Current ROE

Data is always possibly outdated in distributed systems (including multi-threaded applications):

* Dont use atomics, they are slow.
* Use locks, they are fast, especially when uncontended.
* Hold locks as shortly as possible: Lock, copy, unlock, process, propagate, progress
* Strive for thread local data (not TLS) which requires no locking, use inter-thread events to propagate information.
* Process events best-effortly
* Dont demand exactness, but require progress. Try to recover.
* Use event version barriers to throw away outdated async data

User experience:
* Have reasonable defaults
* Require as little configuration as possible
* Provide sensible logging