Cutting P99 Latency 1000X During Connection Storms by Hardening ScyllaDB Admission Control

What this post added

This post details the engineering efforts to harden ScyllaDB's admission control to handle connection storms. It describes the connection establishment pipeline from TCP SYN to query-ready, highlighting the introduction of throttling, caching, optimized hashing, and a dedicated service level to reduce P99 latency by 1000x during connection surges. Specific optimizations include improvements to connection throttling mechanisms, caching of connection-related data to reduce redundant work, and optimization of password hashing. The post also mentions the introduction of a dedicated service level to isolate background traffic.