I think I read blog post related to this problem Elixir and Postgres: A Rarely Mentioned Problem | Lainblog

• Use PostgreSQL 12, if possible
• If you use Ecto, set plan_cache_mode to force_custom_plan in the PostgreSQL config
• If you use PostgreSQL before version 12, set prepare: :unnamed in you Ecto database settings

I think that means when you are using PostgreSQL 12 you can fix this by adding parameters: [plan_cache_mode: "force_custom_plan"] to Ecto config.

Edit: fixed typo paramaters to parameters

Purely on the PG server side. There are a lot of details in the linked issue, but essentially after the fifth request using the prepared query PG’s planner decides to use a generic plan rather than a custom plan for mysterious internal reasons.

Full details are in this comment.

I’m fairly sure those are specific to PG 12 (Oban is 9.6+), and they are global settings. It would be amazing to force a particular plan when preparing the query itself. Maybe it’ll get there one day.

I ended up going with the hack I mentioned in the first post. It doesn’t look too bad, and the performance is as good or better: Avoid prepared statements for all unique queries · oban-bg/oban@756bb4f · GitHub