> From architectural point of view it shouldn't be returned from
> peer.init() but instead from peer.get().  It's not easy to
> implement this for peer.get() though - that's probably why
> peer.init() was used instead.

I think it's okay to do this on a per-module basis. It's like saying
"this upstream module is going to decide itself when requests get
connected to backends". I agree that returning NGX_BUSY from
peer.init() is not the best way to signal this. Perhaps, instead,
adding a flag to ngx_http_upstream_srv_conf_t would be acceptable?