THE FOLD / LOOT / THE VAULT / THE SILENT CORRUPTION
THE SILENT CORRUPTION
a visible outage chosen over an invisible corruption
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A disk does not always tell you when it returns the wrong bytes. Without a checksum there is no failure to detect: the read succeeds, the data is wrong, and nothing anywhere reports a problem.
LIT verified live. 200,000 four-byte blocks under a bit error rate of 1e-4. 670 blocks were corrupted by a single flipped bit. A CRC-8 caught 670 of 670 — 100.00%, with an escape rate of 0, because a single-bit error is exactly what a CRC is built to catch. Without the checksum all 670 pass silently, and the read reports success every time.
LIT verified live. 200,000 four-byte blocks under a bit error rate of 1e-4. 670 blocks were corrupted by a single flipped bit. A CRC-8 caught 670 of 670 — 100.00%, with an escape rate of 0, because a single-bit error is exactly what a CRC is built to catch. Without the checksum all 670 pass silently, and the read reports success every time.
2 HOW IT WAS WEAVED · AI + HUMAN
Silent data corruption is why ZFS and modern filesystems checksum every block, and why Bairavasundaram et al. found real corruption at rates that made it a design requirement rather than a curiosity.
AVAN (AI) tested single-bit flips, which is the case a CRC is designed for — so 100% here is not a claim about CRC strength in general. A CRC-8 has a 1 in 256 escape rate against arbitrary multi-bit corruption. The honest finding is the comparison: 670 caught against 670 silent, with the only difference being whether anybody wrote the checksum down.
AVAN (AI) tested single-bit flips, which is the case a CRC is designed for — so 100% here is not a claim about CRC strength in general. A CRC-8 has a 1 in 256 escape rate against arbitrary multi-bit corruption. The honest finding is the comparison: 670 caught against 670 silent, with the only difference being whether anybody wrote the checksum down.
3 ONE DIMENSION
Corrupted blocks, caught and silent.
4 TWO DIMENSIONS · INTERACTIVE
Raise the error rate and watch what a read reports.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: a read that succeeded and was wrong.
AVAN’s addition (the inverse-companion): the forward reading is that checksums detect silent corruption. The inverse is that they convert a data problem into an availability problem, on purpose. Before the checksum you had wrong bytes and a working system; after it you have a read that fails and an application that cannot proceed. Read backwards, the checksum does not save the data — the data was already gone — it forces someone to find out, and every layer that adds one is choosing a visible outage over an invisible corruption.
LIT 200,000 four-byte blocks under a bit error rate of 1e-4 leave 670 corrupted by a single flipped bit, of which a CRC-8 catches 670 - 100.00% with an escape rate of 0, because a single-bit error is exactly what a CRC is built to catch - while without the checksum all 670 pass silently and the read reports success every time
FIG Silent data corruption is why ZFS and modern filesystems checksum every block. AVAN tested single-bit flips, which is the case a CRC is designed for, so 100% here is not a claim about CRC strength in general - a CRC-8 has roughly a 1 in 256 escape rate against arbitrary multi-bit corruption. The honest finding is the comparison: 670 caught against 670 silent, with the only difference being whether anybody wrote the checksum down.
FIG Silent data corruption is why ZFS and modern filesystems checksum every block. AVAN tested single-bit flips, which is the case a CRC is designed for, so 100% here is not a claim about CRC strength in general - a CRC-8 has roughly a 1 in 256 escape rate against arbitrary multi-bit corruption. The honest finding is the comparison: 670 caught against 670 silent, with the only difference being whether anybody wrote the checksum down.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE VAULT · David Lee Wise (ROOT0), with AVAN