Key facts distilled from CertKata's GCP-PCA question bank. Free to read — no account needed.
An SLI measures a user-facing outcome, not a resource internal - CPU utilization is not a service level indicator. SLOs turn that measurement into a target, error budget consumption turns the target into a priority, and burn-rate alerts are what page a human, so paging is tied to user impact instead of threshold noise. Data pipelines get freshness SLOs for the same reason: a silent stall produces no error. Measure first - SLOs are what focus reliability work.
Google's framework has six pillars: operational excellence, security, reliability, performance optimization, cost optimization, and sustainability. They interact, so a decision rarely improves one without costing another - committed-use discounts cut spend but must not eat the headroom reliability depends on. Resolve a conflict by quantifying the trade-off against a stated target, not by defaulting to one pillar. A review produces prioritized, pillar-mapped improvements rather than a verdict.
IAM decides who may act; a service perimeter decides where data may move. VPC Service Controls bound API-level data movement independently of IAM, so credentials that are perfectly valid still cannot copy data out of the perimeter from an unapproved context. That is the control for exfiltration risk, and it is why analysis compute is placed inside the perimeter rather than granted access through it. Residency controls plus governed API exposure are how sovereignty and sharing coexist.
SLO attainment is the reliability metric leadership can act on, and a gap between the internal view and the customer view means the SLIs measure the wrong thing. A cost-of-downtime model turns an availability debate into a decision: minutes-level RTO requires active-active or hot standby, and should be paid for only where the loss justifies it. Recovery objectives and dependency ordering drive the continuity plan, which is a process with people and communication, validated by game days.
Match the access pattern, not the brand. High-throughput keyed and time-series reads go to Bigtable; large-scale analytical queries go to BigQuery, which is also where the lake lands alongside Cloud Storage. Cloud SQL is the cheapest relational start and Spanner is the destination when global strong consistency is required - bounded staleness trades freshness for read latency. Read-heavy, rarely-changing data belongs in a cache. Shared multi-writer access needs a file service, not a block disk.
This is a preview. Get the full GCP-PCA cheat sheet — every domain, bundled with the practice question bank.
Sign up for the full version