DevOps interview practice by topic.

Signals and the OOM killer, cgroups v2 and namespaces, systemd, /proc, TCP states and TIME_WAIT, DNS resolution order, and the Bash/Python/Go a Google SRE or Meta PE screen actually asks you to write.

IAM least privilege, VPC design, security groups versus NACLs, why peering is not transitive, Transit Gateway routing, multi-account landing zones and multi-region failover across AWS, Azure and GCP.

Terraform and OpenTofu state, remote backends and locking, drift detection, count versus for_each, module design, lifecycle blocks, Ansible idempotence, and the BUSL license split every interviewer now probes.

Control plane internals, the full kubectl apply to running pod path, scheduling and affinity, native sidecars, in-place pod resize, CNI and network policy, CSI and StatefulSet identity, RBAC, operators and service mesh.

Pipeline design and caching, artifact promotion, rolling versus blue-green versus canary and what each costs, progressive delivery with Argo Rollouts and Flagger, ArgoCD and Flux reconciliation, and the DORA metrics behind them.

Metrics versus logs versus traces, PromQL, Prometheus cardinality blowups, OpenTelemetry collectors and temporality, exemplars, SLI and SLO design, error budget math, and burn-rate alerting that survives an on-call rotation.

The highest-signal round on the loop. CrashLoopBackOff with empty logs, exit code 137, ImagePullBackOff, stuck PVC finalizers, 503s from a healthy-looking service, p99 tails, incident command, and the blameless postmortem.

SAST, DAST, SCA and IAST placement in a pipeline, Trivy and Checkov gates, secrets management and rotation with Vault, image signing and SLSA provenance, zero trust in Kubernetes, and compliance as code.

Google-style NALSD with napkin math before boxes, multi-region active-active, disaster recovery under a real RTO, capacity planning, blast radius, and designing a delivery platform for hundreds of services.

Internal developer platforms, golden paths, Backstage catalogs and templates, Crossplane, self-service without a ticket queue, plus cost allocation, showback, Karpenter, spot strategy and commitment management.

The fastest-growing DevOps sub-domain: GPU scheduling with MIG, time-slicing and DRA, vLLM paged attention and continuous batching, KServe and Ray, gang scheduling with Kueue, and the cost model behind inference.

Amazon Leadership Principles in STAR form, the outage story that shows ownership rather than blame, pushing back on a bad deadline, driving a migration nobody wanted, and explaining risk to people who do not write code.