PCDOE question patterns

Cloud Build for CI and Cloud Deploy for promotion, approval and rollback.

Managed services avoid running Jenkins or custom scripts.

Use Cloud Service Mesh golden-signal telemetry, or log-based metrics for log-only signals.

No instrumentation is needed.

Shift traffic to the previous revision or roll back the release, then investigate.

Rollback is the fastest, lowest-risk mitigation.

Federate the CI's OIDC identity and impersonate a service account.

No long-lived keys to leak.

An SLA is the external contract with consequences; keep the SLO stricter.

The SLO gives early warning before the SLA is breached.

Set a realistic SLO below 100% and use the error budget for change.

100% is unaffordable and blocks all releases.

Follow the pre-agreed error budget policy: only fixes and reliability work until budget recovers.

The policy, not opinion, balances velocity and stability.

Page at 14.4x over 1 h and 6x over 6 h; ticket at 1x over 3 days.

Short windows reduce reset time, long windows reduce false pages.

Good requests over valid requests, and fraction under a latency threshold.

They reflect user experience; CPU does not.

Automate or remove the cause; keep toil under 50%.

Toil grows with the service and has no lasting value.

Raise or request quota and alert on usage.

Autoscaling can't exceed project quotas or max replicas.

Create a reservation or future reservation.

Reservations guarantee capacity; CUDs only discount.

Use Flex-start or Calendar mode.

It queues for accelerators when available.

Make the initialization period match real startup time.

The autoscaler otherwise reads unready instances' metrics.

Use folders by environment or team and grant roles to groups.

Policies inherit downward.

Set constraints/gcp.resourceLocations (and Assured Workloads if needed).

Org policy blocks creation elsewhere regardless of IAM.

Host project owns the network, service projects attach.

Central control with project separation.

Create a PSC endpoint.

No CIDR coordination; one-way access.

Add the projects to a scoping project.

Cloud Monitoring views multiple projects from one.

Declare resources as custom resources synced from Git.

Continuous reconciliation reverts drift.

Build once, tag by SHA, promote the rendered release.

Rebuilding breaks traceability.

Set it on the production target.

Rollouts pause until an approver acts.

Define phases and verification jobs, add an automation to repair on failure.

Limits blast radius.

Peer the pool with your VPC.

The default pool has no VPC access.

Reference Secret Manager with secretEnv.

Substitutions and build args leak.

Require attestations signed with Cloud KMS keys.

Deploy-time enforcement of trust.

Use DRYRUN_AUDIT_LOG_ONLY.

Logs violations only.

Cloud Build level 3 provenance attestation, verified at deploy.

Signed metadata about source and builder.

Automatic scanning and rescanning; on-demand scan in CI.

Findings can gate promotion.

Deploy, test the tagged URL, then split traffic.

Instant rollback by shifting back.

Keep the old pool until validated; use maintenance windows.

Surge upgrades are cheaper but less reversible.

Install it and grant logWriter and metricWriter.

Hypervisor metrics exclude memory.

Use PodMonitoring with Managed Service for Prometheus.

Query with PromQL.

Aggregated sink with a retention policy and bucket lock.

Cheapest long-term archive.

Upgrade the bucket or export to BigQuery.

Joins with other data.

Subscriber forwards to Splunk or similar.

Streams in near real time.

Give the sink's service account a role on the destination.

Sinks write as their own identity.

Exclude or sample low severity; filter at source; shorten retention.

Ingestion is the main charge.

Use a log match condition with a rate limit.

No metric needed.

Probe from multiple regions.

Black-box monitoring sees what users see.

Propagate W3C trace context and read spans.

Profiler then finds the function.

IC coordinates; one operations lead changes the system; comms lead updates.

Clear command prevents conflicting changes.

Record impact, timeline, causes, owned action items.

Blame hides information.

Drain the backend or fail over.

Traffic leaves while you investigate.

Autoscaler may be capped by max nodes or quota.

Pending means unschedulable.

60 to 91% off with checkpointing and retries.

Can be preempted with 30 s notice.

Commit the stable minimum only.

Variable load stays on-demand or Spot.

Review cost recommendations in Recommendation Hub.

They estimate savings.

Reserve dedicated capacity sized from measured tokens.

Pay-as-you-go can be throttled under contention.

Cap output tokens, use batch for offline work.

Cost scales with tokens.