Every application is subject to an IP based concurrent request throttle. If a single IP is making more than 30 requests a second, new requests will be dropped. The exact ban period is subject to change, but will be on the order of 30 seconds to a few minutes typically.
And it is highly likely that you see the IP throttle in effect and not the dynamic throttle. The IP based throttle is in effect on the HA-Proxy but that article is from 2010 so I assume by now and by my own experience certain routes are more likely to get you IP banned then others. See How many requests are too many? and The Complete Rate-Limiting Guide, specially the explanation about the Chat rate limit, which seems to work similar to the API.
Basically you have to design and code for two throttle cases
The IP based throttle, kicks in at will
The dynamic throttle, covered by the backoff parameter (assuming you haven't filtered it out)
write_failedthat are relevant.