Sure. I agree various sampling strategies are a good design, as long as they don't add up to much IO. Making that a "user parameter" to be customizable seems

Sure. I agree various sampling strategies are a good design, as long as they don't add up to much IO. Making that a "user parameter" to be customizable seems ideal. What you suggest may make a fine default - if adapted to varying file sizes somehow. E.g., that Nim dups program I mentioned elsewhere could just take a sequence of slices instead of only one..maybe just ignoring out of bounds slices. { And for size clusters where the size < N, screw it - just hash the whole thing, probably. ;-) }