairly small number of workers per single select.