Index JSON data

Applies to: SQL Server 2016 (13.x) and later Azure SQL Database Azure SQL Managed Instance

In SQL Server and SQL Database, JSON is not a built-in data type, and SQL Server does not have custom JSON indexes. You can optimize your queries over JSON documents, however, by using standard indexes.

Database indexes improve the performance of filter and sort operations. Without indexes, SQL Server has to perform a full table scan every time you query data.

Index JSON properties by using computed columns

When you store JSON data in SQL Server, typically you want to filter or sort query results by one or more properties of the JSON documents.

Example

In this example, assume that the AdventureWorks SalesOrderHeader table has an Info column that contains various information in JSON format about sales orders. For example, it contains unstructured data about customer, sales person, shipping and billing addresses, and so forth. You could use values from the Info column to filter sales orders for a customer.

By default, the column Info used does not exist, it can be created in the AdventureWorks database with the following code. Note this does not apply to the AdventureWorksLT series of sample databases.

IF NOT EXISTS(SELECT * FROM sys.columns WHERE object_id = OBJECT_ID('[Sales].[SalesOrderHeader]') AND name = 'Info')
	ALTER TABLE [Sales].[SalesOrderHeader] ADD [Info] NVARCHAR(MAX) NULL
GO
UPDATE h 
SET [Info] =
(
    SELECT [Customer.Name]	= concat(p.FirstName, N' ', p.LastName), 
           [Customer.ID]	= p.BusinessEntityID, 
           [Customer.Type]	= p.[PersonType], 
           [Order.ID]		= soh.SalesOrderID, 
           [Order.Number]	= soh.SalesOrderNumber, 
           [Order.CreationData] = soh.OrderDate, 
           [Order.TotalDue]	= soh.TotalDue
    FROM [Sales].SalesOrderHeader AS soh
         INNER JOIN [Sales].[Customer] AS c ON c.CustomerID = soh.CustomerID
         INNER JOIN [Person].[Person] AS p ON p.BusinessEntityID = c.CustomerID
    WHERE soh.SalesOrderID = h.SalesOrderID FOR JSON PATH, WITHOUT_ARRAY_WRAPPER 
)
FROM [Sales].SalesOrderHeader AS h; 

Query to optimize

Here's an example of the type of query that you want to optimize by using an index.

SELECT SalesOrderNumber,
    OrderDate,
    JSON_VALUE(Info, '$.Customer.Name') AS CustomerName
FROM Sales.SalesOrderHeader
WHERE JSON_VALUE(Info, '$.Customer.Name') = N'Aaron Campbell' 

Example index

If you want to speed up your filters or ORDER BY clauses over a property in a JSON document, you can use the same indexes that you're already using on other columns. However, you can't directly reference properties in the JSON documents.

  1. First, you have to create a "virtual column" that returns the values that you want to use for filtering.
  2. Then you have to create an index on that virtual column.

The following example creates a computed column that can be used for indexing. Then it creates an index on the new computed column. This example creates a column that exposes the customer name, which is stored in the $.Customer.Name path in the JSON data.

ALTER TABLE Sales.SalesOrderHeader
ADD vCustomerName AS JSON_VALUE(Info,'$.Customer.Name')

CREATE INDEX idx_soh_json_CustomerName
ON Sales.SalesOrderHeader(vCustomerName)  

This statement will return the following warning:

Warning! The maximum key length for a nonclustered index is 1700 bytes.
The index 'vCustomerName' has maximum length of 8000 bytes.
For some combination of large values, the insert/update operation will fail.

The JSON_VALUE function might return text values up to 8000 bytes (for example, as the NVARCHAR(4000) type). However, the values that are longer than 1700 bytes cannot be indexed. If you try to enter the value in the indexed computed column that is longer than 1700 bytes, the DML operation will fail. This will be the runtime error.

For better performance try to cast the value that you have exposed using the computed column into the smallest applicable type. Use int and datetime2 types instead of string types.

More info about the computed column

The computed column is not persisted. It's computed only when the index needs to be rebuilt. It does not occupy additional space in the table.

It's important that you create the computed column with the same expression that you plan to use in your queries - in this example, the expression is JSON_VALUE(Info, '$.Customer.Name').

You don't have to rewrite your queries. If you use expressions with the JSON_VALUE function, as shown in the preceding example query, SQL Server sees that there's an equivalent computed column with the same expression and applies an index if possible.

Execution plan for this example

Here's the execution plan for the query in this example.

Screenshot showing the execution plan for this example.

Instead of a full table scan, SQL Server uses an index seek into the nonclustered index and finds the rows that satisfy the specified conditions. Then it uses a key lookup in the SalesOrderHeader table to fetch the other columns that are referenced in the query - in this example, SalesOrderNumber and OrderDate.

Optimize the index further with included columns

If you add required columns in the index, you can avoid this additional lookup in the table. You can add these columns as standard included columns, as shown in the following example, which extends the preceding CREATE INDEX example.

CREATE INDEX idx_soh_json_CustomerName
ON Sales.SalesOrderHeader(vCustomerName)
INCLUDE(SalesOrderNumber,OrderDate)

In this case SQL Server doesn't have to read additional data from the SalesOrderHeader table because everything it needs is included in the nonclustered JSON index. This type of index is a good way to combine JSON and column data in queries and to create optimal indexes for your workload.

JSON indexes are collation-aware indexes

An important feature of indexes over JSON data is that the indexes are collation-aware. The result of the JSON_VALUE function that you use when you create the computed column is a text value that inherits its collation from the input expression. Therefore, values in the index are ordered using the collation rules defined in the source columns.

To demonstrate that the indexes are collation-aware, the following example creates a simple collection table with a primary key and JSON content.

CREATE TABLE JsonCollection
 (
  id INT IDENTITY CONSTRAINT PK_JSON_ID PRIMARY KEY,
  json NVARCHAR(MAX) COLLATE SERBIAN_CYRILLIC_100_CI_AI
  CONSTRAINT [Content should be formatted as JSON]
  CHECK(ISJSON(json)>0)
 ) 

The preceding command specifies the Serbian Cyrillic collation for the JSON column. The following example populates the table and creates an index on the name property.

INSERT INTO JsonCollection
VALUES
(N'{"name":"Иво","surname":"Андрић"}'),
(N'{"name":"Андрија","surname":"Герић"}'),
(N'{"name":"Владе","surname":"Дивац"}'),
(N'{"name":"Новак","surname":"Ђоковић"}'),
(N'{"name":"Предраг","surname":"Стојаковић"}'),
(N'{"name":"Михајло","surname":"Пупин"}'),
(N'{"name":"Борислав","surname":"Станковић"}'),
(N'{"name":"Владимир","surname":"Грбић"}'),
(N'{"name":"Жарко","surname":"Паспаљ"}'),
(N'{"name":"Дејан","surname":"Бодирога"}'),
(N'{"name":"Ђорђе","surname":"Вајферт"}'),
(N'{"name":"Горан","surname":"Бреговић"}'),
(N'{"name":"Милутин","surname":"Миланковић"}'),
(N'{"name":"Никола","surname":"Тесла"}')
GO
  
ALTER TABLE JsonCollection
ADD vName AS JSON_VALUE(json,'$.name')

CREATE INDEX idx_name
ON JsonCollection(vName)

The preceding commands create a standard index on the computed column vName, which represents the value from the JSON $.name property. In the Serbian Cyrillic code page, the order of the letters is 'А','Б','В','Г','Д','Ђ','Е', etc. The order of items in the index is compliant with Serbian Cyrillic rules because the result of the JSON_VALUE function inherits its collation from the source column. The following example queries this collection and sorts the results by name.

SELECT JSON_VALUE(json,'$.name'),*
FROM JsonCollection
ORDER BY JSON_VALUE(json,'$.name')

If you look at the actual execution plan, you see that it uses sorted values from the nonclustered index.

Screenshot showing an execution plan that uses sorted values from the non-clustered index.

Although the query has an ORDER BY clause, the execution plan doesn't use a Sort operator. The JSON index is already ordered according to Serbian Cyrillic rules. Therefore SQL Server can use the nonclustered index where results are already sorted.

However, if you change the collation of the ORDER BY expression - for example, if you add COLLATE French_100_CI_AS_SC after the JSON_VALUE function - you get a different query execution plan.

Screenshot showing a different execution plan.

Since the order of values in the index is not compliant with French collation rules, SQL Server can't use the index to order results. Therefore, it adds a Sort operator that sorts results using French collation rules.

Next steps

Microsoft videos

Note

Some of the video links in this section may not work at this time. Microsoft is migrating content formerly on Channel 9 to a new platform. We will update the links as the videos are migrated to the new platform.

For a visual introduction to the built-in JSON support in SQL Server and Azure SQL Database, see the following videos: