Ankit Pandey

SWE 2 @ Amazon Prime Video
• xWalmart, Xalts, Kutumb
• Career Coach @Topmate

[Ankit Pandey @ankitcode99]

Gamedays aren't about finding bugs. They're about building confidence that when traffic spikes, a dependency flakes, or a region goes down, your system degrades gracefully instead of collapses completely.

That's what makes it fun & challenging to work for PrimeVideo 🚀🚀

[Ankit Pandey @ankitcode99]

I run load tests across 200+ services at Prime Video. Here's what breaks first during a gameday.

It's never the service you tested last month.
It's always the dependency nobody thought to check. 🧵

[Ankit Pandey @ankitcode99]

What happens after a game day matters more than the game day itself. Every finding gets triaged: severity, owner, fix timeline.
But the real value is the institutional knowledge. Your team now knows how the system behaves under pressure before a live event.

[Ankit Pandey @ankitcode99]

I run chaos experiments across 200+ services at Prime Video.
The goal is never to break things. It's to find out what breaks before a live event does it for you.
Next: what a real game day looks like from the inside.
Follow so you don't miss it.

[Ankit Pandey @ankitcode99]

Netflix made chaos engineering famous.
Most engineers took one lesson from it: randomly kill servers and see what breaks.

That's not chaos engineering. That's just breaking things. Here's what it actually is and why the difference matters at scale...
🧵👇

[Ankit Pandey @ankitcode99]

Where most chaos programs die:
- They only test in staging (prod has different load patterns)
- No defined steady state, so results are uninterpretable - Teams treat it as a one-time exercise, not a practice
- Findings go into a doc nobody reads Process beats tools. Always.

[Ankit Pandey @ankitcode99]

The right mental model for agent routing:

→ Resolve intent once (Phase 1)
→ Pin the model for the session (Phase 2)
→ Let cache do the work on every turn after
Pay for the routing decision once. Collect savings on turns 2 through 15.

[Ankit Pandey @ankitcode99]

Uber gave 5,000 engineers Claude Code in Dec 2025.

By April 2026, the entire annual AI budget was gone.

Nothing broke. Engineers used it exactly as intended.

The problem: a docstring lookup and a distributed systems refactor cost the same.

[Ankit Pandey @ankitcode99]

A 1.5B model fine-tuned for routing beat Claude 3.7 Sonnet on accuracy.

28x faster. Not a benchmark trick.

General models aren’t optimized for intent classification. A model built for one job destroys a model prompted for it.

[Ankit Pandey @ankitcode99]

A useful production observability stack should help answer 3 questions quickly:

What happened?
Logs
Why did it happen?
Context and traces

Is it happening right now?
Metrics and alerts
ELK solves a major part of this, but observability is bigger than logs.

Next up : @datadoghq

[Ankit Pandey @ankitcode99]

Your production logs are useless if you can't answer one question:

"What exactly went wrong?"

That becomes difficult when thousands of services generate millions of log lines every day.

This is where an observability stack becomes critical.
🧵 👇🏻👇🏻
#application #observability

[Ankit Pandey @ankitcode99]

Elasticsearch is powerful, but indexing everything blindly gets expensive. A production setup usually needs:
• Log retention policies
• Index lifecycle management
• Sampling for noisy logs
• Appropriate shard sizing
• Separate storage for older data
Observability has a cost.