We are still seeing some issues with this. Now we are trying to enable beta1 with fastRTPS. With a similar node to SecretaryBirds, publishing time at rmw_qos_profile_sensor_data, and the subscribing node subscribing at the same qos, we see the subscribing node using 100% cpu. Ths subscriber is calling spin(). If we try using spin_some or spin once we see the cache fill up and errors. Even when the loop rate is 2000 Hz.
We found a way to get the cpu usage down by subscribing and publishing at qos_default. However, now the publisher slowly climbs up to 100% cpu (takes about 30 seconds to a minute on my machine).
Also, please note that even if we only publish the time 1x per second, we still see the subscriber at 100% cpu.