ep-engine deadlock in write-heavy DGM cases

This is not a new issue, we discussed it many times.

In extremely write-heavy cases we overload servers, memory usage reaches 95% of bucket quota, we eject all replica items... eventually system becomes unusable.

I'm creating this ticket because of XDCR. In 3.0 we can achieve very high throughput of XDCR operations, throttling it doesn't make sense. According to PM team some "users" deploy XDCR within the same data center so this is quite realistic scenario.

Feel free to close this ticket as duplicate of existing bugs. Though I didn't manage to find anything well-defined.

added: this issue returns for most DGM settings < 20%, and even moderate update ops of 3K/node. only current SE/PM advice is to increase memory to avoid, but the value of DGM is lost for many expected conditions.