Well, they say it's a very well supported use case to use it from multiple threads. I've also tried to run it with much higher load than it's usual in prod and hasn't been able to reproduce locally. It seems that it only happens in this weird Azure environment - nobody else ever reported that issue.