Google Cloud services
23 / 24
Google Cloud

Cost Optimisation & the Architecture Framework

Cut spend without cutting capability, and name the pillars a Google architecture review scores against.

The Google Cloud Architecture Framework has five pillars.

A household bill. The savings are not in a cheaper kettle but in the immersion heater nobody remembered was on and the subscription nobody cancelled.

Key Concepts

1
    Operational excellence   deploy, monitor and improve
    Security, privacy and compliance
    Reliability              recover from failure, meet demand
    Cost optimisation        avoid unnecessary spend
    Performance optimisation use the right resources
2
The discounts, and two of them are automatic.
    Sustained Use Discounts   AUTOMATIC. Run a VM most of the month
                              and the rate drops, with no commitment.
    Committed Use Discounts   1 or 3 years, up to ~70%. Resource-based
                              or the more flexible spend-based.
    Spot VMs                  up to 91% off, reclaimed with 30
                              seconds' notice
    Custom machine types      pick exact vCPU and memory, so you stop
                              paying for a size that nearly fits
3
Sustained use being automatic is the GCP-specific point — on other clouds that saving requires a commitment.
4
Custom machine types are the quiet win. A workload needing 6 vCPU and 20 GB fits no standard size, so elsewhere you buy 8 vCPU and 32 GB. Here you buy what you need.
5
Where money actually leaks.
    idle VMs and oversized instances     Recommender flags these
    unattached persistent disks          the VM went, the disk stayed
    old snapshots and untagged images
    BigQuery SELECT * on wide tables     billed per byte scanned
    egress, and cross-zone traffic
    Cloud Composer or a GKE cluster left running for one job
    non-production running overnight
6
BigQuery deserves its own line. Billing by bytes scanned means SELECT * on a wide table can cost more than a day of compute. Partition, cluster, select only the columns needed, and set custom quotas so one query cannot run away.
SELECT *
7
Active Assist and Recommender give concrete rightsizing, idle resource and commitment recommendations — the sensible first pass before any manual review.
8
Labels are the attribution mechanism. Without team, env and service labels, billing export to BigQuery cannot tell anyone which team spent what, and nothing improves while nobody owns the number.
teamenvservice
9
Budgets and alerts fire at thresholds and can trigger a Pub/Sub message, so automation can react rather than an email being ignored.
10
What the interviewer is probing.1. "Which GCP discount needs no commitment?" Probing: the automatic one. Stalls: "None." Moves up: sustained use discounts apply automatically to VMs running most of the month.
11
2. "What is the quiet win on machine sizing?" Probing: custom types. Stalls: "Pick the nearest size." Moves up: custom machine types — a workload needing 6 vCPU and 20GB buys exactly that rather than rounding up to 8 and 32.
12
3. "Where does BigQuery spend get out of hand?" Probing: the scan model. Stalls: "Storing too much." Moves up: bytes scanned — SELECT * on a wide unpartitioned table, which partitioning, clustering and column selection fix.
13
4. "How do you attribute spend to teams?" Probing: labels. Stalls: "Separate projects." Moves up: labels flowing into the billing export, so Cost reports show who spent what and someone owns the number.
Use cases:cost-optimisationcommitted-usesustained-usebigquery-costlabels