Tells a file whose content shifted from one that was edited in place.
Fixed blocks cannot follow an insertion. Insert two megabytes at the front of a file and every block after the insertion moves, so the file looks entirely rewritten even though almost all of its content is unchanged. Measured on a 256 MiB file: a 2 MiB in-place edit moves 0.81% of the blocks and a 2 MiB insertion moves 60.79%, at every block size from 64 KiB to 4 MiB.
That gap is what makes the two cases separable. Above the threshold the file is shifting rather than being edited, and fixed blocks are the wrong tool for it - so the finding is raised and an administrator is offered the two things that would work: whole-object storage, or the content-defined chunker at its measured 4.29 MB/s.
Nothing is changed automatically. A shift is not a fault and the right answer depends on the file type: an appended log wants the chunker, a re-encoded video wants whole-object storage, and a database dump might want neither. Choosing costs storage or CPU, and that is an operator's call.
DEFAULT_THRESHOLD = 0.3
Share of moved blocks above which the content is taken to have shifted.
30%, chosen to sit in the wide gap between the measured in-place case at 0.81% and the measured insertion case at 60.79%. A file type whose real distribution straddles this is exactly what calibration is for.
__construct(float $threshold = \self::DEFAULT_THRESHOLD, int $minimumBlocks = \self::MIN_BLOCKS) : mixed
Constructs a detector.
| float | $threshold | Share of moved blocks that counts as a shift. |
| int | $minimumBlocks | Fewest blocks before the ratio is judged at all. |
When the threshold is not a fraction above zero.