Reconnection Storms Are the Real Load Test

The load problem nobody benchmarks for: a blip disconnects 10k clients; they all reconnect simultaneously:

• TLS handshake surge — CPU-bound; 10k simultaneous TLS negotiations saturate cores
• Subscribe-storm — each reconnect re-subscribes; the auth/rate-limit path becomes the bottleneck
• Thundering herd — exponential backoff without jitter = everyone retries on the same schedule

The mitigations:

• Jittered backoff: 'random(0, 2^retry)' not fixed '2^retry'
• Connection token caching: skip re-auth for recent disconnects
• Pre-warmed listener pools: accept capacity ready before the storm

Stock Picks's infra survived the last major market halt specifically because the reconnect path was designed as the hot path, not the edge case.