Redis Patterns: Squeezing Maximum Performance and Memory

Almost every high-load service eventually faces the need for instant alerts: if a system stumbles, users must be warned within seconds and the on-call engineer woken up immediately.

I build an internal real-time alerting platform (let’s call it Spectra).

Spectra ingests about 3 million metrics per second but stores only the hot subset — roughly 20 k / s. Every minute it evaluates 26 thousand triggers, and around 2 thousand notifications are sent daily through multiple channels.

All inter-service communication goes through Redis clust…