My peers are starting to tease me about becoming a dbcc shrink* expert. (Ha, Ha I said.) Then, I uncovered another design facet about shrinkfile helping with a SQL Server 2000 case.
The situation: dbcc shrinkfile, with EMPTYFILE, against a file with only 128 pages was taking hours to complete. In the end it is by design but it was not a design I anticipated.
Here is what is taking place during the shrink of this file.
- Shrink setups a GAM scan for the file to be shrunk. Starting at the beginning of the file pages are read … 1, 2, 3, ….
- As each page is processed the page type is used to determine the necessary data movement behavior. For data pages this results in a delete and insert pairing to move the row to another file so EMPTYFILE can complete.
- When a BLOB page is encountered I thought that each fragment would be moved to a new page on another file. This does happen but not as I envisioned. Because the BLOB fragments are not doubly linked, a broader scan is required to locate the root of the chains that have fragments that need to be moved occurs.
- When a BLOB page is located a new scan is established. The new scan is an IAM scan over the TEXT/IMAGE (BLOB) chain for the object found on the original page to be moved.
- Each page in the IAM chain is read and each fragment on the page is reviewed.
- Any fragment found that points to a fragment on the file to be EMPTIED is identified and the associated BLOB chain moved.
So to move a BLOB page may require the entire blob chain to be scanned. In this customers case they have a 1TB database and this BLOB index was ~500GB. So to move the BLOB page lots of I/O and scanning take place.
I have not studied the behavior of SQL Server 2005 in regards to this behavior.
You can read more about this subject from my previous posts about shrink*.
References: shrinkfile, shrinkdb, shrinkdatabase
SQL Server Senior Escalation Engineer