gl@luthra_

← back to home

Writing

Notes from building backend and data systems that have to stay fast under real load - distributed systems, observability, performance, and the math that makes them work.

2026 · 06

The Hard Part of Sampling Is the Zero

You can answer a huge aggregate query by reading a sliver of the data and multiplying back. The multiply is the easy part. The hard part is never reporting a confident zero when your sliver missed everything - measure the matches, widen once, then fail closed.

samplingapproximate queriesobservabilitystatistics
2026 · 06

Where Your LLM Tokens Actually Go

You profile CPU and memory. The context window is the most expensive resource in an AI app and almost nobody profiles it. How llmprof flame-graphs every request's tokens, prices the call offline for 1000+ models, and finds the dollars you can reclaim - all locally.

llmobservabilityflame graphscosttokens
2026 · 06

Continuous Profiling: Finding Where the CPU Actually Goes

Metrics say a service is slow; profiling says which function. How always-on sampling profilers work, how to read a flame graph, and what it takes to ingest every runtime into one model.

profilingobservabilityflame graphsperformance
2026 · 06

Anomaly Detection for Alerting, at Nanosecond Cost

How to take metric anomaly detection from 200ms per evaluation to 8.5 nanoseconds and 100k+ timeseries, by moving every expensive thing out of the hot path.

observabilitytime seriesperformancestatistics