You were absolutely right about file descriptors! Simply increasing the hard limit for the number of open descriptors per process completely solves the issue
Running 30s test @ http://127.0.0.1:4000/token
20 threads and 9900 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 127.40ms 52.07ms 1.47s 94.65%
Req/Sec 3.94k 356.81 8.21k 89.22%
Latency Distribution
50% 123.67ms
75% 132.27ms
90% 142.21ms
99% 290.75ms
2331183 requests in 30.10s, 0.94GB read
Requests/sec: 77451.24
Transfer/sec: 32.06MB
