Parquet format in Azure Data Factory and Azure Synapse Analytics

APPLIES TO: Azure Data Factory Azure Synapse Analytics

Follow this article when you want to parse the Parquet files or write the data into Parquet format.

Parquet format is supported for the following connectors:

For a list of supported features for all available connectors, visit the Connectors Overview article.

Using Self-hosted Integration Runtime

Important

For copy empowered by Self-hosted Integration Runtime e.g. between on-premises and cloud data stores, if you are not copying Parquet files as-is, you need to install the 64-bit JRE 8 (Java Runtime Environment), JDK 23 (Java Development Kit), or OpenJDK on your IR machine. Check the following paragraph with more details.

For copy running on Self-hosted IR with Parquet file serialization/deserialization, the service locates the Java runtime by firstly checking the registry (SOFTWARE\JavaSoft\Java Runtime Environment\{Current Version}\JavaHome) for JRE, if not found, secondly checking system variable JAVA_HOME for OpenJDK.

  • To use JRE: The 64-bit IR requires 64-bit JRE. You can find it from here.
  • To use JDK: The 64-but IR requires 64-bit JDK 23. You can find it from here. Be sure to update the JAVA_HOME system variable to the root folder of the JDK 23 installation i.e. C:\Program Files\Java\jdk-23, and add the path to both the C:\Program Files\Java\jdk-23\bin and C:\Program Files\Java\jdk-23\bin\server folders to the Path system variable.
  • To use OpenJDK: It's supported since IR version 3.13. Package the jvm.dll with all other required assemblies of OpenJDK into Self-hosted IR machine, and set system environment variable JAVA_HOME accordingly, and then restart Self-hosted IR for taking effect immediately. To download the Microsoft Build of OpenJDK, see Microsoft Build of OpenJDK�.

Tip

If you copy data to/from Parquet format using Self-hosted Integration Runtime and hit error saying "An error occurred when invoking java, message: java.lang.OutOfMemoryError:Java heap space", you can add an environment variable _JAVA_OPTIONS in the machine that hosts the Self-hosted IR to adjust the min/max heap size for JVM to empower such copy, then rerun the pipeline.

Set JVM heap size on Self-hosted IR

Example: set variable _JAVA_OPTIONS with value -Xms256m -Xmx16g. The flag Xms specifies the initial memory allocation pool for a Java Virtual Machine (JVM), while Xmx specifies the maximum memory allocation pool. This means that JVM will be started with Xms amount of memory and will be able to use a maximum of Xmx amount of memory. By default, the service uses min 64 MB and max 1G.

Dataset properties

For a full list of sections and properties available for defining datasets, see the Datasets article. This section provides a list of properties supported by the Parquet dataset.

Property Description Required
type The type property of the dataset must be set to Parquet. Yes
location Location settings of the file(s). Each file-based connector has its own location type and supported properties under location. See details in connector article -> Dataset properties section. Yes
compressionCodec The compression codec to use when writing to Parquet files. When reading from Parquet files, Data Factories automatically determine the compression codec based on the file metadata.
Supported types are "none", "gzip", "snappy" (default), and "lzo". Note currently Copy activity doesn't support LZO when read/write Parquet files.
No

Note

White space in column name is not supported for Parquet files.

Below is an example of Parquet dataset on Azure Blob Storage:

{
    "name": "ParquetDataset",
    "properties": {
        "type": "Parquet",
        "linkedServiceName": {
            "referenceName": "<Azure Blob Storage linked service name>",
            "type": "LinkedServiceReference"
        },
        "schema": [ < physical schema, optional, retrievable during authoring > ],
        "typeProperties": {
            "location": {
                "type": "AzureBlobStorageLocation",
                "container": "containername",
                "folderPath": "folder/subfolder",
            },
            "compressionCodec": "snappy"
        }
    }
}

Copy activity properties

For a full list of sections and properties available for defining activities, see the Pipelines article. This section provides a list of properties supported by the Parquet source and sink.

Parquet as source

The following properties are supported in the copy activity *source* section.

Property Description Required
type The type property of the copy activity source must be set to ParquetSource. Yes
storeSettings A group of properties on how to read data from a data store. Each file-based connector has its own supported read settings under storeSettings. See details in connector article -> Copy activity properties section. No

Parquet as sink

The following properties are supported in the copy activity *sink* section.

Property Description Required
type The type property of the copy activity sink must be set to ParquetSink. Yes
formatSettings A group of properties. Refer to Parquet write settings table below. No
storeSettings A group of properties on how to write data to a data store. Each file-based connector has its own supported write settings under storeSettings. See details in connector article -> Copy activity properties section. No

Supported Parquet write settings under formatSettings:

Property Description Required
type The type of formatSettings must be set to ParquetWriteSettings. Yes
maxRowsPerFile When writing data into a folder, you can choose to write to multiple files and specify the max rows per file. No
fileNamePrefix Applicable when maxRowsPerFile is configured.
Specify the file name prefix when writing data to multiple files, resulted in this pattern: <fileNamePrefix>_00000.<fileExtension>. If not specified, file name prefix will be auto generated. This property does not apply when source is file-based store or partition-option-enabled data store.
No

Mapping data flow properties

In mapping data flows, you can read and write to parquet format in the following data stores: Azure Blob Storage, Azure Data Lake Storage Gen2 and SFTP, and you can read parquet format in Amazon S3.

Source properties

The below table lists the properties supported by a parquet source. You can edit these properties in the Source options tab.

Name Description Required Allowed values Data flow script property
Format Format must be parquet yes parquet format
Wild card paths All files matching the wildcard path will be processed. Overrides the folder and file path set in the dataset. no String[] wildcardPaths
Partition root path For file data that is partitioned, you can enter a partition root path in order to read partitioned folders as columns no String partitionRootPath
List of files Whether your source is pointing to a text file that lists files to process no true or false fileList
Column to store file name Create a new column with the source file name and path no String rowUrlColumn
After completion Delete or move the files after processing. File path starts from the container root no Delete: true or false
Move: [<from>, <to>]
purgeFiles
moveFiles
Filter by last modified Choose to filter files based upon when they were last altered no Timestamp modifiedAfter
modifiedBefore
Allow no files found If true, an error is not thrown if no files are found no true or false ignoreNoFilesFound

Source example

The below image is an example of a parquet source configuration in mapping data flows.

Parquet source

The associated data flow script is:

source(allowSchemaDrift: true,
    validateSchema: false,
    rowUrlColumn: 'fileName',
    format: 'parquet') ~> ParquetSource

Sink properties

The below table lists the properties supported by a parquet sink. You can edit these properties in the Settings tab.

Name Description Required Allowed values Data flow script property
Format Format must be parquet yes parquet format
Clear the folder If the destination folder is cleared prior to write no true or false truncate
File name option The naming format of the data written. By default, one file per partition in format part-#####-tid-<guid> no Pattern: String
Per partition: String[]
As data in column: String
Output to single file: ['<fileName>']
filePattern
partitionFileNames
rowUrlColumn
partitionFileNames

Sink example

The below image is an example of a parquet sink configuration in mapping data flows.

Parquet sink

The associated data flow script is:

ParquetSource sink(
    format: 'parquet',
    filePattern:'output[n].parquet',
    truncate: true,
    allowSchemaDrift: true,
    validateSchema: false,
    skipDuplicateMapInputs: true,
    skipDuplicateMapOutputs: true) ~> ParquetSink

Data type mapping for Parquet

When reading data from the source connector in Parquet format, the following mappings are used from Parquet data types to interim data types used by the service internally.

Parquet type Interim service data type
BOOLEAN Boolean
INT_8 SByte
INT_16 Int16
INT_32 Int32
INT_64 Int64
INT96 DateTime
UINT_8 Byte
UINT_16 UInt16
UINT_32 UInt32
UINT_64 UInt64
DECIMAL Decimal
FLOAT Single
DOUBLE Double
DATE Date
TIME_MILLIS TimeSpan
TIME_MICROS Int64
TIMESTAMP_MILLIS DateTime
TIMESTAMP_MICROS Int64
STRING String
UTF8 String
ENUM Byte array
UUID Byte array
JSON Byte array
BSON Byte array
BINARY Byte array
FIXED_LEN_BYTE_ARRAY Byte array

When writing data to the sink connector in Parquet format, the following mappings are used from interim data types used by the service internally to Parquet data types.

Interim service data type Parquet type
Boolean BOOLEAN
SByte INT_8
Int16 INT_32
Int32 INT_32
Int64 INT_64
Byte INT_32
UInt16 INT_32
UInt32 INT_64
UInt64 DECIMAL
Decimal DECIMAL
Single FLOAT
Double DOUBLE
Date DATE
DateTime INT96
DateTimeOffset INT96
TimeSpan INT96
String UTF8
GUID UTF8
Byte array BINARY

To learn about how the copy activity maps the source schema and data type to the sink, see Schema and data type mappings.

Parquet complex data types (e.g. MAP, LIST, STRUCT) are currently supported only in Data Flows, not in Copy Activity. To use complex types in data flows, do not import the file schema in the dataset, leaving schema blank in the dataset. Then, in the Source transformation, import the projection.