Shan Valleru's Blog

Every era of the cloud got a new tenant, and every new tenant got a new primitive.

Giga sells voice agents for enterprise customer support.

An AI agent is a while loop wrapped around a model call.

Your company announced an AI strategy.

You’re hiring a senior engineer.

You shipped the new platform.

Your engineering team is working hard.

Kubernetes was built to orchestrate containerized workloads at scale.

In 2023, GPT-4 cost $30/$60 per million tokens.

Kubernetes works until it doesn’t.

In 2020, OpenAI researchers noticed something strange.

I don’t write most of the words.

How do 86 billion neurons, each doing something simple, produce the feeling of being me?

There’s a progression happening in how developers work with AI.

There’s a certain kind of CEO who has done the reading.

Every efficiency initiative eventually targets the same thing: slack.

Every MBA student learns Porter’s Five Forces in their first strategy class.

Your platform team offers CI/CD, Kubernetes, observability, and secrets management.

Toyota revolutionized manufacturing by relentlessly eliminating waste—muda in Japanese.

Gresham’s Law: “Bad money drives out good.”

Your platform saves developers 2 hours per deployment.

Your cloud capacity is inventory.

“Don’t put all your eggs in one basket.”

Finance sees the platform team: 8 engineers, $2M/year, no direct revenue.

Everyone knows “cloud scales.”

Year 1: You hire 5 engineers to build an internal platform.

Prometheus is free.

Your sales team closes a landmark deal.

Your company’s Kubernetes cluster is an asset.

Your ML team needs GPUs.

Containers revolutionized deployment.

We had gang-scheduled jobs that performed DNS lookups at startup.

You deploy your controller with 3 replicas for high availability.

Your application has plenty of CPU headroom—at least according to your metrics.

A pod is created.

You delete a Deployment.

You’re building a Kubernetes controller.

You’ve got 2,000 Services in your cluster.

You apply a Pod manifest.

There’s no such thing as a container.

You went multi-cluster to escape the limits of a single control plane.

You run curl my-service:8080 from a pod, and it just works.

Every Kubernetes controller depends on watches.

Your cluster has 12 admission webhooks.

You’ve got 20 clusters.

Your custom controller watches for a label change.

If you’ve operated Kubernetes at scale, you’ve encountered the infamous 5,000 node recommendation.

If you’ve spent any time managing stateful applications on Kubernetes, you know the pain: manually scaling, handling backups, managing upgrades, recovering from failures.

Go’s concurrency model is often reduced to “goroutines are cheap, channels are cool.”

Python’s concurrency story is often summarized as “there’s a GIL, so use multiprocessing.”