It is tempting to treat orchestration as the platform. Schedule the work, restart the containers, scale the workers, and the system looks modern.
But the hard part is still knowing what is happening. Is the workload blocked on input data, storage throughput, a hot partition, an overloaded dependency, a bad rollout, or a retry pattern that looks like useful work? Without observability, orchestration can make a broken system more efficient at hiding the reason it is broken.
For data and AI systems, I want dashboards that explain workload shape: queue depth, lag, spill, retry rate, saturation, tail latency, and cost. Automation should come after those signals are trustworthy.