I've noticed on MySQLWorkbench that you can choose how to store your indexes before forward engineering your design. The storage types are:

  1. BTREE
  2. RTREE
  3. HASH

Researching this, I found some information that was pretty much over my head, so I'm looking for practical information on what the difference is between these and/or why you should choose one over another.

Also, I have never chosen a storage type before, so I assume MySQL is choosing a default storage type (BTREE?)

3 Answers 3



BTree (in fact B*Tree) is an efficient ordered key-value map. Meaning:

  • given the key, a BTree index can quickly find a record,
  • a BTree can be scanned in order.
  • it's also easy to fetch all the keys (and records) within a range.

e.g. "all events between 9am and 5pm", "last names starting with 'R'"


RTree is a spatial index which means that it can quickly identify close values in 2 or more dimensions. It's used in geographic databases for queries such as:

all points within X meters from (x,y)


Hash is an unordered key-value map. It's even more efficient than a BTree: O(1) instead of O(log n).

But it doesn't have any concept of order so it can't be used for sort operations or to fetch ranges.

As a side note, originally, MySQL only allowed Hash indexes on MEMORY tables; but I'm not sure if that has been changed over the years.

  • Does MySQL support Rtrees?
    – Pacerier
    Commented Jul 5, 2012 at 23:14
  • 4
    yes, they're called SPATIAL INDEX (dev.mysql.com/doc/refman/5.0/en/spatial-extensions.html)
    – Javier
    Commented Jul 5, 2012 at 23:37
  • Cool, thanks =) Are there other structures besides these 3, or planned structures in the near future?
    – Pacerier
    Commented Jul 6, 2012 at 3:40
  • Memory tables support btree indexes as well
    – Amareswar
    Commented Oct 20, 2012 at 20:11
  • 1
    @Amareswar, right. Maybe my answer can be read both ways, but what i meant was that HASH indexes were only allowed on MEMORY tables, not on 'normal' tables.
    – Javier
    Commented Oct 22, 2012 at 14:05

The accepted answer contains good quality of information. index type selection is limited by the storage engine applied to the database table, here is extra information about the index types supported by the storage engines in MySQL. See Table 16.1 Storage Engines Feature Summary in Alternative Storage Engines in MySQL 8.0 documentation

In short , here are the storage engines that support these 3 index types.

  • B-tree index : MyISAM, Memory, InnoDB
  • Hash index : Memory
  • R-tree (Geospatial) index: MyISAM , InnoDB(since MySQL 5.7)

side note:

  • InnoDB internally utilizes hash indexes.
  • This also describes difference between B-tree and hash index in term of usage limit at application level

Spatial indices like R-Tree allow you to query with inequalities on multiple columns efficiently

E.g., with a spatial index you would be able to efficiently query all points that are inside a rectangle such as:

create table t(id integer primary key, x integer, y integer)
select * from mytable where x >= 1 and x < 10 and y >= 2 and y < 20

which has inequalities on both x and y.

The more basic and common B-tree index only allows you to efficiently speed up inequalities in one dimension, even if you try to use composite indices on x and y.

E.g. a x-y composite B-tree index would:

  • efficiently speed up:
    • x = 1 and y = 2: equalities in both columns
    • x = 1 and y > 2: equality on first column plus inequality on the second column
  • very limited speed up:
    • x > 1 and y > 2: inequalities on both columns, including the first
    • x > 1 and y = 2: inequality on the first column
    • y > 2: this is equivalent to x > -infinity and y > 2, so it is just the worse case possible for the compound B-tree index. Note however, this case could be solved efficiently with a B-tree index.

A spatial index would efficiently handle all the above queries however.

The trade-off is that R-Tree indices are more expensive to maintain.

Example of why a composite index B-tree search would be slow

This is well explained at: https://dba.stackexchange.com/questions/249848/why-we-cant-have-more-than-one-inequality-condition-in-mysql-indexing/249909#249909

One way to visualize a B-tree is to see how it would sort the rows. After all, it is a structure analogous to a binary search tree, just with more entries per node to speed up disk access:

enter image description here

Image source

Consider the following x-y composite index, which sorts all rows by (x, y) tuples lexicographically:






And remember that the in-disk ordering might have nothing to do with this. In particular, there could be other columns with completely different values.

Now suppose we want to find:

x > 0 and y > 4

The only speedup operation we can do is a binary search on the above index.

First it uses the index to binary tree searches for (1, 5), which is a speedup over a full scan.

Then it follows the index ordering and grabs every larger y for x = 1. So far so good.

The problem is what happens next.

Note how in this case, there are no y > 4 for x = 2 and x = 4.

However, there is no way to immediately tell that from the index and jump straight to x = 5!

What the search has to do is simply: I'm done with x = 1, so now give me the next larger x. So it keeps traversing the index tree linearly until the next value.

Then it finds the first (2, 2) and it knows: OK, there exists x = 2. Now it has two choices:

  • keep following the index until y = 5 is found
  • binary search down to (2, 5)

Which is better depends on how many rows in total there are in the DB, since the fresh binary search is log(n), so would not be worth if unless there's a large number of values with x = 2 and y < 5.

With either of those, it decides that there were no results for x = 2, so we've just spent some time scanning a bunch of invalid rows.

So it continues the above procedure, basically scanning through the entire index.

x = 4 like x = 2 is scanned uselessly and has no reults.

Then it continues going through the index and finds x = 5, eventually reaching (5, 5), and we finally have some results.

So as we can see, this requires a lot of stepping over ranges that may not contain any results of interest, which is why this compound B-tree index can only produce a limited speedup if we are searching over a large x range with many empty y hits.

An R-tree implementation of a spatial index looks more like this:

enter image description here

Image source

So we intuitively understand that it actually tries to split up a 2D space into a bunch of balanced rectangles, and because of that is able to efficiently query arbitrary rectangular areas.

SQLite minimal benchmark

I'm not very familiar with MySQL, but the concepts should be analogous.

I'm going to create two test databases with 10 million points in a straight line with inclination 2:

  • one with a x-y B-tree index, SQLite's default index
  • one with a x-y R-tree index, SQLite's built-in spatial index

And then let's ask the question:

how many points are there between x >= 1000000 and x < 2000000 and y >= 4000000 and y < 6000000

with a query:

time sqlite3 100mr.sqlite 'select count(*) from t where x >= 10000000 and x < 90000000 and y >= 50000000 and y < 60000000'
  • B-tree: query: 6.4s, index creation time: 4 mins, DB file size: 3.9 GB
  • R-tree: query: 0.7s, index creation time: 30 mins, DB file size: 5.9 GB

The test databases were generated as follows:


from pathlib import Path
import csv
import sqlite3

f = '100mr.sqlite'
n = 100000000
connection = sqlite3.connect(f)
cursor = connection.cursor()
cursor.execute('CREATE VIRTUAL TABLE t using rtree(id, x, x2, y, y2)')
cursor.executemany('INSERT INTO t VALUES (?, ?, ?, ?, ?)', ((None, str(i), str(i), str(i*2), str(i*2)) for i in range (n)))


rm -f "$f"
time sqlite3 "$f" 'create table t(id integer, x integer, y integer)'
time sqlite3 "$f" 'insert into t select value as id, value as x, value * 2 as y from generate_series(0,99999999)'
time sqlite3 "$f" 'create index txy on t(x, y)'

So we see that in the case of this implementation, R-tree search was way faster, but at the cost of a much slower index creation time.

Tested on Ubuntu 23.04, Python 3.11.2, Lenovo ThinkPad P51, SSD: Samsung MZVLB512HAJQ-000L7 512GB SSD, 3 GB/s nominal speed, csvkit==1.0.7, sqlite 3.40.1.


Hash maps vs binary search trees is a classic algorithm question: https://stackoverflow.com/questions/4128546/advantages-of-binary-search-trees-over-hash-tables as these are the two most common ways of implementing a map data structure.