You manage a Dataproc cluster and need faster job completion while minimizing cost and preserving in-progress work. What change do you make?

Graceful decommissioning allows preemptible workers to finish their current tasks before being shut down, preserving in-progress work and preventing data loss. This balances cost savings (from using preemptible VMs) with job completion reliability. Increasing cluster size with non-preemptible workers would be more expensive. Forcibly decommissioning preemptible workers would interrupt jobs and lose work. While Cloud Monitoring (formerly Stackdriver) can trigger scripts, relying on external scripting for graceful shutdown is less integrated and more complex than Dataproc's built-in graceful decommissioning feature.