0

What is an InnoDB Page checksum exactly? Does a Page checksum only detect issues with underlying storage when writing or reading to a page/block?

3

(I think this is what you are asking about.)

InnoDB "blocks" are 16KB. But most disk subsystems work in smaller units -- 4KB or 512 bytes. For InnoDB data to remain intact, the disk needs to write all 16KB as a unit -- either all written or none. What can happen is that the pieces of the 16KB are written one after another, and the power fails part way through. This causes a "torn page" ("page" referring to the 16KB block).

To recover from a torn page, InnoDB does two things. It checksums each block in order to discover it, and it uses a "double write" for recovery. The block (or at least certain critical blocks) are written twice - to some relatively constant spot, then to the desired location on disk (in the data or index).

When recovering from a crash, the "double write buffer" is checked to see if it leads to a "torn page"; in which case, it is repaired.

The double write, as its name implies, is costly. (I have no metrics on how costly; I suspect the cost depends heavily on HDD vs SSD and RAID controllers.) Turning it off is a way to gain some more speed, but with risk. FusionIO was (it's been bought out) the only drive manufacturer who guaranteed 16KB atomic writes; I hope others have added this feature.

RAID with Battery Backed Write Cache should make the double write virtually zero cost.

A different checksum... The Percona Toolkit uses a "checksum" for data. Since rows are not necessarily laid out identically between Master and Slave, checksumming the files is not useful to seeing if the tables match. I suspect it involves reading the rows in a repeatable order, and checksumming each row or clump of rows.

"rsync" does a similar thing, but at the file level.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.