Starkquarry
Independent coverage of infrastructure, networking and data practice.
The Operator's Guide to Load Testing
September 14, 2026
Synthetic load generation consistently fails to replicate production outages because artificial traffic profiles lack real-world dynamics. Sending constant, uniformly spaced queries against a solitary URI tests that specific route in isolation. Real crashes stem from synchronized client retry spikes following a brief hiccup or morning crawler sweeps hitting unprimed caches.
Capturing and re-executing actual production traffic patterns—incorporating true endpoint ratios, user pauses, and diverse headers—delivers far more predictive value than inflating synthetic concurrency. Pay keen attention to packet arrival clustering: production traffic arrives in sharp surges rather than tidy Poisson curves, quickly swamping buffers.
A Practical Guide to API Rate Limiting
August 16, 2026
Rate limiting is one of those features everyone agrees is important and almost nobody designs deliberately. The naive per-IP token bucket works until you meet carrier-grade NAT, where a hundred thousand mobile users share a handful of addresses and your limiter punishes them as a single entity.…
Structuring DNS for Reliability
July 26, 2026
Domain resolution outages produce an awkward paradox: server instances remain perfectly healthy while clients perceive complete downtime. Insulating against platform-level registrar outages demands secondary authoritative DNS providers operating over separate routing infrastructure.…
Practical Notes on Postgres Connection Pooling
September 16, 2026
Each incoming PostgreSQL connection instantiates an isolated operating system process with dedicated memory overhead, necessitating a dedicated pooling proxy in growing deployments. Classic session pooling acts transparently, whereas high-throughput transaction pooling frequently disrupts stateful f…
Cache Invalidation Patterns That Survive Traffic
August 7, 2026
Everyone quotes the two hard things joke; fewer people ship caches engineered against stampede. The pattern that matters most is request collapsing: when a hot key expires, exactly one worker rebuilds while everyone else serves slightly stale data, which turns a cache miss from a database incident i…
More reading
- Reading Latency Percentiles Without Fooling Yourself — Operations, August 7, 2026
- Data Residency Basics for Global Teams — Compliance, May 10, 2026
- A Field Guide to Graceful Degradation — Engineering, September 20, 2026
About us
Our contributors have spent years on-call for large platforms. This site collects the playbooks, postmortems and reference material we wish someone had handed us earlier.