Query Azure Cosmos DB data with a serverless SQL pool in Azure Synapse Link

A serverless SQL pool allows you to analyze data in your Azure Cosmos DB containers that are enabled with Azure Synapse Link in near real time without affecting the performance of your transactional workloads. It offers a familiar T-SQL syntax to query data from the analytical store and integrated connectivity to a wide range of business intelligence (BI) and ad-hoc querying tools via the T-SQL interface.

For querying Azure Cosmos DB, the full SELECT surface area is supported through the OPENROWSET function, which includes the majority of SQL functions and operators. You can also store results of the query that reads data from Azure Cosmos DB along with data in Azure Blob Storage or Azure Data Lake Storage by using create external table as select (CETAS). You can't currently store serverless SQL pool query results to Azure Cosmos DB by using CETAS.

In this article, you'll learn how to write a query with a serverless SQL pool that will query data from Azure Cosmos DB containers that are enabled with Azure Synapse Link. You can then learn more about building serverless SQL pool views over Azure Cosmos DB containers and connecting them to Power BI models in this tutorial. This tutorial uses a container with an Azure Cosmos DB well-defined schema. You can also check out the Learn module on how to Query Azure Cosmos DB with SQL Serverless for Azure Synapse Analytics.

Note

You can’t use managed identity to access an Azure Cosmos DB container from serverless SQL pool.

Prerequisites

  • Make sure that you have prepared Analytical store:
  • Make sure that you have applied all best practices, such as:
    • Ensure that your Azure Cosmos DB analytical storage is in the same region as serverless SQL pool.
    • Ensure that the client application (Power BI, Analysis service) is in the same region as serverless SQL pool.
    • If you are returning a large amount of data (bigger than 80GB), consider using caching layer such as Analysis services and load the partitions smaller than 80GB in the Analysis services model.
    • If you are filtering data using string columns, make sure that you are using the OPENROWSET function with the explicit WITH clause that has the smallest possible types (for example, don't use VARCHAR(1000) if you know that the property has up to 5 characters).

Overview

Serverless SQL pool enables you to query Azure Cosmos DB analytical storage using OPENROWSET function.

  • OPENROWSET with inline key. This syntax can be used to query Azure Cosmos DB collections without need to prepare credentials.
  • OPENROWSET that referenced credential that contains the Azure Cosmos DB account key. This syntax can be used to create views on Azure Cosmos DB collections.

To support querying and analyzing data in an Azure Cosmos DB analytical store, a serverless SQL pool is used. The serverless SQL pool uses the OPENROWSET SQL syntax, so you must first convert your Azure Cosmos DB connection string to this format:

OPENROWSET( 
       'CosmosDB',
       '<SQL connection string for Azure Cosmos DB>',
       <Container name>
    )  [ < with clause > ] AS alias

The SQL connection string for Azure Cosmos DB specifies the Azure Cosmos DB account name, database name, database account master key, and an optional region name to the OPENROWSET function. Some of this information can be taken from the standard Azure Cosmos DB connection string.

Converting from the standard Azure Cosmos DB connection string format:

AccountEndpoint=https://<database account name>.documents.azure.com:443/;AccountKey=<database account master key>;

The SQL connection string has the following format:

'account=<database account name>;database=<database name>;region=<region name>;key=<database account master key>'

The region is optional. If omitted, the container's primary region is used.

Important

There is another optional parameter in connection string called endpoint. The endpoint param is needed for accounts that do not match the standard *.documents.azure.com format. For example, if your Azure CosmosDB account ends with .documents.azure.us, make sure that you add endpoint=<account name>.documents.azure.us in the connection string.

The Azure Cosmos DB container name is specified without quotation marks in the OPENROWSET syntax. If the container name has any special characters, for example, a dash (-), the name should be wrapped within square brackets ([]) in the OPENROWSET syntax.

Important

Make sure that you're using some UTF-8 database collation, for example, Latin1_General_100_CI_AS_SC_UTF8, because string values in an Azure Cosmos DB analytical store are encoded as UTF-8 text. A mismatch between text encoding in the file and collation might cause unexpected text conversion errors. You can easily change default collation of the current database by using the T-SQL statement alter database current collate Latin1_General_100_CI_AI_SC_UTF8.

Note

A serverless SQL pool doesn't support querying an Azure Cosmos DB transactional store.

Sample dataset

The examples in this article are based on data from the European Centre for Disease Prevention and Control (ECDC) COVID-19 Cases and COVID-19 Open Research Dataset (CORD-19), doi:10.5281/zenodo.3715505.

You can see the license and the structure of data on these pages. You can also download sample data for the ECDC and CORD-19 datasets.

To follow along with this article showcasing how to query Azure Cosmos DB data with a serverless SQL pool, make sure that you create the following resources:

  • An Azure Cosmos DB database account that's Azure Synapse Link enabled.
  • An Azure Cosmos DB database named covid.
  • Two Azure Cosmos DB containers named Ecdc and Cord19 loaded with the preceding sample datasets.

You can use the following connection string for testing purpose: Account=synapselink-cosmosdb-sqlsample;Database=covid;Key=s5zarR2pT0JWH9k8roipnWxUYBegOuFGjJpSjGlR36y86cW0GQ6RaaG8kGjsRAQoWMw1QKTkkX8HQtFpJjC8Hg==. Note that this connection will not guarantee performance because this account might be located in remote region compared to your Synapse SQL endpoint.

Explore Azure Cosmos DB data with automatic schema inference

The easiest way to explore data in Azure Cosmos DB is by using the automatic schema inference capability. By omitting the WITH clause from the OPENROWSET statement, you can instruct the serverless SQL pool to autodetect (infer) the schema of the analytical store of the Azure Cosmos DB container.

SELECT TOP 10 *
FROM OPENROWSET( 
       'CosmosDB',
       'Account=synapselink-cosmosdb-sqlsample;Database=covid;Key=s5zarR2pT0JWH9k8roipnWxUYBegOuFGjJpSjGlR36y86cW0GQ6RaaG8kGjsRAQoWMw1QKTkkX8HQtFpJjC8Hg==',
       Ecdc) as documents

In the preceding example, we instructed the serverless SQL pool to connect to the covid database in the Azure Cosmos DB account MyCosmosDbAccount authenticated by using the Azure Cosmos DB key (the dummy in the preceding example). We then accessed the Ecdc container's analytical store in the West US 2 region. Since there's no projection of specific properties, the OPENROWSET function will return all properties from the Azure Cosmos DB items.

Assuming that the items in the Azure Cosmos DB container have date_rep, cases, and geo_id properties, the results of this query are shown in the following table:

date_rep cases geo_id
2020-08-13 254 RS
2020-08-12 235 RS
2020-08-11 163 RS

If you need to explore data from the other container in the same Azure Cosmos DB database, you can use the same connection string and reference the required container as the third parameter:

SELECT TOP 10 *
FROM OPENROWSET( 
       'CosmosDB',
       'Account=synapselink-cosmosdb-sqlsample;Database=covid;Key=s5zarR2pT0JWH9k8roipnWxUYBegOuFGjJpSjGlR36y86cW0GQ6RaaG8kGjsRAQoWMw1QKTkkX8HQtFpJjC8Hg==',
       Cord19) as cord19

Explicitly specify schema

While automatic schema inference capability in OPENROWSET provides a simple, easy-to-use querience, your business scenarios might require you to explicitly specify the schema to read-only relevant properties from the Azure Cosmos DB data.

The OPENROWSET function enables you to explicitly specify what properties you want to read from the data in the container and to specify their data types.

Let's imagine that we've imported some data from the ECDC COVID dataset with the following structure into Azure Cosmos DB:

{"date_rep":"2020-08-13","cases":254,"countries_and_territories":"Serbia","geo_id":"RS"}
{"date_rep":"2020-08-12","cases":235,"countries_and_territories":"Serbia","geo_id":"RS"}
{"date_rep":"2020-08-11","cases":163,"countries_and_territories":"Serbia","geo_id":"RS"}

These flat JSON documents in Azure Cosmos DB can be represented as a set of rows and columns in Synapse SQL. The OPENROWSET function enables you to specify a subset of properties that you want to read and the exact column types in the WITH clause:

SELECT TOP 10 *
FROM OPENROWSET(
      'CosmosDB',
      'Account=synapselink-cosmosdb-sqlsample;Database=covid;Key=s5zarR2pT0JWH9k8roipnWxUYBegOuFGjJpSjGlR36y86cW0GQ6RaaG8kGjsRAQoWMw1QKTkkX8HQtFpJjC8Hg==',
       Ecdc
    ) with ( date_rep varchar(20), cases bigint, geo_id varchar(6) ) as rows

The result of this query might look like the following table:

date_rep cases geo_id
2020-08-13 254 RS
2020-08-12 235 RS
2020-08-11 163 RS

For more information about the SQL types that should be used for Azure Cosmos DB values, see the rules for SQL type mappings at the end of the article.

Create view

Creating views in the master or default databases is not recommended or supported. So you need to create a user database for your views.

Once you identify the schema, you can prepare a view on top of your Azure Cosmos DB data. You should place your Azure Cosmos DB account key in a separate credential and reference this credential from OPENROWSET function. Do not keep your account key in the view definition.

CREATE CREDENTIAL MyCosmosDbAccountCredential
WITH IDENTITY = 'SHARED ACCESS SIGNATURE', SECRET = 's5zarR2pT0JWH9k8roipnWxUYBegOuFGjJpSjGlR36y86cW0GQ6RaaG8kGjsRAQoWMw1QKTkkX8HQtFpJjC8Hg==';
GO
CREATE OR ALTER VIEW Ecdc
AS SELECT *
FROM OPENROWSET(
      PROVIDER = 'CosmosDB',
      CONNECTION = 'Account=synapselink-cosmosdb-sqlsample;Database=covid',
      OBJECT = 'Ecdc',
      SERVER_CREDENTIAL = 'MyCosmosDbAccountCredential'
    ) with ( date_rep varchar(20), cases bigint, geo_id varchar(6) ) as rows

Do not use OPENROWSET without explicitly defined schema because it might impact your performance. Make sure that you use the smallest possible sizes for your columns (for example VARCHAR(100) instead of default VARCHAR(8000)). You should use some UTF-8 collation as default database collation or set it as explicit column collation to avoid UTF-8 conversion issue. Collation Latin1_General_100_BIN2_UTF8 provides best performance when you filter data using some string columns.

When you query the view, you may encounter errors or unexpected results. This probably means that the view references columns or objects that were modified or no longer exist. You need to manually adjust the view definition to align with the underlying schema changes. Have in mind that this can happen both when using automatic schema inference in the view and when explicitly specifying the schema.

Query nested objects

With Azure Cosmos DB, you can represent more complex data models by composing them as nested objects or arrays. The autosync capability of Azure Synapse Link for Azure Cosmos DB manages the schema representation in the analytical store out of the box, which includes handling nested data types that allow for rich querying from the serverless SQL pool.

For example, the CORD-19 dataset has JSON documents that follow this structure:

{
    "paper_id": <str>,                   # 40-character sha1 of the PDF
    "metadata": {
        "title": <str>,
        "authors": <array of objects>    # list of author dicts, in order
        ...
     }
     ...
}

The nested objects and arrays in Azure Cosmos DB are represented as JSON strings in the query result when the OPENROWSET function reads them. You can specify the paths to nested values in the objects when you use the WITH clause:

SELECT TOP 10 *
FROM OPENROWSET( 
       'CosmosDB',
       'Account=synapselink-cosmosdb-sqlsample;Database=covid;Key=s5zarR2pT0JWH9k8roipnWxUYBegOuFGjJpSjGlR36y86cW0GQ6RaaG8kGjsRAQoWMw1QKTkkX8HQtFpJjC8Hg==',
       Cord19)
WITH (  paper_id    varchar(8000),
        title        varchar(1000) '$.metadata.title',
        metadata     varchar(max),
        authors      varchar(max) '$.metadata.authors'
) AS docs;

The result of this query might look like the following table:

paper_id title metadata authors
bb11206963e831f… Supplementary Information An eco-epidemi… {"title":"Supplementary Informati… [{"first":"Julien","last":"Mélade","suffix":"","af…
bb1206963e831f1… The Use of Convalescent Sera in Immune-E… {"title":"The Use of Convalescent… [{"first":"Antonio","last":"Lavazza","suffix":"", …
bb378eca9aac649… Tylosema esculentum (Marama) Tuber and B… {"title":"Tylosema esculentum (Ma… [{"first":"Walter","last":"Chingwaru","suffix":"",…

Learn more about analyzing complex data types like Parquet files and containers in Azure Synapse Link for Azure Cosmos DB or nested structures in a serverless SQL pool.

Important

If you see unexpected characters in your text like MÃÂ&copy;lade instead of Mélade, then your database collation isn't set to UTF-8 collation. Change collation of the database to UTF-8 collation by using a SQL statement like ALTER DATABASE MyLdw COLLATE LATIN1_GENERAL_100_CI_AS_SC_UTF8.

Flatten nested arrays

Azure Cosmos DB data might have nested subarrays like the author's array from a CORD-19 dataset:

{
    "paper_id": <str>,                      # 40-character sha1 of the PDF
    "metadata": {
        "title": <str>,
        "authors": [                        # list of author dicts, in order
            {
                "first": <str>,
                "middle": <list of str>,
                "last": <str>,
                "suffix": <str>,
                "affiliation": <dict>,
                "email": <str>
            },
            ...
        ],
        ...
}

In some cases, you might need to "join" the properties from the top item (metadata) with all elements of the array (authors). A serverless SQL pool enables you to flatten nested structures by applying the OPENJSON function on the nested array:

SELECT
    *
FROM
    OPENROWSET(
      'CosmosDB',
      'Account=synapselink-cosmosdb-sqlsample;Database=covid;Key=s5zarR2pT0JWH9k8roipnWxUYBegOuFGjJpSjGlR36y86cW0GQ6RaaG8kGjsRAQoWMw1QKTkkX8HQtFpJjC8Hg==',
       Cord19
    ) WITH ( title varchar(1000) '$.metadata.title',
             authors varchar(max) '$.metadata.authors' ) AS docs
      CROSS APPLY OPENJSON ( authors )
                  WITH (
                       first varchar(50),
                       last varchar(50),
                       affiliation nvarchar(max) as json
                  ) AS a

The result of this query might look like the following table:

title authors first last affiliation
Supplementary Information An eco-epidemi… [{"first":"Julien","last":"Mélade","suffix":"","affiliation":{"laboratory":"Centre de Recher… Julien Mélade {"laboratory":"Centre de Recher…
Supplementary Information An eco-epidemi… [{"first":"Nicolas","last":"4#","suffix":"","affiliation":{"laboratory":"","institution":"U… Nicolas 4# {"laboratory":"","institution":"U…
Supplementary Information An eco-epidemi… [{"first":"Beza","last":"Ramazindrazana","suffix":"","affiliation":{"laboratory":"Centre de Recher… Beza Ramazindrazana {"laboratory":"Centre de Recher…
Supplementary Information An eco-epidemi… [{"first":"Olivier","last":"Flores","suffix":"","affiliation":{"laboratory":"UMR C53 CIRAD, … Olivier Flores {"laboratory":"UMR C53 CIRAD, …

Important

If you see unexpected characters in your text like MÃÂ&copy;lade instead of Mélade, then your database collation isn't set to UTF-8 collation. Change collation of the database to UTF-8 collation by using a SQL statement like ALTER DATABASE MyLdw COLLATE LATIN1_GENERAL_100_CI_AS_SC_UTF8.

Azure Cosmos DB to SQL type mappings

Although Azure Cosmos DB transactional store is schema-agnostic, the analytical store is schematized to optimize for analytical query performance. With the autosync capability of Azure Synapse Link, Azure Cosmos DB manages the schema representation in the analytical store out of the box, which includes handling nested data types. Since a serverless SQL pool queries the analytical store, it's important to understand how to map Azure Cosmos DB input data types to SQL data types.

Azure Cosmos DB accounts of SQL (Core) API support JSON property types of number, string, Boolean, null, nested object, or array. You would need to choose SQL types that match these JSON types if you're using the WITH clause in OPENROWSET. The following table shows the SQL column types that should be used for different property types in Azure Cosmos DB.

Azure Cosmos DB property type SQL column type
Boolean bit
Integer bigint
Decimal float
String varchar (UTF-8 database collation)
Date time (ISO-formatted string) varchar(30)
Date time (UNIX timestamp) bigint
Null any SQL type
Nested object or array varchar(max) (UTF-8 database collation), serialized as JSON text

Full fidelity schema

Azure Cosmos DB full fidelity schema records both values and their best match types for every property in a container. The OPENROWSET function on a container with full fidelity schema provides both the type and the actual value in each cell. Let's assume that the following query reads the items from a container with full fidelity schema:

SELECT *
FROM OPENROWSET(
      'CosmosDB',
      'account=MyCosmosDbAccount;database=covid;region=westus2;key=C0Sm0sDbKey==',
       Ecdc
    ) as rows

The result of this query will return types and values formatted as JSON text:

date_rep cases geo_id
{"date":"2020-08-13"} {"int32":"254"} {"string":"RS"}
{"date":"2020-08-12"} {"int32":"235"} {"string":"RS"}
{"date":"2020-08-11"} {"int32":"316"} {"string":"RS"}
{"date":"2020-08-10"} {"int32":"281"} {"string":"RS"}
{"date":"2020-08-09"} {"int32":"295"} {"string":"RS"}
{"string":"2020/08/08"} {"int32":"312"} {"string":"RS"}
{"date":"2020-08-07"} {"float64":"339.0"} {"string":"RS"}

For every value, you can see the type identified in an Azure Cosmos DB container item. Most of the values for the date_rep property contain date values, but some of them are incorrectly stored as strings in Azure Cosmos DB. Full fidelity schema will return both correctly typed date values and incorrectly formatted string values. The number of cases is information stored as an int32 value, but there's one value that's entered as a decimal number. This value has the float64 type. If there are some values that exceed the largest int32 number, they would be stored as the int64 type. All geo_id values in this example are stored as string types.

Important

The OPENROWSET function without a WITH clause exposes both values with expected types and the values with incorrectly entered types. This function is designed for data exploration and not for reporting. Don't parse JSON values returned from this function to build reports. Use an explicit WITH clause to create your reports. You should clean up the values that have incorrect types in the Azure Cosmos DB container to apply corrections in the full fidelity analytical store.

To query Azure Cosmos DB for MongoDB accounts, you can learn more about the full fidelity schema representation in the analytical store and the extended property names to be used in What is Azure Cosmos DB Analytical Store?.

Query items with full fidelity schema

While querying full fidelity schema, you need to explicitly specify the SQL type and the expected Azure Cosmos DB property type in the WITH clause.

In the following example, we'll assume that string is the correct type for the geo_id property and int32 is the correct type for the cases property:

SELECT geo_id, cases = SUM(cases)
FROM OPENROWSET(
      'CosmosDB'
      'account=MyCosmosDbAccount;database=covid;region=westus2;key=C0Sm0sDbKey==',
       Ecdc
    ) WITH ( geo_id VARCHAR(50) '$.geo_id.string',
             cases INT '$.cases.int32'
    ) as rows
GROUP BY geo_id

Values for geo_id and cases that have other types will be returned as NULL values. This query will reference only the cases with the specified type in the expression (cases.int32).

If you have values with other types (cases.int64, cases.float64) that can't be cleaned in an Azure Cosmos DB container, you would need to explicitly reference them in a WITH clause and combine the results. The following query aggregates both int32, int64, and float64 stored in the cases column:

SELECT geo_id, cases = SUM(cases_int) + SUM(cases_bigint) + SUM(cases_float)
FROM OPENROWSET(
      'CosmosDB',
      'account=MyCosmosDbAccount;database=covid;region=westus2;key=C0Sm0sDbKey==',
       Ecdc
    ) WITH ( geo_id VARCHAR(50) '$.geo_id.string', 
             cases_int INT '$.cases.int32',
             cases_bigint BIGINT '$.cases.int64',
             cases_float FLOAT '$.cases.float64'
    ) as rows
GROUP BY geo_id

In this example, the number of cases is stored either as int32, int64, or float64 values. All values must be extracted to calculate the number of cases per country/region.

Troubleshooting

Review the self-help page to find the known issues or troubleshooting steps that can help you to resolve potential problems with Azure Cosmos DB queries.

Next steps

For more information, see the following articles: