Professional Cloud DevOps Engineer · Domain 4: Implementing observability practices and troubleshooting issues · 22% of the exam
Task 4.3: Managing metrics, dashboards, and alerts.
Metrics Explorer, dashboards and PromQL, alerting policies on SLIs, SLOs and cost, third-party notification channels and AI-assisted metrics interpretation.
Try each one before opening the answer. Every option is explained, with the Google Cloud documentation page that proves it.
Question 1 · choose 2
A service in Cloud Monitoring has a 99.9% availability SLO over 30 days. The SRE team wants the on-call engineer paged quickly when an outage is eating the error budget fast, and a ticket opened when a smaller problem keeps consuming the budget a little too fast for a long time. Which TWO SLO alerting policies should it create? (Choose TWO.)
ABurn rate above 10 over a 1-hour lookback period, notifying the paging channel
BBurn rate above 10 over a 24-hour lookback period, notifying the paging channel
CBurn rate above 1 over a 30-day lookback period, notifying the ticket queue
DBurn rate above 2 over a 24-hour lookback period, notifying the ticket queue
EError budget remaining below 0%, notifying both the paging channel and the ticket queue
Show the answer and why
ABurn rate above 10 over a 1-hour lookback period, notifying the paging channel
Correct
A fast-burn alert uses a short lookback and a high threshold, about 10 times the baseline, so a sudden outage is caught within minutes.
BBurn rate above 10 over a 24-hour lookback period, notifying the paging channel
Incorrect
A high threshold over a long window lets a severe outage burn budget for hours before the alert can fire, which defeats a fast-burn page.
CBurn rate above 1 over a 30-day lookback period, notifying the ticket queue
Incorrect
Burn-rate alerts cannot use a compliance period longer than 24 hours, so a 30-day lookback is not available.
DBurn rate above 2 over a 24-hour lookback period, notifying the ticket queue
Correct
A slow-burn alert uses a longer lookback to smooth out short spikes and a threshold slightly above the baseline, such as 2 times over 24 hours.
EError budget remaining below 0%, notifying both the paging channel and the ticket queue
Incorrect
A budget below zero already means the SLO is violated, so the alert arrives after the damage instead of warning while budget remains.
Burn rate compares how fast the error budget is consumed with the rate that would use it up exactly at the end of the period. Pair a fast-burn policy (short lookback, high threshold, page) with a slow-burn policy (lookback of up to 24 hours, threshold just above 1, ticket); lookback periods longer than 24 hours are not supported for these alerts.
To control logging costs, a team created a PromQL alerting policy with this
condition:
sum(rate(logging_googleapis_com:billing_bytes_ingested{monitored_resource="global"}[10m])) > 200000
Even on normal days the alert keeps opening and closing, although total
ingestion over each day is steady. What change fixes the condition?
AUse a lookback window longer than one hour in the query, such as [2h]
BKeep the 10-minute window and raise the threshold tenfold so the hourly bursts stay under it
CSet the condition's retest window to 0 seconds so one sample decides the state
DShorten the evaluation interval so the policy checks the query every 15 seconds
Show the answer and why
AUse a lookback window longer than one hour in the query, such as [2h]
Correct
Logging reports billing usage in hourly batches, so a window shorter than an hour sees bursts and gaps; Google's example uses [2h].
BKeep the 10-minute window and raise the threshold tenfold so the hourly bursts stay under it
Incorrect
The short-window rate still swings between a burst and nothing every hour, so a threshold high enough to stop the flapping also hides real increases; the documented fix is a longer lookback.
CSet the condition's retest window to 0 seconds so one sample decides the state
Incorrect
0 seconds is already the default, and it lets a single measurement trigger a notification, which adds to the flapping.
DShorten the evaluation interval so the policy checks the query every 15 seconds
Incorrect
Evaluating more often does not change the hourly shape of the data; it only re-checks the same short window more often.
billing_bytes_ingested arrives in hourly batches. A PromQL rate over a short window alternates between a burst and nothing, so a cost alert on it flaps; a lookback longer than an hour gives a stable rate to compare with the threshold.
Critical alerting policies in Cloud Monitoring notify a PagerDuty channel. The SRE team wants a second channel on the same policies that still delivers if the path to PagerDuty has a problem, and an internal tool must read those notifications and open incident records automatically. Which channel type should the team add?
APub/Sub
BWebhooks
CSlack
DSMS
Show the answer and why
APub/Sub
Correct
Email or Pub/Sub is the recommended redundant channel next to PagerDuty, and Pub/Sub (or webhooks) is the recommended channel for automation tools.
BWebhooks
Incorrect
Webhooks suit automation, but PagerDuty, webhooks, Slack and the mobile app share one internal delivery service, so a webhook is no backup for PagerDuty.
CSlack
Incorrect
Slack shares the same delivery service as PagerDuty, and Google recommends against using it with automation tools because the format can change.
DSMS
Incorrect
SMS isn't fully reliable or available everywhere, and its messages are not meant to be parsed by automation tools.
Cloud Monitoring recommends more than one channel type for redundancy. Mobile app, PagerDuty, webhook and Slack notifications share a single point of failure, so the backup should be email or Pub/Sub; for tools that read or act on notifications, webhook or Pub/Sub is recommended. Only Pub/Sub meets both needs.
One alerting policy watches disk usage on 40 VMs. Next Saturday from 02:00 to 06:00, one of them, db-07, will be rebuilt and will trip the policy. The on-call team wants no alerts or pages for db-07 during that window, while the other 39 VMs stay covered, and nobody should have to remember to undo anything afterward. What should the team set up?
ADisable the alerting policy on Saturday at 02:00 and enable it again once the rebuild is finished
BRemove the notification channels from the policy for the window, then add them back after the rebuild
CAdd a log exclusion filter that drops db-07's logs during the window, so that the alert has no data to evaluate
DA snooze on the policy for Saturday 02:00 to 06:00, with a label-based filter that matches only db-07's instance ID
Show the answer and why
ADisable the alerting policy on Saturday at 02:00 and enable it again once the rebuild is finished
Incorrect
Disabling the policy stops alerts for all 40 VMs, and someone must remember to enable it again.
BRemove the notification channels from the policy for the window, then add them back after the rebuild
Incorrect
Alerts would still be created for every VM with no one notified, and the channels must be restored by hand.
CAdd a log exclusion filter that drops db-07's logs during the window, so that the alert has no data to evaluate
Incorrect
Exclusion filters apply to log ingestion; they do not stop a policy that evaluates disk usage metrics.
DA snooze on the policy for Saturday 02:00 to 06:00, with a label-based filter that matches only db-07's instance ID
Correct
A snooze has a start time and duration, and on a single policy a label-based filter limits it to matching resources while alerts for the other VMs continue.
Snoozes stop Cloud Monitoring from creating alerts and sending notifications for matching alerting policies during a scheduled period, for example for planned maintenance. For a single policy, a label-based filter can limit the snooze to one resource. Unlike disabling a policy, a snooze ends on its own.
An alerting policy has two conditions on Compute Engine VMs: CPU utilization above 90% for 10 minutes, and memory utilization above 90% for 10 minutes. The team wants an alert only when the same VM breaches both, but last night it was paged because vm-a had high CPU while vm-b had high memory. Which combiner value should the policy use?
AOR, so that the policy opens one alert for whichever condition is met first on any VM
BAND_WITH_MATCHING_RESOURCE, so that one resource must meet both conditions before alerts open
CAND, the current default, with the retest window of both conditions raised from 10 minutes to 30 minutes
DAND, with the two conditions moved into two separate policies linked to the same notification channel
Show the answer and why
AOR, so that the policy opens one alert for whichever condition is met first on any VM
Incorrect
OR opens an alert when any resource meets any condition, which pages even more often.
BAND_WITH_MATCHING_RESOURCE, so that one resource must meet both conditions before alerts open
Correct
This most stringent combiner requires that the same resource causes each condition to be met, identified by matching resource labels after aggregation.
CAND, the current default, with the retest window of both conditions raised from 10 minutes to 30 minutes
Incorrect
AND opens alerts when all conditions are met even by different resources; a longer window does not tie them to one VM.
DAND, with the two conditions moved into two separate policies linked to the same notification channel
Incorrect
Separate policies each alert on their own, so a single breach on any VM would page.
A policy with several conditions (up to six) uses a combiner: OR opens an alert when any condition is met; AND, the default, when all are met even by different resources; AND_WITH_MATCHING_RESOURCE only when the same resource meets every condition, which requires the conditions to monitor the same resource type and keep the same resource labels after aggregation.
A dashboard chart uses PromQL to show the CPU utilization of production VMs,
which carry the user label env=prod. The team writes:
avg(compute.googleapis.com/instance/cpu/utilization{env="prod"})
Cloud Monitoring rejects the query. Which query is correct?
Under the UTF-8 rules, a metric name with dots and slashes is quoted inside the braces, and a user metadata label is referred to with the metadata_user_ prefix.
Label names with dots are not valid unquoted, and the documented form for a user metadata label is metadata_user_env.
PromQL in Cloud Monitoring follows the UTF-8 specification: metric names such as kubernetes.io/container/cpu/limit_utilization are quoted and placed inside the braces, and metadata labels are written as metadata_system_* or metadata_user_*, for example sum("compute.googleapis.com/instance/cpu/utilization") by (metadata_user_env).
In Metrics Explorer, an engineer charts a cumulative metric that counts the bytes a service has sent since it started. The line only climbs, which makes it impossible to see traffic peaks. The engineer wants the chart to show bytes per second. Which alignment function should the engineer configure?
Anext older, so that each alignment period keeps only the most recent sample of the counter
Bdelta, so that the chart shows how many bytes were sent in each alignment period
Crate, so that the cumulative count becomes a gauge of bytes per second for each period
Dpercentile with the 95th value, so that short bursts stand out from the steady growth of the line
Show the answer and why
Anext older, so that each alignment period keeps only the most recent sample of the counter
Incorrect
next older suits gauge metrics, such as uptime check results; it does not turn a cumulative count into a per-second value.
Bdelta, so that the chart shows how many bytes were sent in each alignment period
Incorrect
delta gives the change per alignment period, not per second, so its values shift whenever the period changes.
Crate, so that the cumulative count becomes a gauge of bytes per second for each period
Correct
rate converts a cumulative or delta metric into a gauge, like a delta divided by the alignment period, giving a unit such as bytes per second.
Dpercentile with the 95th value, so that short bursts stand out from the steady growth of the line
Incorrect
Percentile functions apply to distribution-valued metrics, not to a cumulative counter of bytes.
Metrics Explorer alignment functions transform each time series per alignment period. rate turns cumulative or delta metrics into a per-second gauge, delta gives the change per period, next older keeps the latest sample of a gauge, and percentile picks a value from distribution metrics.