A nightly batch job is exhausting database connections and the API is timing out. How do you fix it?

Stop the immediate harm by throttling or pausing the batch job, then separate the workloads so they cannot compete. Point the batch job at a read replica if it is read heavy, give it its own connection pool with a hard ceiling, and chunk the work with pauses so it never monopolizes the database. Longer term, reserve capacity for interactive traffic and alert on connection pool saturation before users notice.

First, stop the bleeding. Pause or throttle the batch job, because the interactive path is what customers feel and the batch can run an hour later without anyone caring. The tempting fix is to raise the maximum connections, and that usually makes it worse, since every connection has memory and scheduling cost and you end up with a database thrashing instead of one that is merely busy. So the real fix is isolation. If the job is read heavy it moves to a read replica, which takes the load off the primary entirely. It gets its own credentials and its own pool with a hard ceiling, sized so that even fully saturated it leaves plenty of headroom for the API. And the job itself gets chunked, so it processes in batches with a short pause and a statement timeout, rather than one enormous transaction that holds locks and blows out replication lag. Then prevention: alert on pool utilization and long running queries at a threshold that fires before it becomes user visible, so we hear about it from monitoring rather than from support tickets.