“To put that in perspective, the total cpu time consumed by pingora-origin is 40,000 compute-seconds per second.”
One small function that strips internal headers ate 1.7% of pingora-origin’s CPU, about 680 of 40,000 saturated cores. Cloudflare swapped the loop for a custom trie and cut it to 0.34%. More than 90% of lookups are misses, and the trie rejects misses fast. At 35 million requests per second, one tidy loop costs hundreds of cores.