With a large projects repository withh hundreds of simulateanous builds we encountered several times builds lost , impossible to stop or restart withoutt any kind of message in the logs.... Default docker-java implementation uses Netty  as a communication layer  and Netty handling for timeouts requires tricky code.....

So I tried to drop the default transport mode to use Apache HC5 instead . The results (in production) are quite good and we do not loose builds anymore.

But our code is not conformant to quality level as expectedby the Jenkins devs & users.