Repo.preload looses prefix when nested preloads are requested.

[hickscorp]

Environment

• Elixir version (elixir -v): 1.10.4
• Database and version (PostgreSQL 9.4, MongoDB 3.2, etc.): Postgres 12
• Ecto version (mix deps): 3.5.1
• Database adapter and version (mix deps): Ecto SQL 3.5
• Operating system: Xubuntu 20.04

Current behavior

We are working on an app that uses multi-tenancy.

At some point, deep in our function call tree, we're trying to preload stuff. The weirdest thing happens:

• This works: c |> Repo.preload(:agent)
• This works: c |> Repo.preload(:notifier)
• This does not work: c |> Repo.preload(:agent, :notifier) - Ecto tries to find the notifier association back in the public prefix...
• But then this works: c |> Repo.preload([:agent:, :notifier], prefix: "whatever")
• And also this works: c |> Repo.preload(:agent) |> Repo.preload(:notifier)

Expected behavior

Ecto should stay on the prefix for the structure passed on the first argument of preload.

Notes

Our Repo.default_options looks in process storage to find which prefix to work on. We are wondering whether Repo.preload is trying to either bypass default_options past one preload, or spawn other processes (which won't inherit storage from current process, and therefore won't have the prefix set when calling Repo.default_options). Which one is it? :)

Also something worth mentioning, when the bug occurs all bindings from my IEX shell are lost.

[josevalim]

Can you please fill in the issues template?

[hickscorp]

Done @josevalim

[hickscorp]

Well that was fast. Thanks @josevalim !