Went through the Datadog senior SWE loop in May 2026, targeting their infrastructure/agent team. Sharing because I couldn't find enough specifics when I was prepping.
The system design round was 60 minutes. One interviewer, pretty conversational, not adversarial. The prompt was open-ended: something along the lines of design a distributed metrics collection system that can handle millions of time-series data points per second. Very on-brand for Datadog, not a surprise.
What they actually cared about: Ingestion layer: how do you handle bursty writes, what's your buffer strategy, do you partition by metric name or by host or something else Storage backend: columnar vs row-based, they pushed back on anything generic like 'use Postgres'. You need to know why time-series databases (think InfluxDB or something similar to what Datadog uses internally) make sense here Cardinality explosions: this is their actual problem. If you don't bring up high-cardinality tags (like container IDs in ephemeral infra) on your own, they'll push you there and it's a big signal Query path: how does a customer query P99 latency for the last hour across 10k hosts without melting the db
I had a decent answer for ingestion and storage but got caught flat-footed on the cardinality question. The interviewer was clearly an expert and you could tell when you said something off. They weren't rude, just precise.
One thing I noticed: they didn't care much about the whiteboard being polished. Rough boxes and arrows were fine. The depth of your verbal reasoning mattered a lot more.
Time management was a challenge. I spent too long on the write path and barely touched on observability of the system itself (ironic given it's Datadog). Build in time to talk about monitoring your own design.
Overall difficulty: harder than the average senior system design I've seen. They build this stuff. You can't fake it.