Traefik's throughput suddenly ramps up slower after running for a while

I run Traefik and another container in Docker Swarm both using Docker overlay networks. The other container serves larger payloads (think of streaming data or large static files).

When I start Traefik to serve as a reverse proxy, it performs very well. But after running for some time (non-determinate, usually a few hours), when making new requests, the throughput from Traefik ramps up much slower than bypassing Traefik.

I measured this using curl and captured pcaps, and plotted the throughput as shown in the attachement. The first row shows Traefik just starting, second row was after a while (this time is about 1 hour). The first & second column show throughput via Traefik, the thrid column is bypassing traefik.

I can't find any relevant information and would like to know if anyone else has seen anything similar.

Interesting finding. Is it still happening? Do you use the latest Traefik version?

How do you do the bypass? Same or different port? Same or different (Docker) network?

Can you explain how you tested and generated the graphs to be able to reproduce?