If we have InnoDB pages there are two ways to learn how many records they contain:

  • PAGE_N_RECS field in the page header
  • Count records while walking over the list of records from infimum to supremum

In some previous revision of the recovery tool a short summary was added to a dump which is produced by the constraints_parser.

But if a page is lost and page_parser hasn’t found it, all records from this page are lost. In other words per-page recovery statistics gives us little idea about whether or not a recovered table is complete.

To cover this flaw a new tool index_check is introduced in the revision 80

As you might know InnoDB stores a table in a clustered index called PRIMARY.

The PRIMARY index is a B+tree structure. It has a nice feature that all leaf nodes have pointers to a next leaf node, so following the pointers we can read whole index in the primary key order. InnoDB extends the structure and stores also a pointer to the previous node. The previous node of the head is NULL as well as the next node of the tail.

Based on this knowledge we can check if we have all elements of the list. If we can get from any InnoDB page to the beginning and the end of the list then our set of pages is complete.

This is exactly what index_chk does – it reads files from a directory with InnoDB pages (that is produced by page_parser) and tries to walk over the list of pages back and forth.

Let’s review an example. I took some corrupt InnoDB tablespace and split it with a page_parser:

Now let’s take some index and check if it has all pages:

Bad news, a page with id 140 is missing!

Indeed the previous page before page#145 is page#140, but it’s missing.

For a cheerful ending let’s check an index that has all pages:

Thus, a table is fully recovered if and only if two conditions are met:

  • Command grep “Lost records: YES” table.dump | grep -v “Leaf page: NO” outputs nothing
  • ./index_chk -f pages-1374669586/FIL_PAGE_INDEX/<inde_id of=”” my=”” table=””> reports OK