Liveness, and it has to mean something.

This is the ALB's health check (infra/ecs.tf), so what it returns decides whether a task keeps receiving traffic. It used to return ok unconditionally — so a task whose database connection had died stayed in the target group and kept answering requests it could not serve, and the load balancer had no way to route around it.

The query is deliberately the cheapest one there is. A health check runs every 30 seconds per task forever; anything that touches real tables would be a permanent load, and anything slow would turn a busy database into a failed deploy.

It is also cached for a few seconds, because this endpoint is unauthenticated and has to be — the ALB cannot present a credential. Probing per request meant anyone who could reach the load balancer could open a database connection per request, for free, without signing in: a health check that can be used to exhaust the pool it exists to report on. The cache makes a flood cost one SELECT 1 every _HEALTH_TTL no matter how fast it arrives, and the ALB's 30-second interval is far longer than that, so the check it gets is always freshly measured.