Try each one before opening the answer. Every option is explained, with the Google Cloud documentation page that proves it.
Question 1 · choose 2
In one project and region, 40 N2 VMs run around the clock and are expected to stay for at least three years. Another 20 N2 VMs run only during the first 10 days of each month for month-end processing, which must not be interrupted once it starts. The FinOps lead wants the lowest cost for both groups. Which TWO choices should the team make? (Choose TWO.)
AA 3-year resource-based commitment that covers the vCPUs and memory of all 60 VMs
BA 3-year resource-based commitment that covers the vCPUs and memory of the 40 steady VMs
CRun the 20 month-end VMs on demand and let sustained use discounts apply to them
DRun the 20 month-end VMs as Spot VMs to get the deepest discount for the busy days
EBuy a sustained use discount in advance for the 20 month-end VMs, to be used during the 10 busy days
Show the answer and why
AA 3-year resource-based commitment that covers the vCPUs and memory of all 60 VMs
Incorrect
Commitments are billed monthly for the whole term whether or not the resources are used, so the share for the month-end VMs is paid for during the 20 idle days.
BA 3-year resource-based commitment that covers the vCPUs and memory of the 40 steady VMs
Correct
Resource-based commitments suit predictable, steady-state use, and a 3-year term has a higher discount than a 1-year term.
CRun the 20 month-end VMs on demand and let sustained use discounts apply to them
Correct
N2 VMs that run for more than 25% of a billing month get sustained use discounts automatically, and 10 days is about a third of the month.
DRun the 20 month-end VMs as Spot VMs to get the deepest discount for the busy days
Incorrect
Spot VMs can be preempted at any time, which breaks the rule that month-end processing must not be interrupted.
EBuy a sustained use discount in advance for the 20 month-end VMs, to be used during the 10 busy days
Incorrect
Sustained use discounts are not purchased; Compute Engine calculates and applies them automatically from usage.
Match each pricing model to the usage pattern: a resource-based commitment for usage that is steady all month for years, and on-demand capacity for the part-time group, which still earns sustained use discounts automatically once it runs more than a quarter of the month. Spot fits only work that tolerates preemption.
A GKE service writes about 2 TB a month of INFO request logs into its project's _Default log bucket, and Cloud Logging is now a large share of the bill. The team wants to store only about a tenth of the INFO entries, keep every WARNING and higher entry, and keep a user-defined log-based metric in the same project that counts every request from the INFO entries accurate. What should it configure?
AAn exclude_logs rule in the Ops Agent that drops nine in ten INFO entries before they leave the nodes
BAn exclusion filter on the _Default sink: severity<WARNING AND sample(insertId, 0.1)
CAn exclusion filter on the _Default sink: severity<WARNING AND sample(insertId, 0.9)
DThe _Default sink disabled, with a new sink for severity>=WARNING only
Show the answer and why
AAn exclude_logs rule in the Ops Agent that drops nine in ten INFO entries before they leave the nodes
Incorrect
The Ops Agent is the collector for Compute Engine VMs, and anything dropped before ingestion would also be missing from the request-count metric.
BAn exclusion filter on the _Default sink: severity<WARNING AND sample(insertId, 0.1)
Incorrect
sample() selects the given fraction, so this filter excludes about 10% of INFO entries and stores the other 90%.
CAn exclusion filter on the _Default sink: severity<WARNING AND sample(insertId, 0.9)
Correct
The filter excludes about 90% of entries below WARNING, so about a tenth is stored, and user-defined log-based metrics are calculated from all logs the project receives, regardless of exclusion filters.
DThe _Default sink disabled, with a new sink for severity>=WARNING only
Incorrect
This stores no INFO entries at all, so the team loses the tenth it wants to keep for analysis.
Exclusion filters on a sink stop matching entries from being routed and stored. Combined with sample(insertId, FRACTION), which selects roughly that fraction of entries, an exclusion can drop most low-severity entries while keeping a sample. Project-level user-defined log-based metrics still count every received entry, because they are calculated from all logs the Logging API receives for the project, regardless of inclusion and exclusion filters.
A company bought a Provisioned Throughput order for the Gemini model behind its document-summary agent. Finance now sees pay-as-you-go charges for the same model on busy days and wants the agent's model spend to be the fixed order price only, accepting that some requests fail at peaks. What should the developers do?
ASet the X-Vertex-AI-LLM-Request-Type header to shared on the agent's requests
BSet the X-Vertex-AI-LLM-Request-Type header to dedicated on the agent's requests
CLeave the requests unchanged, because spillover traffic is included in the order's fixed price
DCreate an alerts-only budget for the project at the order's monthly price
Show the answer and why
ASet the X-Vertex-AI-LLM-Request-Type header to shared on the agent's requests
Incorrect
shared sends requests straight to pay-as-you-go and bypasses the order, which increases the charges finance wants to avoid.
BSet the X-Vertex-AI-LLM-Request-Type header to dedicated on the agent's requests
Correct
dedicated uses only Provisioned Throughput; requests that exceed the order return a 429 error instead of spilling over to pay-as-you-go.
CLeave the requests unchanged, because spillover traffic is included in the order's fixed price
Incorrect
By default, a request that exceeds the remaining Provisioned Throughput quota is processed and billed at the pay-as-you-go rate as spillover.
DCreate an alerts-only budget for the project at the order's monthly price
Incorrect
An alerts-only budget sends notifications but doesn't stop usage or billing when the amount is reached.
With an active Provisioned Throughput order, requests consume the order first and, by default, overflow to pay-as-you-go as spillover. The X-Vertex-AI-LLM-Request-Type header controls this per request: dedicated means Provisioned Throughput only (429 on overage), shared means pay-as-you-go only.
The Managed Service for Prometheus bill of a GKE platform has tripled. Teams scrape their Pods every 10 seconds, and many exporters publish hundreds of metrics that no dashboard or alert uses. The platform team must cut the cost without losing the metrics that dashboards and alerts depend on. What should it do?
AShorten how long Cloud Monitoring keeps the Prometheus metric data, to the minimum the dashboards need
BLengthen the scrape intervals to 30 or 60 seconds and filter the collected metrics to the ones in use
CAdd an exclusion filter on the _Default log sink for the Prometheus exporters' entries
DLower the read volume by giving the dashboards a longer auto-refresh interval in Cloud Monitoring
Show the answer and why
AShorten how long Cloud Monitoring keeps the Prometheus metric data, to the minimum the dashboards need
Incorrect
The service charges for samples ingested and for read requests, so how long data is kept is not what drives this bill.
BLengthen the scrape intervals to 30 or 60 seconds and filter the collected metrics to the ones in use
Correct
The number of samples ingested is the main cost; longer sampling periods cut samples (by 66% from 10 to 30 seconds) and filtering removes unused time series.
CAdd an exclusion filter on the _Default log sink for the Prometheus exporters' entries
Incorrect
Log sink exclusions control log entries in Cloud Logging, not the metric samples ingested by Managed Service for Prometheus.
DLower the read volume by giving the dashboards a longer auto-refresh interval in Cloud Monitoring
Incorrect
Read requests are a smaller part of the charges; samples ingested are the primary contributor, and unused metrics are still ingested.
Managed Service for Prometheus charges mainly for samples ingested. The documented levers are fewer time series (filtering metrics with keep or drop rules), fewer samples per series (a longer scrape interval) and limits on high-cardinality metrics.
A Cloud Workstations configuration for 30 developers has a quick start pool of 10 and auto-sleep after 2 hours of inactivity. The bill shows VM charges through nights and weekends, when nobody works. Developers accept waiting a little longer for a workstation to start. Which change cuts this cost?
ALower the persistent home disk from 200 GB to 50 GB, since disk is the largest part of an idle workstation's charges
BRaise the auto-sleep time to 8 hours with the quick start pool unchanged, so that workstations restart less often
CDisable quick start workstations, so that no pre-started VMs are kept running when they are not in use
DSwitch the persistent home disk from Balanced to SSD, so that workstations boot faster and run for less time
Show the answer and why
ALower the persistent home disk from 200 GB to 50 GB, since disk is the largest part of an idle workstation's charges
Incorrect
The charges in question are VM charges at night; a smaller home disk does not stop VMs from running.
BRaise the auto-sleep time to 8 hours with the quick start pool unchanged, so that workstations restart less often
Incorrect
A longer auto-sleep keeps inactive workstations running longer, which adds VM time instead of removing it.
CDisable quick start workstations, so that no pre-started VMs are kept running when they are not in use
Correct
Quick start keeps a pool of VMs pre-started for faster startup, and the project is billed for those VMs; disabling it trades startup speed for lower cost.
DSwitch the persistent home disk from Balanced to SSD, so that workstations boot faster and run for less time
Incorrect
SSD costs more than Balanced, and faster boots do not remove the charges for VMs that run with no one using them.
In a workstation configuration, quick start keeps a number of VMs in a pre-started state so that workstations start faster, and the project pays for those VMs. Auto-sleep, two hours of inactivity by default, shuts down idle workstations. Disabling quick start, or lowering its pool size, is the documented trade of startup time for cost.
A cost review finds hundreds of Compute Engine persistent disks across development projects that no VM seems to use. The FinOps team wants a Google-generated list of these unused disks, with the recommended action of keeping a backup before each disk is removed. Which recommender should it use?
AThe idle persistent disk recommender, which recommends backing up and removing unused disks
BThe idle VM recommender, which lists unused VMs together with the disks attached to them
CThe idle custom image recommender, which finds images and disks that were created and then never used
DThe VM machine type recommender, which resizes VMs and frees the persistent disks they no longer need
Show the answer and why
AThe idle persistent disk recommender, which recommends backing up and removing unused disks
Correct
The idle persistent disk recommender targets disks and its recommendation is to back up and remove unused disks.
BThe idle VM recommender, which lists unused VMs together with the disks attached to them
Incorrect
The idle VM recommender recommends removing unused VMs; the disks in question are not used by any VM.
CThe idle custom image recommender, which finds images and disks that were created and then never used
Incorrect
That recommender recommends removing unused custom images, not persistent disks.
DThe VM machine type recommender, which resizes VMs and frees the persistent disks they no longer need
Incorrect
Machine type recommendations resize VMs; they do not report unused persistent disks.
Active Assist includes recommenders for idle resources: idle VMs (remove unused VMs), idle persistent disks (back up and remove unused disks), idle IP addresses, idle custom images, idle GKE clusters and idle Cloud SQL instances. Each targets one resource type and suggests a specific action.
To cut costs, a batch Job in a GKE Autopilot cluster now runs on Spot Pods:
spec:
nodeSelector:
cloud.google.com/gke-spot: "true"
terminationGracePeriodSeconds: 60
The workers save a checkpoint when they receive SIGTERM, which takes about
40 seconds. After preemptions, many checkpoints are incomplete. What
explains this, and what should the team do?
AThe nodeSelector is ignored in Autopilot; add a toleration for the Spot taint so that the 60-second grace period applies
BSpot Pods get at most 15 seconds of grace on preemption, whatever is requested; checkpoint more often so each save fits
CAutopilot caps every Pod's grace period at 30 seconds; raise the Job's resource requests so that the checkpoint runs faster
DSet terminationGracePeriodSeconds to 120, so that the Pods get twice the time the checkpoint needs before they stop
Show the answer and why
AThe nodeSelector is ignored in Autopilot; add a toleration for the Spot taint so that the 60-second grace period applies
Incorrect
A nodeSelector on cloud.google.com/gke-spot is the documented way to require Spot Pods in Autopilot, and tolerations do not extend the grace period.
BSpot Pods get at most 15 seconds of grace on preemption, whatever is requested; checkpoint more often so each save fits
Correct
The maximum grace period for Spot Pods during preemption is 15 seconds, and requesting more does not grant more.
CAutopilot caps every Pod's grace period at 30 seconds; raise the Job's resource requests so that the checkpoint runs faster
Incorrect
The 15-second limit applies to Spot Pods at preemption, and larger requests do not lift it.
DSet terminationGracePeriodSeconds to 120, so that the Pods get twice the time the checkpoint needs before they stop
Incorrect
Requesting more than 15 seconds does not grant more than 15 seconds when a Spot Pod is preempted.
Spot Pods in Autopilot cost less but can be preempted. Require them with a nodeSelector or node affinity on cloud.google.com/gke-spot=true. On preemption they get at most 15 seconds of grace whatever terminationGracePeriodSeconds says, so workloads must tolerate interruption, for example by checkpointing regularly.
A team plans a resource-based commitment in project ml-prod, region us-central1, for vCPUs, memory and 8 NVIDIA A100 GPUs of the A2 machine series, and wants to attach a reservation to it so that the capacity is assured. Which reservation can be attached to the commitment?
AA reservation in the project ml-dev in us-central1 for A2 VMs with 8 A100 GPUs, shared with ml-prod
BA reservation in ml-prod in us-east1 for A2 VMs with 8 A100 GPUs, with auto-delete disabled
CA reservation in ml-prod in us-central1 for A2 VMs with 4 A100 GPUs, to be doubled once demand grows
DA reservation in ml-prod in us-central1 for A2 VMs with 8 A100 GPUs and auto-delete disabled
Show the answer and why
AA reservation in the project ml-dev in us-central1 for A2 VMs with 8 A100 GPUs, shared with ml-prod
Incorrect
Reservations attached to a resource-based commitment must be in the same project as the commitment.
BA reservation in ml-prod in us-east1 for A2 VMs with 8 A100 GPUs, with auto-delete disabled
Incorrect
The attached reservation must be in the same region as the commitment.
CA reservation in ml-prod in us-central1 for A2 VMs with 4 A100 GPUs, to be doubled once demand grows
Incorrect
When a commitment covers GPUs, the attached reservations must specify exactly the same numbers and types of GPUs.
DA reservation in ml-prod in us-central1 for A2 VMs with 8 A100 GPUs and auto-delete disabled
Correct
It matches the commitment's project, region, machine series and exact GPU count and type, and has auto-delete disabled, as attached reservations require.
Reservations attached to resource-based commitments must be in the same project and region as the commitment, use the same machine family series (any machine type in it), have auto-delete disabled, and, if the commitment includes GPUs or applicable Local SSD disks, specify exactly the same numbers and types of those resources.