Precisely this. In theory, there will never be unexpected latency to the database, so you don't need to worry about overwhelming it and can just immediately retry. In theory, your instance's CPU won't throttle due to a shared tenant on the host running an AVX-512 instruction, so the profiling you did in staging is valid and can be trusted.