Did you know that Windows Azure billing is done by the clock hour? This means that if you scaled up an extra instance at 3:30pm and scaled down that instance at 4:30pm, you will incur charges for two hours: for 2-3pm clock hour, and for 3-4pm clock hour. Thus, it may make sense to scale down only when the clock hour nears its completion. To take advantage of this, latest version of AzureWatch introduces a new flag to allow for rules to be evaluated only during certain times in a clock hour. Why scale-down in the beginning or middle of the hour, if the full hour was already paid for?
While Microsoft does not guarantee a specific time span within which instances will actually be deallocated after a scale-down request, Windows Azure is pretty good about deallocating instances only within a few minutes time. In light of this, it might be prudent to not let the scale-down window to be too near the end of the clock hour, this can help avoid overruns into a new clock hour.