Skip to content

Professional Cloud DevOps Engineer · Domain 1: Establishing and bootstrapping foundational infrastructure and organizational governance · 18% of the exam

Task 1.4: Managing multiple runtime environments (e.g., staging, production).

Ephemeral environments, configuration and policy across environments, GKE fleets, serverless environments, safe patching and upgrades, and isolating agent execution with GKE Agent Sandbox.

Study it

Sample questions

Try each one before opening the answer. Every option is explained, with the Google Cloud documentation page that proves it.

Question 1 · choose 1

A retailer's production checkout runs on a regional GKE Standard cluster in the Regular release channel. For six weeks over the holidays, no Pod may be evicted or rescheduled because of a GKE upgrade, but the security team still wants the cluster control plane to receive patch versions during that time. What should the platform team configure for the six weeks?

  1. AA maintenance exclusion with the No upgrades scope
  2. BA maintenance exclusion with the No minor upgrades scope
  3. CA weekly maintenance window from 02:00 to 06:00 on Sundays
  4. DA maintenance exclusion with the No minor or node upgrades scope
Show the answer and why
  • AA maintenance exclusion with the No upgrades scope

    Incorrect

    This default scope blocks every automatic upgrade, including control plane patch upgrades, which the security team wants to keep.

  • BA maintenance exclusion with the No minor upgrades scope

    Incorrect

    This scope still allows automatic patch upgrades on the nodes, and a node upgrade restarts node VMs and reschedules their Pods.

  • CA weekly maintenance window from 02:00 to 06:00 on Sundays

    Incorrect

    A window only decides when automatic maintenance may run; node upgrades would still happen inside it during the six weeks.

  • DA maintenance exclusion with the No minor or node upgrades scope

    Correct

    This scope blocks minor upgrades and all node upgrades, while control plane patch upgrades stay allowed.

GKE maintenance exclusion scopes differ in what they still allow. No upgrades blocks everything (for at most 90 days); No minor upgrades still patches the control plane and nodes; No minor or node upgrades patches only the control plane, which avoids the evictions and rescheduling that node upgrades cause.

Question 2 · choose 1

A coding assistant on GKE Autopilot runs Python that its LLM writes. Each user session gets its own Pod, isolated from the node's kernel, which must be ready in under a second when the session starts; an idle session must be suspendable without losing its persistent volume. The team wants to manage all of this with Kubernetes APIs. Which approach fits?

  1. AA one-replica StatefulSet per session with the gVisor runtime class, scaled to zero replicas while the session is idle
  2. BAgent Sandbox, with sessions claiming Sandboxes from a SandboxWarmPool built on a SandboxTemplate
  3. CAgent Sandbox, with the application creating a Sandbox resource directly for each session at the moment the session starts
  4. DA Deployment of gVisor Pods with spare replicas kept warm, each session labelling one ready replica as its own
Show the answer and why
  • AA one-replica StatefulSet per session with the gVisor runtime class, scaled to zero replicas while the session is idle

    Incorrect

    GKE Sandbox isolates the kernel and the volume survives scaling to zero, but each new Pod waits for standard scheduling and startup instead of being handed out pre-warmed in under a second.

  • BAgent Sandbox, with sessions claiming Sandboxes from a SandboxWarmPool built on a SandboxTemplate

    Correct

    Agent Sandbox gives each agent a dedicated, kernel-isolated Pod, hands out pre-warmed Pods from a warm pool in under a second, and can suspend a Sandbox while keeping its persistent volumes.

  • CAgent Sandbox, with the application creating a Sandbox resource directly for each session at the moment the session starts

    Incorrect

    A Sandbox created on demand still waits for a new Pod to start; the sub-second hand-out comes from pre-warmed Sandboxes in a warm pool that a SandboxClaim draws from.

  • DA Deployment of gVisor Pods with spare replicas kept warm, each session labelling one ready replica as its own

    Incorrect

    Warm replicas are fast and kernel-isolated, but a Deployment's Pods have no per-session persistent volume or stable identity, and there is no way to suspend one session and later resume it with its volume.

Agent Sandbox on GKE is built for agents that run untrusted, LLM-generated code: each Sandbox is a dedicated Pod isolated with gVisor or microVMs; the SandboxClaim, SandboxWarmPool and SandboxTemplate resources hand out pre-warmed Pods in under a second; and setting a Sandbox's operating mode to Suspended deletes the Pod but keeps the resource and its persistent volumes.

Question 3 · choose 1

A company runs dev, staging and production GKE clusters, all enrolled in the Regular release channel. It wants every automatic minor upgrade to reach the dev clusters first, then staging, then production, with a fixed number of days between groups to check workloads on the new version. What should the platform team configure?

  1. AThe blue-green node upgrade strategy with a soak time on the production node pools
  2. BMaintenance windows that open on the dev clusters one week earlier than on production
  3. CThe dev clusters in the Rapid channel and the production clusters moved to the Stable channel
  4. DFleets per environment and a rollout sequence of them, with a soak time after each one
Show the answer and why
  • AThe blue-green node upgrade strategy with a soak time on the production node pools

    Incorrect

    A blue-green soak time pauses one node pool's upgrade before old nodes are removed; it does not order upgrades across groups of clusters.

  • BMaintenance windows that open on the dev clusters one week earlier than on production

    Incorrect

    Windows only say when automatic maintenance may run on each cluster; they do not make production wait for the same version to be validated upstream.

  • CThe dev clusters in the Rapid channel and the production clusters moved to the Stable channel

    Incorrect

    Each channel receives versions on its own schedule; rollout sequencing needs the clusters in the same release channel so that one version moves through the groups.

  • DFleets per environment and a rollout sequence of them, with a soak time after each one

    Correct

    Rollout sequencing upgrades groups of clusters in the defined order, with the soak time between groups to validate workloads before production.

GKE rollout sequencing defines an ordered sequence of cluster groups, such as fleets mapped to environments, all in the same release channel. When GKE sets a new automatic upgrade target, it upgrades each group in turn and waits the configured soak time, up to 30 days, before the next group. For new sequences Google recommends the custom stages version of the feature.

Question 4 · choose 1

A production fleet has three GKE clusters. The backend team runs on clusters 1 and 2 in the namespace backend; the frontend team runs on clusters 2 and 3 in two namespaces of its own. Each team must reach only its namespaces on its clusters and see logs and metrics for just those, and the platform team does not want to hand-write RBAC on every cluster. What should it set up?

  1. AThe namespaces created on all three clusters, relying on fleet namespace sameness to grant each team access everywhere
  2. BA team scope per team holding its clusters, with fleet namespaces created inside each scope
  3. CA separate fleet for each team, so that the clusters each team uses form a fleet of their own
  4. DRoleBindings for each team applied with kubectl to the namespaces on each of its clusters
Show the answer and why
  • AThe namespaces created on all three clusters, relying on fleet namespace sameness to grant each team access everywhere

    Incorrect

    Sameness treats same-named namespaces as one across the whole fleet, so each team would get its namespace on every cluster, not only on its own.

  • BA team scope per team holding its clusters, with fleet namespaces created inside each scope

    Correct

    Team scopes group the clusters a team may use, and fleet namespaces in a scope give team members access only on those clusters, with team-scoped logs and metrics.

  • CA separate fleet for each team, so that the clusters each team uses form a fleet of their own

    Incorrect

    Cluster 2 is shared by both teams, and a cluster belongs to one fleet, so a fleet per team cannot express the shared cluster.

  • DRoleBindings for each team applied with kubectl to the namespaces on each of its clusters

    Incorrect

    This is the per-cluster manual RBAC work that the platform team wants to avoid, and it gives no team-scoped view of logs and metrics.

Fleet team management uses team scopes, each associated with one or more member clusters, and fleet namespaces created in those scopes. Team members get access to their namespaces only on clusters in their scope, a cluster can belong to several scopes, and teams can view logs, utilization and metrics for their scope.

Question 5 · choose 1

A GKE Standard node pool of GPU nodes is being upgraded, but the project has no spare GPU quota, so GKE cannot create any extra node during the upgrade and the upgrade stalls. The training workload can tolerate losing one node at a time while it is recreated. Which surge upgrade settings should the team use for this node pool?

  1. AmaxSurge=1 and maxUnavailable=0, the default balanced setting
  2. BmaxSurge=2 and maxUnavailable=1, so that two nodes upgrade together
  3. CmaxSurge=0 and maxUnavailable=1, so that nodes are replaced one by one
  4. DAutoscaled blue-green upgrades, so that GKE drains the old nodes in steps
Show the answer and why
  • AmaxSurge=1 and maxUnavailable=0, the default balanced setting

    Incorrect

    With these values GKE creates a new surge node before removing an old one, which needs the GPU capacity the project does not have.

  • BmaxSurge=2 and maxUnavailable=1, so that two nodes upgrade together

    Incorrect

    Any maxSurge above zero still needs extra nodes, so the upgrade keeps waiting for capacity.

  • CmaxSurge=0 and maxUnavailable=1, so that nodes are replaced one by one

    Correct

    In a resource-constrained environment, setting maxSurge to 0 and maxUnavailable to 1 upgrades with fixed capacity: GKE takes one existing node down and recreates it.

  • DAutoscaled blue-green upgrades, so that GKE drains the old nodes in steps

    Incorrect

    Blue-green strategies create the new nodes alongside the old ones, which also needs spare quota and capacity.

Surge upgrades are the default node upgrade strategy. maxSurge sets how many extra nodes GKE may create, and maxUnavailable how many existing nodes it may disrupt at once. When quota or capacity for extra nodes is missing, use maxSurge=0 with maxUnavailable greater than zero to upgrade with fixed capacity.

Question 6 · choose 1

A platform team adds a Policy Controller constraint that requires a cost-center label on every Deployment in its staging and production clusters. For the first month, developers must get a message at the moment they apply a manifest that breaks the rule, but nothing may be blocked yet. Which enforcementAction should the constraint use?

  1. Awarn
  2. Bdryrun
  3. Cdeny
  4. DNo enforcementAction field, so the constraint only audits
Show the answer and why
  • Awarn

    Correct

    warn does not block the request but returns an immediate message about the violation at admission time; audit also records it.

  • Bdryrun

    Incorrect

    dryrun records violations for audit without blocking, but gives the developer no message when the manifest is applied.

  • Cdeny

    Incorrect

    deny rejects requests that violate the constraint, which the team does not want during the first month.

  • DNo enforcementAction field, so the constraint only audits

    Incorrect

    deny is the default enforcement action, applied even when the field is left out, so violating requests would be blocked.

Policy Controller constraints support three enforcement actions: deny, the default, blocks violating requests; dryrun reports violations in audit results and logs without blocking; warn behaves like dryrun but also returns an immediate warning at admission time. Teams often start with dryrun or warn before switching to deny.

Practice domain 1 →Practice all domains →