63 posts
Engineering ResourcesAugust 18, 2026by Formation
Failover isn't just "promote a follower" — done carelessly, it produces split-brain and zombie-leader failures worse than the outage it was meant to prevent. Here's what fencing tokens, quorum election, and STONITH actually protect against.
Engineering ResourcesAugust 18, 2026by Formation
Read replicas add read capacity and failover copies, but every write still funnels through one leader—so replicas alone won't fix a write bottleneck or guarantee fresh reads.
Interview PreparationAugust 18, 2026by Formation
Most CAP theorem explanations conflate availability with freshness, CAP consistency with ACID consistency, and treat CP/AP as permanent labels. Here's how to state the trade-off correctly during a network partition.
Engineering ResourcesAugust 10, 2026by Formation
Holding a million WebSocket connections is easy; delivering one message to users scattered across servers is the hard part. Here's why a broker alone doesn't solve fanout, and what a presence registry actually buys you.
Engineering ResourcesAugust 10, 2026by Formation
WebSockets are the most efficient option per message, but idle connections still consume memory and file descriptors — for a large, mostly-idle client population, polling or SSE can end up cheaper overall.
Engineering ResourcesAugust 10, 2026by Formation
"Server push" is a misleading name: browsers behind NAT and firewalls can't be dialed directly, so polling, long polling, SSE, and WebSockets are all client-initiated channels the server holds open differently.
Interview PreparationAugust 3, 2026by Formation
A "Design Spotify" prompt mixes local controls, routine backend work, and hidden distributed-systems problems. Here's how to scope requirements so your limited interview time goes to what actually deserves architectural depth.
Engineering ResourcesAugust 3, 2026by Formation
Horizontal scaling isn't a more mature version of vertical scaling — it's a change in system topology with real routing, state, and connection-budget costs. Here's how to decide which trade-off actually fits your measured bottleneck.
Engineering ResourcesAugust 3, 2026by Formation
"We'd add more servers" is an unfalsifiable scaling answer. Setting latency, throughput, availability, fault-model, and cost targets first turns an assumed 10 million DAU into a checkable peak RPS and in-flight concurrency via Little's Law.
Interview PreparationAugust 3, 2026by Formation
System design vocabulary is easy to reproduce, so a familiar diagram alone doesn't differentiate a candidate; the stronger signal is whether each architectural choice is tied to the actual workload, product risk, and operating constraints. The post walks through a notification-platform prompt to show what that judgment looks like in practice.
