From another topic discussion I found a link to this one(posted by @axelson) and there Cowboy may be the culprit:
You may want to check if it relates with your problem or not.
Exadra37
derek-zhou
zhangzhen
After one-week investigation that includes testing and tracing with Extrace, I found that the Cowboy http server hung from a random time point due to an unknown reason, the subsequent requests were rejected in somewhere whether it is a GET request or a POST request, even a GET / request for the home page. The “This site can’t be reached” message was shown in a browser. Tracing indicates that handling stops before going into the router. Such hanging lasts in some cases for a short time such as 5 mins, in other cases for a long time such as 40 mins. After hanging the system works again. What magic can break the Cowboy down? Is it ETS?
1
Exadra37
Just a wild guess…
If swap is enabled in your Linux server I would suggest you to try to disable it or at least monitor if when this issues happen your server is using swap.
I would imagine that when swap is being used that things just run much slower, not to hang, but we see so many weird things happening in our life’s as developers that in my opinion it will not hurt to try to explore the swap possibility.
0
zhangzhen
Thank you for your advice. br-tower is developed and tested on my Hackintosh ( Intel NUC Hades Canyon).
0
Exadra37
Mac’s also seem to use swap files, thus I guess that an Hackintosh will also use it?
Disable/Enable swap on Mac:
0
derek-zhou
If I were you, I will put all instrumentation methods I can think of on the system, for both the OS and BEAM. Then analysis when this thing happen, is there anything unusual also happening. Then post here with your finding.
Tricky bugs are rarely solved by taking wild guesses and shooting in the dark. You need data and evidences, and lots of it.
4
zhangzhen
The culprit is max open files per os process. It is set to 256 by default on MacOS Catalina. When it is changed to 4096, there is no hanging anymore. But I cannot identify the location where the cow_boy server waits for any available file descriptor to establish the TCP connection.
2
hauleth
Because it isn’t done by Cowboy, but by Ranch. And it probably doesn’t wait in Erlang code, it just call the OS to provide one for them via gen_tcp:accept/2.
5