Hop 1 of 8, Lookup

Recruiter view
Recruiter view

I keep things up

Anshul Singh Chauhan

DevOps/cloud engineer. Production Azure integrations at ATH Infosystems. EKS, Terraform and CI/CD in my own projects.

New Delhi, India Aiming for Japan by 2029–30, studying for JLPT N5.

Short version for recruiters or press R

For recruiters

Recorded
  • Cut a production integration's run time by about 93%, from over 16 minutes to about 1.
  • Took critical vulnerabilities from 12 to 0 in 2 weeks.
  • Restored 21 of 23 overwritten identity values across 17 people, with 0 errors. The other 2 were already right.
Book a time by emailResume (PDF)GitHubLinkedIn

Or write to anshulchauhanworks@gmail.com

See the full journey

Live

You got here in ms, measured just now.This page measures itself in your browser. This site's Terraform, pipeline and scans are public.

Live

An EKS lab with tiers split by taints and affinity. TLS ends at the ALB, and here's what's still exposed.

Recorded

Autoscaling reacts late. Jump the dial and watch latency climb before pods arrive.

Simulated Runs in your browser on the repo's real config. Assumptions are listed below, dated 2 Nov 2026.

This demo runs in your browser and needs JavaScript. The picture shows its resting state.

  • Ready
  • Ready

200: 2 pods' limit 600: 6 pods' limit

Jump past the first mark. Turn it slowly and the pods keep up.
Latency
21 ms
Ready pods
2 of 2
Timed out
0
New pods took
no climb yet
Peak latency
no climb yet
Simulated time
0 s
Clock speed
5× real time
A floor with six pod slots in a row, two of them holding ready pods, and an empty queue pad where traffic enters.A floor with six pod slots in a row, two of them holding ready pods, and an empty queue pad where traffic enters.

This picture shows the resting state. The pods list is live.

Steady. 2 pods are handling 10 requests a second.

What breaks here

  • The CPU target is measured against the 100m request, not the 500m limit. At target, six pods use 360m of a 3,000m combined limit, so the autoscaler runs out of pods long before they run out of headroom.
  • The readiness probe calls /ok, which doesn't check MongoDB. A pod can be Ready with the database down.

Assumptions

From the repo
Minimum 2 pods, maximum 6, CPU target 60%. CPU request 100m, limit 500m. Readiness probe on /ok: first check at 5 s, then every 5 s.
Kubernetes defaults
The controller runs every 15 s with a tolerance of 0.1. Scale-down waits 300 s. Scale-up adds the larger of 4 pods or 100% every 15 s. Pods not yet ready count as 0% on the way up.
Assumed, 2 Nov 2026
Metrics every 15 s, the value in metrics-server's shipped manifest. 5 ms of CPU per request, 20 ms base latency, 12 s for a container to serve, 30 s client timeout, 5m idle CPU, 1 s to create a replacement. MongoDB and node capacity aren't modelled. Latency values depend on these assumptions.

Redis down means 503: the limiter fails closed before the cache can fail open.

Simulated

Production work and postmortems.

Recorded

Every hour, a Lambda checks for idle EC2. This is the exact Slack alert it builds.

Simulated

200 OK. Resume, links, book a time.

Live