Pushdown computations in PolyBase

Applies to: SQL Server 2016 (13.x) and later versions

Pushdown computation improves the performance of queries on external data sources. Beginning in SQL Server 2016 (13.x), pushdown computations were available for Hadoop external data sources. SQL Server 2019 (15.x) introduced pushdown computations for other types of external data sources.


To determine whether or not PolyBase pushdown computation is benefiting your query, see How to tell if external pushdown occurred.

Enable pushdown computation

The following articles include information about configuring pushdown computation for specific types of external data sources:

This table summarizes pushdown computation support on different external data sources:

Data Source Joins Projections Aggregations Filters Statistics
Generic ODBC Yes Yes Yes Yes Yes
Oracle Yes+ Yes Yes Yes Yes
SQL Server Yes Yes Yes Yes Yes
Teradata Yes Yes Yes Yes Yes
MongoDB* No Yes Yes*** Yes*** Yes
Hadoop No Yes Some** Some** Yes
Azure Blob Storage No No No No Yes

* Azure Cosmos DB pushdown support is enabled via the Azure Cosmos DB API for MongoDB.

** See Pushdown computation and Hadoop providers.

*** Pushdown support for aggregations and filters for the MongoDB ODBC connector for SQL Server 2019 was introduced with SQL Server 2019 CU18.

+ Oracle supports pushdown for joins but you might need to create statistics on the join columns to achieve pushdown.


Pushdown computation can be blocked by some T-SQL syntax. For more information, review Syntax that prevents pushdown.

Pushdown computation and Hadoop providers

PolyBase currently supports two Hadoop providers: Hortonworks Data Platform (HDP) and Cloudera Distributed Hadoop (CDH). There are no differences between the two features in terms of pushdown computation.

To use the computation pushdown functionality with Hadoop, the target Hadoop cluster must have the core components of HDFS, YARN, and MapReduce, with the job history server enabled. PolyBase submits the pushdown query via MapReduce and pulls status from the job history server. Without either component, the query fails.

Some aggregation must occur after the data reaches SQL Server. But a portion of the aggregation occurs in Hadoop. This method is common in computing aggregations in massively parallel processing systems.

Hadoop providers support the following aggregations and filters.

Aggregations Filters (binary comparison)
Count_Big NotEqual
Sum LessThan
Avg LessOrEqual
Max GreaterOrEqual
Min GreaterThan
Approx_Count_Distinct Is

Key beneficial scenarios of pushdown computation

With PolyBase pushdown computation, you can delegate computation tasks to external data sources. This reduces the workload on the SQL Server instance and can significantly improve performance.

SQL Server can push joins, projections, aggregations, and filters to external data sources to take advantage of remote compute and restrict the data sent over the network.

Pushdown of joins

In many cases, PolyBase can facilitate pushdown of the join operator for the join of two external tables on same external data source, which will greatly improve performance.

If the join can be done at the external data source, this reduces the amount of data movement and improves the query's performance. Without join pushdown, the data from the tables to be joined must be brought locally into tempdb, then joined.

In the case of distributed joins (joining a local table to an external table), unless there is a filter on the joined external table, all of the data in the external table must be brought locally into tempdb in order to perform the join operation. For example, the following query has no filtering on the external table join condition, which will result in all of the data from the external table being read.

SELECT * FROM LocalTable L
JOIN ExternalTable E on L.id = E.id

Since the join is on E.id column of the external table, if a filter condition is added to that column, the filter can be pushed down thereby reducing the number of rows read from the external table.

SELECT * FROM LocalTable L
JOIN ExternalTable E on L.id = E.id
WHERE E.id = 20000

Select a subset of rows

Use predicate pushdown to improve performance for a query that selects a subset of rows from an external table.

In this example, SQL Server initiates a map-reduce job to retrieve the rows that match the predicate customer.account_balance < 200000 on Hadoop. Because the query can complete successfully without scanning all of the rows in the table, only the rows that meet the predicate criteria are copied to SQL Server. This saves significant time and requires less temporary storage space when the number of customer balances < 200000 is small in comparison with the number of customers with account balances >= 200000.

SELECT * FROM customer WHERE customer.account_balance < 200000;
SELECT * FROM SensorData WHERE Speed > 65;  

Select a subset of columns

Use predicate pushdown to improve performance for a query that selects a subset of columns from an external table.

In this query, SQL Server initiates a map-reduce job to preprocess the Hadoop delimited-text file so that only the data for the two columns, customer.name and customer.zip_code, will be copied to SQL Server.

SELECT customer.name, customer.zip_code
FROM customer
WHERE customer.account_balance < 200000;

Pushdown for basic expressions and operators

SQL Server allows the following basic expressions and operators for predicate pushdown.

  • Binary comparison operators (<, >, =, !=, <>, >=, <=) for numeric, date, and time values.
  • Arithmetic operators (+, -, *, /, %).
  • Logical operators (AND, OR).
  • Unary operators (NOT, IS NULL, IS NOT NULL).

The operators BETWEEN, NOT, IN, and LIKE might be pushed down. The actual behavior depends on how the query optimizer rewrites the operator expressions as a series of statements that use basic relational operators.

The query in this example has multiple predicates that can be pushed down to Hadoop. SQL Server can push map-reduce jobs to Hadoop to perform the predicate customer.account_balance <= 200000. The expression BETWEEN 92656 AND 92677 is also composed of binary and logical operations that can be pushed to Hadoop. The logical AND in customer.account_balance AND customer.zipcode is a final expression.

Given this combination of predicates, the map-reduce jobs can perform all of the WHERE clause. Only the data that meets the SELECT criteria is copied back to SQL Server.

SELECT * FROM customer 
WHERE customer.account_balance <= 200000 
AND customer.zipcode BETWEEN 92656 AND 92677;

Supported functions for pushdown

SQL Server allows the following functions for predicate pushdown.

String functions

  • LEN
  • LIKE

Mathematical functions

  • ABS
  • ACOS
  • ASIN
  • ATAN
  • COS
  • EXP
  • SIGN
  • SIN
  • SQRT
  • TAN

General functions


* Using with COLLATE can prevent pushdown in some scenarios. For more information, see Collation conflict.

Date & time functions


Syntax that prevents pushdown

The following T-SQL functions or syntax prevents pushdown computation:

  • RAND

Pushdown support for the FORMAT and TRIM syntax was introduced in SQL Server 2019 (15.x) CU10.

Filter clause with variable

When specifying a variable in a filter clause, by default this prevents pushdown of the filter clause. For example, if you run the following query, the filter clause will not be pushed down:

DECLARE @BusinessEntityID INT

SELECT * FROM [Person].[BusinessEntity]  
WHERE BusinessEntityID = @BusinessEntityID;

To achieve pushdown of the variable, you need to enable query optimizer hotfixes functionality. This can be done in any of the following ways:

  • Instance Level: Enable trace flag 4199 as a startup parameter for the instance
  • Database Level: In the context of the database that has the PolyBase external objects, execute ALTER DATABASE SCOPED CONFIGURATION SET QUERY_OPTIMIZER_HOTFIXES = ON

This limitation applies to execution of sp_executesql. The limitation also applies to utilization of some functions in the filter clause.

The ability to pushdown the variable was first introduced in SQL Server 2019 CU5.

Collation conflict

Pushdown might not be possible with data with different collations. Operators like COLLATE can also interfere with the outcome. Equal collations or binary collations are supported. For more information, see How to tell if pushdown occurred.

Pushdown for parquet files

Starting in SQL Server 2022 (16.x), PolyBase introduced support for parquet files. SQL Server is capable of performing both row and column elimination when performing pushdown with parquet. With parquet files, the following operations can be pushed down:

  • Binary comparison operators (>, >=, <=, <) for numeric, date, and time values.
  • Combination of comparison operators (> AND <, >= AND <, > AND <=, <= AND >=).
  • In list filter (col1 = val1 OR col1 = val2 OR vol1 = val3).
  • IS NOT NULL over column.

Presence of the following prevents pushdown for parquet files:

  • Virtual columns.
  • Column comparison.
  • Parameter type conversion.

Supported data types

  • Bit
  • TinyInt
  • SmallInt
  • BigInt
  • Real
  • Float
  • VARCHAR (Bin2Collation, CodePageConversion, BinCollation)
  • NVARCHAR (Bin2Collation, BinCollation)
  • Binary
  • DateTime2 (default and 7-digit precision)
  • Date
  • Time (default and 7-digit precision)
  • Numeric *

* Supported when parameter scale aligns with column scale, or when parameter is explicitly cast to decimal.

Data types that prevent parquet pushdown

  • Money
  • SmallMoney
  • DateTime
  • SmallDateTime


Force pushdown

SELECT * FROM [dbo].[SensorData]
WHERE Speed > 65

Disable pushdown

SELECT * FROM [dbo].[SensorData]
WHERE Speed > 65