Patrick Richter

KVM for AI workloads makes sense when you need proper isolation, but most AI agents don't need hypervisor overhead unless they're handling sensitive data.

Been there with autonomous deployments failing on flaky tests. Infrastructure drift is usually the culprit, not the code generation itself.

"We need someone to review our infrastructure setup."

"Around 40k monthly spend, previous consultant said it was optimized."

Opened the console. Single app. 23 subnets across 8 regions. Load balancers with no targets. RDS in every AZ.

"Optimized for what exactly?"

Still need to deal with the banking maze there though. Had clients struggle with opening accounts even with the visa sorted.

CNI abstractions help but you still need that networking foundation when troubleshooting pod connectivity issues in production.

That's a massive bet on AI infrastructure paying off long term. Wonder how much of that goes to custom silicon vs just more data centers.

The orchestration overhead alone can double your compute costs if you're not careful with resource requests and limits.

Fair point, game engines have their own workflow patterns that don't always translate to typical software practices.

Classic scope creep. At least you stopped before adding a config file and CLI flags.

Apprenticeship programs are solid for breaking into SRE. Much better than bootcamps since you actually learn production systems under guidance.

State file drift between local and pipeline? Usually points to different backend configs or someone manually nuked resources through console.

Classic AWS error messages. Had Terraform think an entire VPC was gone because of one missing IAM permission last month.

Yeah, the cognitive load is real. Takes me about 2 weeks to find a decent workflow in each new city.

Zone files are beautiful but try explaining SOA serial increments to a junior dev who just wants to add a CNAME.

The moved blocks are a game changer for refactoring without the old state mv dance. Still miss being able to grep through HCL sometimes though.

Azure's policy sunset timing is brutal when you're already stretched thin. LLM governance feels like the wild west right now.

Been there with the credential sprawl nightmare. Nothing like LLM confidence meeting production terraform to remind you why we have approval gates.

Yeah, the docker-compose to k8s YAML translation dance gets old fast. Half the time the "simple" examples skip the networking bits that actually matter.

Nah, running VMs in k8s is still a mess. KubeVirt exists but honestly Nomad's multi-driver approach just works better for mixed workloads.

Been there with midnight hotfixes that never made it back to Terraform. Running drift detection in CI helps catch it early before it becomes a mess.

Exactly. Most people underestimate how much mental energy goes into constantly adapting to new places without good habits.

True for real-time inference, but most AI workloads are batch or near-real-time where cloud elasticity still wins. Depends what you're building.

This is why I always run plan first and actually read the diff. Too many people just apply blindly and wonder why their prod went sideways.

Exactly this. Learned more about actual system reliability from a brief stint at a tiny ISP than from years of enterprise "best practices."

$85/hr for senior data architect seems low for current market, especially with that tech stack.

Finally, proper plan/apply separation for K8s deployments. Been waiting for this workflow outside of Terraform for ages.

Half my best skills came from jobs I stumbled into rather than planned for. Sometimes accidental experience beats strategic career moves.

Works until you hit a real outage that costs actual money. Most startups learn this the expensive way.

Exactly why I moved to freelancing. I keep the skills sharp but someone else gets the 3am pages.

Fair point, specialization beats being spread thin. I just like having the ops skills when third party stuff inevitably breaks at 3am.