dipanshu

22 ... talk about tech and lifestyle ... building #kubernetes cost optimiser

[dipanshu @_dipansh]

Your code is slow. You blame Python. You're wrong.

You're probably using a 5-star chef (CPU) to do a burger-flipper's (GPU) job.

Read more at - https://t.co/GHGcHLHZHa

[dipanshu @_dipansh]

Kubernetes “downtime” is usually just your app whispering sweet nothings to a dead Pod IP.

[dipanshu @_dipansh]

Me in dev: "NodePort is so easy! Just minikube service my-app and it works."

Me in prod: "WHY IS PORT 30007 OPEN ON EVERY. SINGLE. NODE."

It's the "quick and dirty" dev tool. Keep it that way. Don't let it near production.

Read more at: https://t.co/pMWI8WYf5l

[dipanshu @_dipansh]

Stop. Putting. Retry. Logic. In. Your. App. Code.

Your app's job is business logic, not figuring out exponential backoff for a flaky auth-service. That's platform-level noise.

Abstract it. Move it to a layer built for it (a service mesh).

https://t.co/8BkSXOh4KO

[dipanshu @_dipansh]

Netflix's most brutal, efficient optimization?

They don't waste time fixing corrupt video sources.

They just reject them.

"Garbage in, garbage out" is an infra principle. If the source is bad, fail the pipeline.

Stop building complex error concealment for upstream's failures.

Read More at: https://t.co/u9ZozbycVf

[dipanshu @_dipansh]

@Matt_Pinner weird biology

[dipanshu @_dipansh]

PM: "Let's add video previews! It'll be easy!"

Me, dying inside, knowing they're about to ask me to serve a 4GB .wmv file to an iPhone on 4G.

The "fast vs. cheap" transcoding problem is a classic infra trap.

Read more: https://t.co/B5vByR67de

[dipanshu @_dipansh]

After analyzing terabytes of cost data from 30+ K8s clusters, I can tell you this:

Nobody has a node problem. They have a request problem.

Your bill isn't high because you chose the wrong EC2 instance. It's high because 400 services are requesting 2 CPUs and using 0.1, and your logging sidecar is shipping 5TB a day.

Stop optimizing the instance type.
Go look at Opencost and find the why.

[dipanshu @_dipansh]

Platform Engineering is just what we called good DevOps until it needed a marketing budget.

You're not a platform engineer if you just run Terraform for other teams.

You're a platform engineer when you give devs a self-service API (like the 25+ I built for my last agent), sane defaults, and a clear cost model.

Developers are your customers, not your ticket queue.

[dipanshu @_dipansh]

@Ajatshatru_28 everyone know what its actual usage is .. and i am not talking about cooking

[dipanshu @_dipansh]

@ShishirShelke1 animations this time it just shit

[dipanshu @_dipansh]

@Nythis110929 whts the font

[dipanshu @_dipansh]

Stop blaming your app for latency spikes during autoscaling.

The problem is "connection churn." Your load balancer is dropping all its connections and recreating them every single time a pod is added. It's a flaw in "perfect" balancing.

Here’s the fix Google found.

https://t.co/ASTDbcYzFo

[dipanshu @_dipansh]

Before:
> kubectl get secret ... -o yaml
> Copy c3VwZXJzZWNyZXQxMjM=
> New terminal
> echo "..." | base64 -d
> Repeat.

After:
kubectl decode ...

I got fed up and built the After.

Here's the post on why it's better. https://t.co/UR84mlfrwW

[dipanshu @_dipansh]

We spent years building complex, multi-queue, state-tracking LRU variants to optimize cache eviction.

Then SIEVE comes along, uses a few simple FIFO queues, and beats them all.

Turns out dumber is just smarter. We've been over-engineering this the whole time.

Broke down why:

https://t.co/b3PnwJcurt

[dipanshu @_dipansh]

We've all been forced to choose:

> A simple cache policy (LRU) that's slow.
> A fast cache policy (FIFO) that's dumb.
> A smart cache policy (ARC) that's crazy complex.

What if you could have all three?

Meet SIEVE, a new cache eviction algorithm from NSDI '24.

It's literally simpler than LRU.

The magic?

It's a tiny tweak to CLOCK: when an object "survives" eviction, just... leave it alone. Don't promote it.

Paper: https://t.co/L16Jicr0FK

[dipanshu @_dipansh]

Stop copy-pasting giant values.yaml files.

A well-built custom Helm chart with sensible defaults is the difference between a 10-minute deploy and a 3-hour debug session hunting for a typo.

[dipanshu @_dipansh]

Reducing your Docker image size isn't just about faster CI builds.

It's about security.

A smaller image means a smaller attack surface. Stop shipping your entire OS. Start using multi-stage builds.

[dipanshu @_dipansh]

Newbies use Terraform to create resources.

Pros use Terraform to manage state.

The first just saves you clicks.

The second saves your job when production drifts.

[dipanshu @_dipansh]

Stop staring at your Grafana dashboard hoping for an answer.

Metrics (Prometheus ) tell you what is broken.

Logs (Loki ) tell you why it broke.

You need both, and you need them correlated. Otherwise, you're just guessing.