The Reason It Is Adopted And The Reason It Exists
Write-back and write-behind are two names for the same pattern, and both appear in the literature and in library documentation. This lesson uses write-behind throughout, because the behaviour it names is the one that matters: the write goes to the cache, and the store is updated behind it.
The previous lesson treated it as write-through with the acknowledgement moved earlier, which is the usual description and is where the discussion normally stops. The win is real and it is the small half.
A write-behind cache holds a dirty set and flushes it on a schedule. If the same key is written ten times between two flushes, the store receives one write.
Writing the same key again replaces what is already there, so only the last value is ever flushed, and the reduction is exactly the number of repeats inside the interval.
What A Flush Interval Buys And What It Costs
A flush every one second cut store writes from 400,000 to 167,132, which is 2.39 times less work, in exchange for one second of exposure.
The curve is steep and then flat. Most of what buffering can do for you is done within a second, because a Zipf head is rewritten within a second and the tail is not rewritten at all however long you wait.
