Here's data from a 20 hour run with a ceramic resonator:
- The average frequency is off by 2021ppm (too slow)
- The frequency varies by ~10ppm per degree-C (hotter = slower)
- for a fixed temperature the frequency stability is +/-8ppm or better (here measurement accuracy was limited by the timing and temperature resolution of the data acquisition)
The temperature variation is only about 5x worse than typical datasheet specifications for a crystal. So as long as the average frequency is calibrated some reasonable timekeeping is possible.
PS. Can anybody suggest a way to timestamp data with greater than 10ms resolution? I discovered that Windows checks the serial port "only" every 10ms. Tsk, so slow! 
[/list]