Important

You are viewing the archived documentation of TiDB, which no longer receives updates. It is recommended that you use the latest LTS version of the TiDB database.

Index Selection

Reading data from storage engines is one of the most time-consuming steps during the SQL execution. Currently, TiDB supports reading data from different storage engines and different indexes. Query execution performance depends largely on whether you select a suitable index or not.

This document introduces how to select an index to access a table, and some related ways to control index selection.

Access tables

Before introducing index selection, it is important to understand the ways TiDB accesses tables, what triggers each way, what differences each way makes, and what the pros and cons are.

Operators for accessing tables

Operator	Trigger Conditions	Applicable Scenarios	Explanations
PointGet / BatchPointGet	When accessing tables in one or more single point ranges.	Any scenario	If triggered, it is usually considered as the fastest operator, since it calls the kvget interface directly to perform the calculations rather than calls the coprocessor interface.
TableReader	None	Any scenario	It is generally considered as the least efficient operator that scans table data directly from the TiKV layer. It can be selected only if there is a range query on the `_tidb_rowid` column, or if there are no other operators for accessing tables to choose from.
TableReader	A table has a replica on the TiFlash node.	There are fewer columns to read, but many rows to evaluate.	Tiflash is column-based storage. If you need to calculate a small number of columns and a large number of rows, it is recommended to choose this operator.
IndexReader	A table has one or more indexes, and the columns needed for the calculation are included in the indexes.	When there is a smaller range query on the indexes, or when there is an order requirement for indexed columns.	When multiple indexes exist, a reasonable index is selected based on the cost estimation.
IndexLookupReader	A table has one or more indexes, and the columns needed for calculation are not completely included in the indexes.	Same as IndexReader.	Since the index does not completely cover calculated columns, TiDB needs to retrieve rows from a table after reading indexes. There is an extra cost compared to the IndexReader operator.

Note

The TableReader operator is based on the _tidb_rowid column index, and TiFlash uses a column storage index, so the selection of index is the selection of an operator for accessing tables.

Index selection rules

TiDB provides a heuristic rule named skyline-pruning based on the cost estimation of each operator for accessing tables. It can reduce the probability of wrong index selection caused by wrong estimation.

Skyline-pruning

Skyline-pruning is a heuristic filtering rule for indexes. To judge an index, the following three dimensions are needed:

Whether it needs to retrieve rows from a table when you select the index to access the table (that is, the plan generated by the index is IndexReader operator or IndexLookupReader operator). Indexes that do not retrieve rows from a table are better on this dimension than indexes that do.
Select whether the index satisfies a certain order. Because index reading can guarantee the order of certain column sets, indexes that satisfy the query order are superior to indexes that do not satisfy on this dimension.
How many access conditions are covered by the indexed columns. An “access condition” is a where condition that can be converted to a column range. And the more access conditions an indexed column set covers, the better it is in this dimension.

For these three dimensions, if an index named idx_a is not worse than the index named idx_b in all three dimensions and one of the dimensions is better than idx_b, then idx_a is preferred.

Selection based on cost estimation

After using the skyline-pruning rule to rule out inappropriate indexes, the selection of indexes is based entirely on the cost estimation. The cost estimation of accessing tables requires the following considerations:

The average length of each row of the indexed data in the storage engine.
The number of rows in the query range generated by the index.
The cost for retrieving rows from a table.
The number of ranges generated by index during the query execution.

According to these factors and the cost model, the optimizer selects an index with the lowest cost to access the table.

Common tuning problems with cost estimation based selection

The estimated number of rows is not accurate?
This is usually due to stale or inaccurate statistics. You can re-execute the analyze table statement or modify the parameters of the analyze table statement.
Statistics are accurate, and reading from TiFlash is faster, but why does the optimizer choose to read from TiKV?
At present, the cost model of distinguishing TiFlash from TiKV is still rough. You can decrease the value of tidb_opt_seek_factor parameter, then the optimizer prefers to choose TiFlash.
The statistics are accurate. Index A needs to retrieve rows from tables, but it actually executes faster than Index B that does not retrieve rows from tables. Why does the optimizer choose Index B?
In this case, the cost estimation may be too large for retrieving rows from tables. You can decrease the value of tidb_opt_network_factor parameter to reduce the cost of retrieving rows from tables.

Control index selection

The index selection can be controlled by a single query through Optimizer Hints.

USE_INDEX / IGNORE_INDEX can force the optimizer to use / not use certain indexes.
READ_FROM_STORAGE can force the optimizer to choose the TiKV / TiFlash storage engine for certain tables to execute queries.