Design and 'what would you do' questions
Practise these as flashcards. One or two lines each; the ones you miss come back first.
Start freeClarify uptime, data sensitivity and peak load. Then: two web VMs behind an L7 load balancer with health checks and TLS termination, app tier separate from the database, database with replication and backups, an internal DNS name, firewall rules per tier, monitoring on the golden signals and host metrics, central logs, cert expiry alerts, a patch group, and a runbook. Trade-offs: VMs vs containers, shared vs dedicated database.
Inventory first: hardware, apps, dependencies, owners. Pilot with 5–10. Automated builds so every machine is identical. Waves by site and criticality, with rollback (keep the old image for two weeks). A validation checklist per machine, a communication plan, and metrics per wave. If you've done one, tell it with your own numbers.
Get data before opinions. Check LB latency per backend (is it one server?), server CPU, memory and disk I/O, app logs for slow calls, database wait stats, and the network path. Build a timeline of what changed. Narrow it down by elimination and share findings with evidence, not blame.
Inventory every cert and expiry date. Group by renewal method: internal CA, public CA, ACME. Automate what you can. Schedule the rest in change windows by expiry date. Add monitoring at 30, 14 and 7 days so it never sneaks up again.
Back it up first. Inventory OS, packages, running services, listening ports, cron jobs, mounts, and who connects to it (LB pools, DNS records, firewall rules). Read configs and logs. Write the doc as you go and change nothing until you understand the dependencies.
A golden image or template, automated install (cloud-init, Ubuntu autoinstall, Kickstart, or sysprep with unattend), Ansible for post-build configuration, a request form that feeds the automation, and a validation script that checks the standard. Measure before and after.
Fault tolerance: no interruption at all (redundant power supplies, RAID). High availability: automatic recovery in seconds to minutes (LB pools, cluster failover). Disaster recovery: recovery at another site after a major event, within a planned RTO and RPO.
First 30 days: learn the environment, read runbooks, shadow on-call, meet app owners, map what is and isn't monitored. Next 30: pick the top two pain points, usually alert noise and manual builds, and fix one. By day 90: propose a roadmap with measurable targets and own a patch cycle end to end.
Start with RTO/RPO from the business. Then: infra as code so the second region can be built or kept warm, data replication (SQL failover groups, GRS storage, ASR for VMs), global traffic routing (Front Door or Traffic Manager) with health-based failover, secrets and certs available in both regions, DNS and firewall rules pre-staged, and a tested runbook. Run a real failover test at least yearly; an untested DR plan is a guess.
Don't just say yes or no. Find what's truly needed for day one (MVP), run long-lead items in parallel (cert, firewall, DNS, CAB), use standard templates, and make the risks explicit in writing: what's skipped (load test, DR) and who accepts it. Get the business owner to sign the risk acceptance, and schedule the remaining work as follow-ups.
Ask: is the app stateless, is the vendor supporting containers, does the team have the skills to run them, what are the scaling and release-frequency needs, and what's the cost of running VMs (patching, hardening) vs a platform. Legacy, stateful or vendor-locked apps stay on VMs; new stateless services go to PaaS or containers. Say the trade-off out loud.
Measure first: categorize a month of tickets and find the top five types by volume and time. Automate or self-service the biggest (Ansible job, Terraform module, ServiceNow catalog item), fix root causes of repeat incidents, and track ticket volume before and after. Show the hours saved in numbers.