dnsimple / erldns / 15780289435 / 1
84%
main: 84%

DEFAULT BRANCH: main
Ran
Files 34
Run time 1s
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 76.463% (+0.6%) from 75.862%
15780289435.1

push

github

web-flow 
Backpressure (#240)

* Add change to changelog

* Acceptor pool is not using any feature of wpool so make it more lightweight

* Correct Ranch parallelism

This flag is per configured connection supervisor and not global,
hence we don't need to multiply here by parallelism.

The value chosen is by the following calculation: if a light request
takes 1ms, and we queue N of them in a scheduler, the last one would be
delayed by N ms, which would probably be about to hit the ingress
timeout. So that's the maximum number of requests per scheduler allowed.

* Introduce mechanism to ignore expired requests

Every request, when picked, will compare the current timestamp against
the originally emitted one, and if the timeout has expired (by default,
the configured udp and tcp ingress timeouts), the worker will simply
ignore this request entirely and not even answer, as most likely the
client has already considered this request and timed-out.

* Introduce load-shedding mechanism

Listeners will get packets from the network in batches, and before
taking the next batch, they will check the current state of the
schedulers utilisation. When utilisation is above 90%, listeners will
probabilistically delay taking the next batch, leaving it to the
kernel's network stack to start dropping packets entirely.

* Introduce an API to hard restart all listeners

1150 of 1504 relevant lines covered (76.46%)

2570.64 hits per line

Source Files on job 15780289435.1