Mysql indexes

From wikinotes

Indexes use various datastructures to store data to prevent full scans of all rows in the table.




Jeremy Cole: InnoDB BTree


An index is a search-optimized datastructure. By default, a BTree.
A query can only use one index at a time, and the index is chosen by the optimizer.

Internally, the structure of a query actually uses two indexes.

  • The clustered index is the entire table contents, indexed by the primary key
  • The secondary index is what your optimizer may choose.
    It optimizes a search on a non-primary-key index. The identified rows they use the clustered-index to retrieve the table contents.


# create index
CREATE INDEX index_users_on_created_at ON (users);

# show indexes

# force FROM to use index
FROM users FORCE INDEX (index_users_on_created_at)

# force JOIN to use index
JOIN projects FORCE INDEX (index_projects_on_user_id)
ON ...

Index Types


A regular index, on a single column. Defaults to BTREE.
NULL values are indexed for fast searching, NULL value rows are grouped together at the start of the tree.

Multi-Column Index

A multi-column index is a regular index,
where the values of multiple columns are concatenated together to form one string.

Index Column Order

The order of the columns is significant,
there are debates over the optimal ordering, generally least-specific to most-specific is correct.

- (unspecific -> specific)   (ex. id, province, country)
  - arguably most scalable. you can reuse the same index when breaking up query into batches.

- (specific -> unspecfic)
  - creates larger decision-tree.
  - more work on initial query, making it difficult to decompose into multiple queries.
  - less likely to select on most-specific rows alone

# but perhaps it is best to consider which columns will be used without the others (likely least specific)
# avoiding the write-cost of adding a row-change to 2x indexes.

Single-Column Queries

You can take advantage of a multi-column index for single-column queries,
if your query includes items in the leftmost-to-rightmost order.

For example

# we have a multi-column index on (country, province, city)

SELECT ... WHERE country = "CA";                     # CAN use index, query on leftmost column
SELECT ... WHERE country = "CA" AND province = "ON"  # CAN use index, query on each of the two leftmost columns

SELECT ... WHERE proince = "ON";                     # CANNOT use index, queries 2nd column without 1st

Beware of Range Queries

As soon as you hit a fork in the index, you can no longer use it for later columns.

For example

INDEX (country, province, city)

SELECT .. WHERE country IN ("CA", "US");                                  # uses index (1st-column)
SELECT .. WHERE country="CA" AND province IN ("NB", "NS")                 # uses index (1st, 2nd column)
SELECT .. WHERE country="CA" AND province IN ("NB", "NS") AND city="Foo"  # uses index, but only (1st, 2nd column)

NULL in a multi-column index

NULL IS represented within a multicolumn index.
NULL is an empty-string within the concatenated set of rows that forms the multicolumn-index value,
But a bitmask (over each column?) indicates if it is NULL or "".

Index Datastructure Types


A binary tree. Default for persisted storage engines.

Optimize Queries that are

- sorted
- match on field prefixes (but not suffixes)
- range bound (less than 10, between C and F)

But not queries

- match on field suffixes


A hash table. Default for the memory engine.
hash-collisions are accounted for, but each key with the same hash will need to be checked, making the query more expensive.
Indexes have a small memory footprint.

Optimize queries that are

- based on the full value (ex. 'IN, NOT IN, =')

But not queries that

- match on only a single key from a multi-key index
  (since both keys are hashed together)
- are range bound. (ex. less than 10, betweeen C and F)
  (hash-tables are un-ordered, so range queries are not optimized)


Fulltext indexes are optimized for finding keywords within blocks of text.
They are used when the MATCH AGAINST ".." operator is used.
It may still be beneficial to have a second index with a different type on the same column.