This is expected behaviour as of RC-1.
The use case for the Clock interface was to allow an extra performance optimization for users who want to squeeze the last drop of performance out of the logging subsystem. The trade-off is that the faster CachedClock implementation shows jumps of 10-16 milliseconds, so you lose accuracy in the log timestamps.
However this optimization only becomes visible after other bottlenecks have been removed. The performance bottleneck of Async appender is on the queue, and optimizing the timestamping mechanism would not make a visible difference in performance.
If your use case for using a custom Clock is not performance related, can you clarify what it is you are trying to achieve?
Perhaps it would make sense to use the Clock interface consistently everywhere in Log4j.