http://www.sqlservercentral.com/blogs/steve_jones/2007/06/05/dbcc-internals/ Printed 2017/01/20 01:18PM
- needs consistent view of the db. Needs to do this without locking. In
SQL Server 2000 log analysis is used. Almost everything is logged and
once the database is read, the log is read and inside checkdb, recovery
is performed to reconcile the differences. If there was a lot of
activity, this could take a long time. While the data portion is
multi-threaded, the t-log analysis is single threaded. There are
problems because not everything is logged. So it guesses in some cases.
Could cause false positives.
In SQL Server 2005, this was changed and they use COWs (copy on write).
An internal snapshot is generated and used for a consistent view. Still
with a high load, there are problems (just like a snapshot). Tempdb is
not used (nor can you create a snapshot on tempdb).
- no online checks on tempdb. recovery cannot be run and a snapshot cannot be created.
- only works if you have databases on NTFS.
- internal snapshot is an alternate file stream on each db file, but it will not show up in explorer.
- any space used in the alternate stream is returned to the drive after checkdb.
- If you run checkdb on a snapshot (if you're low on space), no alternate stream/snapshot is created.
- What does it do?
- primitive checks on critical tables (5 tables)
- works by getting to the leaf level of each of
these tables. Read every page, no cycles in the linkages, page audit of
- if there's a problem it's done.
- allocation checks (= checkalloc) - checks consistency of
allocation maps. Ensures that the maps are accurately stored. IAM, GAM,
etc. ensure no two tables have the same map set, ensure that the GAM
and the IAM match up correctly. Check PFS pages - is page allocated, is
allocation page, etc.
- so far this is fast
- If running repair, it's run on allocation problems.
- logical system table chceks and repairs
- eq of checktable on these. Works in parallel
- check indexes are correct, b-trees are correct, page audit, etc.
- If page checksums are turned on, this is checked.
- logical checks of user tables (and repairs)
- this runs in parallel as well
- service broker checks
- metadata checks (check catalog, wasn't included in ss2k)
- indexed views and XML index checks (and repairs). Determines
if the content of the indexed view matches the definition of the view.
Can determine if there are extra views or missing rows. Makes new index
to check. Similar thing with XML.
You can check percent complete column to figure out which stage and progress in the DMV (dm_exec_requests).
- only first two checks
- then reads every page to do checksum checks and page audits.
Leaf Index Checks
- Page audit (buffer pool does torn page or checksum checks). Can eat a checksum failure and continue on. Reports the issues.
- look at page header, check that data is what is expected for that page.
- Record audit - gross audit of structure of all records. Are offsets correctly set.
- B-tree check
- for index and leaf pages.
- check linkages
- are key ranges correct?
- partitioning - is the record correct for the partitioning function.
- key ordering - records on the page are not stored in order.
They are stored where they fit, so there's a key ordering store for the
records on a page.
- per record checks
- complex columns - checks the pointer to the values. ensures that what's stored in the row is what's stored elsewhere.
- computed columns - make sure persisted values are correct according to the formula.
- Data purity - not all types have range checking. Datetime has
minutes after midnight. Must be < 1440, but the stored value could
be higher in theory. So checks are performed to be sure the stored
value is within the range.
- 1 to 1 mapping - ensures that a nonclustered index maps to one base table row.
- Page counts
- counts in header are correct.
No blocking of the table on SS2k and beyond.
How often do you run this? It depends.
checks could approach n<sup>2</sup> complexity. Fastest way
is N * log(n). Rather than following links, as pages are read, various
facts are generated about each page.
- A (actual) - we read A
- A (sinbling) b - who does A point to
- P (parent) A - when A is read.
Where it's an end page, nulls are generated, For parent pages, a
psuedo-parent is generated. A, B, O, are really the page IDs, index
IDs, etc. these are sorted and then checked. Every node must have 3
facts (can be null) or there's a missing link. If there are too many
items, like 2 parents, then there's a linkage issue. This is done so
pages can be read in allocation order, which is the fastest way to read
them. Other checks use a similar method with different facts.
Index matching. It's getting complex, but the idea is that a
bitmap is generated. As the table (heap or clustered index), is read, a
hash is done and a bit is flipped. As the index is read, it's hashsed
and it should map up to the same value and the bit is flipped back. So
as long as a bit is flipped twice, it's good. so it digs through and
can reuser the bitmap for multiple tables. As long as everything flips
twice, it's ok. If at the end they do find a set bit a "deep
drive" is performed, which is very time consuming to find the
corruption. This first check is 30% of the resources of the deep dive
and is fast. So if checkdb starts to take a long time, it's bad.
Before SP2, you couldn't tell a deep-dive was being run. Post SP2, an error message is now logged.
- purpose is to make the db consistent. NOT protect data. Data is not
necessarily deleted, but the repairs are efficient, fast repairs and
data can be lost.
- mor eintrusive repairs ar edone first. Like allocation map for an
index repaired before a missing row. The more drastic ones often fix
less drastic ones.
- Not everything is repaired, or known. You have to check.
- Not feasible to make it online. Too complex.
CHECKDB doesn't run faster after index rebuilds.
If you run CHECKDB and find errors, does it run again with the repair
option? Yes, it runs twice. Because something could have changed.
Space required ro run checkdb.
- facts require more space, so tempdb space is needed to store
facts. Run with_estimate_only, to get an idea of the space for facts.
can internally do single index row inserts and deletes.
If a value in a row is corrupt, the row can be deleted.
If a pointer is corrupt, the entire page can be deleted.
to be safe, you can start a transaction, run repair, and if data is deleted, you can rollback if needed.
checkdb does not respect key relationships (PK, FK). It is
structurally complete, but you could have orphaned rows. Need to make
Some things are unrepairable
- system table clustered index problems
- PFS pages
- Data purity errors.
Beware of third party kernel filters not working with CHECKDB that doesn't handle alternate streams.
Not recommended to have anti-virus on SQL Server because it can interfere with IOs