Google Cloud ACE: Ensuring successful operation Practice Questions

The two correct answers are: (1) The pod has a NodeSelector or affinity rule that doesn't match any available nodes (pods won't schedule if their affinity constraints exclude available nodes), and (2) The GKE Autopilot cluster's quota limit has been reached and Autopilot cannot provision additional nodes (in Autopilot, even with apparent capacity, quota limits can prevent node creation). Option 0 is wrong because the user reported available node capacity, so resource constraints wouldn't typically cause pending state. Option 3 is wrong because image availability doesn't block scheduling; the image is pulled after a pod is scheduled, so missing images delay pod startup, not scheduling. Option 4 is wrong because PersistentVolume unavailability would cause pod creation errors, not scheduling delays; pods would typically fail to mount rather than remain pending indefinitely.

The most important consideration is to quiesce the disk—stop writes and flush the cache—to ensure the snapshot captures a consistent state, especially for databases. This prevents data corruption in the snapshot. Option 1 is the best answer because it directly addresses data integrity. Option 2 is incorrect; you don't need to unmount the disk to snapshot it; Compute Engine can snapshot mounted disks, and unmounting a database disk is risky. Option 3 is a practical consideration but not the 'most important' from a data integrity perspective. Option 4 is false—public IP is not required for snapshotting; it's an internal operation.

The correct answer is to enable read replicas and configure the application to route read queries to replicas. Since queries are already optimized, the issue is likely read-heavy workload distribution. Read replicas offload read traffic without requiring instance upscaling. Option 0 is wrong because scaling up immediately increases costs; you should first consider distributing the load. Option 2 is wrong because storage configuration and SSD disks improve I/O latency but don't address CPU-bound compute workloads. Option 3 is wrong because database clones are point-in-time copies for backup/dev purposes, not for distributing live production workloads; that requires replication.

Option 3 is correct. When creating a machine image from an instance, a reboot occurs. However, systemd (or init systems) will automatically restart services configured with 'Restart=always' or similar policies, minimizing application downtime. Option 0 is partially true (shutdown occurs) but ignores automatic restart behavior. Option 1 is true about the reboot but doesn't account for automatic restarts. Option 2 is false—a reboot does occur during image creation. The key insight is understanding how systemd handles service restarts after a reboot.

Recreating the VPC connector (option 2) will require re-establishing the connection and may necessitate instance-level changes or validation. Option 0 is incorrect—encryption key rotation in Cloud Spanner is handled transparently and doesn't require a restart. Option 1 is incorrect—scaling node count is a non-disruptive operation in Cloud Spanner. Option 3 is incorrect—IAM permission updates don't require instance changes; they apply immediately. The key distinction is that VPC connector recreation is infrastructure-level and can affect instance connectivity.