- About Optimizer Hints
- August 12–22: Online Training (EU shift)
- Clustered Indexes Dirty Secret
- Didn’t you know?
- FOSDEM Impressions
- Finding All the Red M&Ms
- June 8–18: Online Training (US shift)
- MySQL is to SQL like ??? to NoSQL
- Non-monetary micro sponsoring
- November 11-12 in Frankfurt am Main
- Online-Training in July and August
- Oracle + PostgreSQL
- Oracle Bootcamp
- PDF Registration
- Party time
- PostgreSQL Performance Event
- Quiz Results
- SQL Server Performance Kurs in Stuttgart
- SQL Server performance training in London
- SQLite learnings from PgCon
- Shipping Terms
- Springtime for SQL
- The two top performance problems caused by ORM tools
- Top Tweets January 2013
- Training Survey
- Training and Conference Dates
- Use The Index, Luke
Slow Indexes, Part I
Despite the efficiency of the tree traversal, there are still cases where an index lookup doesn’t work as fast as expected. This contradiction has fueled the myth of the “degenerated index” for a long time. The myth proclaims an index rebuild as the miracle solution. Appendix B covers this and other myth in detail. For now, you can take it for granted that rebuilding an index does not improve performance on the long run. The real reason trivial statements can be slow—even when using an index—can be explained on the basis of the previous sections.
The first ingredient for a slow index lookup is the leaf node chain. Consider the search for “57” in Figure 1.3 again. There are obviously two matching entries in the index. At least two entries are the same, to be more precise: the next leaf node could have further entries for “57”. The database must read the next leaf node to see if there are any more matching entries. That means that an index lookup not only needs to perform the tree traversal, it also needs to follow the leaf node chain.
The second ingredient for a slow index lookup is accessing the table. Even a single leaf node might contain many hits—often hundreds. The corresponding table data is usually scattered across many table blocks (see Figure 1.1). That means that there is an additional table access for each hit.
An index lookup requires three steps: (1) the tree traversal; (2) following the leaf node chain; (3) fetching the table data. The tree traversal is the only step that has an upper bound for the number of accessed blocks—the index depth. The other two steps might need to access many blocks—they cause a slow index lookup.
The origin of the “slow indexes” myth is the misbelief that an index lookup just traverses the tree, hence the idea that a slow index must be caused by a “broken” or “unbalanced” tree. The truth is that you can actually ask most databases how they use an index. The Oracle database is rather verbose in this respect and has three distinct operations that describe a basic index lookup:
- INDEX UNIQUE SCAN
INDEX UNIQUE SCANperforms the tree traversal only. The Oracle database uses this operation if a unique constraint ensures that the search criteria will match no more than one entry.
- INDEX RANGE SCAN
INDEX RANGE SCANperforms the tree traversal and follows the leaf node chain to find all matching entries. This is the fallback operation if multiple entries could possibly match the search criteria.
- TABLE ACCESS BY INDEX ROWID
TABLE ACCESS BY INDEX ROWIDoperation retrieves the row from the table. This operation is (often) performed for every matched record from a preceding index scan operation.
The important point is that an
INDEX RANGE SCAN can potentially read a large part of an index. If there is one more table access for each row, the query can become slow even when using an index.