Databricks on AWS

Destination AWS Databricks #

On this page, you’ll find step-by-step instructions on how to set up your AWS Databricks instance with Arcion.

The extracted replicant-cli will be referred to as the $REPLICANT_HOME directory in the following steps.

I. Obtain the JDBC Driver for Databricks #

Replicant requires the Databricks JDBC Driver as a dependency. To obtain the appropriate driver, follow the steps below:

  • Go to the Databricks JDBC Driver download page and download the driver.
  • From the downloaded ZIP, locate and extract the DatabricksJDBC42.jar file.
  • Put the DatabricksJDBC42.jar file inside $REPLICANT_HOME/lib directory.

II. Set up connection configuration #

  1. From $REPLICANT_HOME, navigate to the sample connection configuration file:

    vi conf/conn/databricks.yaml
    
  2. The configuration file has two parts:

    • Parameters related to target Databricks server connection.
    • Parameters related to stage configuration.

    Note: All communications with Databricks happen through port 443, the standard port for HTTPS. So all data is encrypted and secure with SSL by default.
    You can store your connection credentials in a secrets management service and tell Replicant to retrieve the credentials. For more information, see Secrets management.

    Otherwise, you can put your credentials in plain form like the sample below:

    type: DATABRICKS_DELTALAKE
    
    host: HOSTNAME
    port: PORT_NUMBER
    
    url: "jdbc:spark://<host>:<port>/<database-name>;transportMode=http;ssl=1;httpPath=<http-path>;AuthMech=3" #You can copy this URL from Databricks cluster info page
    
    username: "USERNAME"                         
    password: "PASSWORD"                            
    max-connections: 30 #Maximum number of connections Replicant can open in Databricks
    max-retries: 100 #Number of times any operation on the source system will be re-attempted on failures.
    retry-wait-duration-ms: 1000 #Duration in milliseconds replicant should wait before performing then next retry of a failed operation
    

    Replace the following:

    • HOSTNAME: the hostname of your Databricks host
    • PORT_NUMBER: the port number of the Databricks cluster
    • USERNAME: the username that connects to your Databricks server
    • PASSWORD: the password associated with USERNAME
    For Databricks Unity Catalog, set the connection type to DATABRICKS_LAKEHOUSE. For more information, see Databricks Unity Catalog Support.

    It is mandatory to use DATABRICKS_DBFS or an external stage like S3 to hold the data files and load them on the target database from there. The stage section allows specifying the details Replicant needs to connect to and use a given stage.

    • type[v21.06.14.1]: The stage type. Allowed stages are S3, AZURE, GCP, and DATABRICKS_DBFS.

      For Databricks Unity Catalog, set type to DATABRICKS_LAKEHOUSE. For more information, see Databricks Unity Catalog Support.

    • root-dir: Specify a directory on stage which can be used to stage bulk-load files.

    • conn-url[v21.06.14.1]: Specify the connection URL for stage. For example, for S3 as stage, specify bucket-name; for AZURE as stage, specify the container name.

    • use-credentials: Applicable only for DATABRICKS_DBFS as type. Indicates whether to use the provided connection credentials. When true, you must set host, port, username, and password as described in the section Parameters Related to Target Databricks server connection.

      Default: By default, this parameter is set to false.

    • key-id: This config is valid for S3 as stage type only. Represents Access Key ID for AWS account hosting S3.

    • account-name[v21.06.14.1]: This config is valid for AZURE type only. Represents name of the ADLS storage account.

    • secret-key[v21.06.14.1]: This config is valid for S3 and AZURE type only. For example, Secret Access Key for AWS account hosting S3 or ADLS account.

    • credential-file-path: For GCP as stage type only. Represents the absolute path to the service account key file. For more information, see GCP as stage.

    • assume-role: For S3 as stage type only. Use this parameter to configure ARNs (Amazon Resource Name) for roles. It has the following configurable fields:

      • enable: true or false. Whether to enable this option.
      • role-arn: The ARN specifying the role—for example arn:aws:iam::123456789012:role/S3Access.
      • session-duration-s: The duration of AWS STS credentials in seconds.

      For an example, see S3 as stage.

    Databricks DBFS as stage #

    Below is a sample for using Databricks DBFS as stage:

    stage:
      type: DATABRICKS_DBFS
      root-dir: "replicate-stage/databricks-stage"
      use-credentials: false
    

    S3 as stage #

    Below is a sample for using S3 as stage:

    stage:
      type: S3
      root-dir: "replicate-stage/s3-stage"
      key-id: "<S3_access key>"
      conn-url: "replicate-stage"
      secret-key: "<S3 secret key>"
      assume-role:
        enable: true
        role-arn: "<Role_ARN>"
        session-duration-s: 900
    

    GCP as stage #

    When hosting Target Databricks on Google Cloud Platform (GCP), it’s possible to use GCP Storage as the staging area. To use GCP Storage as the staging area, follow the steps below:

    Below is sample for using GCP as stage:

    stage:
      type: GCP
      root-dir: '<stage_directory>'
      conn-url: '<gcp_bucket_name>'
      credential-file-path: '<absolute_file_path_to_service_account_key_file>'
    

III. Configure mapper file (optional) #

If you want to define data mapping from your source to Databricks AWS, specify the mapping rules in the mapper file. For more information on how to define the mapping rules and run Replicant CLI with the mapper file, see Mapper configuration and Mapper configuration in Databricks.

IV. Set up Applier configuration #

  1. From $REPLICANT_HOME, navigate to the applier configuration file:

    vi conf/dst/databricks.yaml
    
  2. The configuration file has two parts:

    • Parameters related to snapshot mode.
    • Parameters related to realtime mode.

    For snapshot mode, make the necessary changes as follows:

    snapshot:
      threads: 16 #Maximum number of threads Replicant should use for writing to the target
    
      #If bulk-load is used, Replicant will use the native bulk-loading capabilities of the target database
      bulk-load:
        enable: true
        type: FILE
        serialize: true|false #Set to true if you want the generated files to be applied in serial/parallel fashion
    

    There are some additional parameters available that you can use in snapshot mode:

    snapshot:
      enable-optimize-write: true
      enable-auto-compact:  true
      enable-unmanaged-delta-table: false
      unmanaged-delta-table-location:
      init-sk: false
      per-table-config:
        init-sk: false
        shard-key:
        enable-optimize-write: true
        enable-auto-compact: true
        enable-unmanaged-delta-table: false
        unmanaged-delta-table-location:
    

    These parameters are specific to Databricks as destination. More details about these parameters are as follows:

    • enable-optimize-write: Databricks dynamically optimizes Apache Spark partition sizes based on the actual data, and attempts to write out 128 MB files for each table partition. This is an approximate size and can vary depending on dataset characteristics.

      Default: By default, this parameter is set to true.

    • enable-auto-compact: After an individual write, Databricks checks if files can be compacted further. If so, it runs an OPTIMIZE job to further compact files for partitions that have the most number of small files. The job is run with 128 MB file sizes instead of the 1 GB file size used in the standard OPTIMIZE.

      Default: By default, this parameter is set to true.

    • enable-unmanaged-delta-table: An unmanaged table is a Spark SQL table for which Spark manages only the metadata. The data is stored in the path provided by the user. So when you perform DROP TABLE <example-table>, Spark removes only the metadata and not the data itself. The data is still present in the path you provided.

      Default: By default, this parameter is set to false.

    • unmanaged-delta-table-location: The path where data for the unmanaged table is to be stored. It can be a Databricks DBFS path (for example FileStore/tables), or an S3 path (for example, s3://replicate-stage/unmanaged-table-data) where the S3 bucket is accessible to Databricks.

    • init-sk: Partition-key on the source table is represented as a shard-key by replicant. By default the target table does not include this sharding information. If init-sk is true we add the shard-key/partition key to target table create SQL. Shard-key replication is disabled by default because DML replication with partitioned tables in Databricks is very slow if the partition key has a high distinct count.

      Default: By default, this parameter is set to false.

    • per-table-config: This configuration allows you to specify various properties for target tables on a per table basis.

      • init-sk: Partition-key on the source table is represented as a shard-key by replicant. By default, the target table does not include this sharding information. If init-sk is true we add the shard-key/partition key to target table create SQL. Shard-key replication is disabled by default because DML replication with partitioned tables in\ databricks is very slow if the partition key has a high distinct count.

        Default: By default, this parameter is set to false.

      • shard-key: Shard key to be used for partitioning the target table.

      • enable-optimize-write: Databricks dynamically optimizes Apache Spark partition sizes based on the actual data, and attempts to write out 128 MB files for each table partition. This is an approximate size and can vary depending on dataset characteristics.

        Default: By default, this parameter is set to true.

      • enable-auto-compact: After an individual write, Databricks checks if files can be compacted further. If so, it runs an OPTIMIZE job to further compact files for partitions that have the most number of small files. The job is run with 128 MB file sizes instead of the 1 GB file size used in the standard OPTIMIZE.

        Default: By default, this parameter is set to true.

      • enable-unmanaged-delta-table: An unmanaged table is a Spark SQL table for which Spark manages only the metadata. The data is stored in the path provided by the user. So when you perform DROP TABLE <example-table>, Spark removes only the metadata and not the data itself. The data is still present in the path you provided.

        Default: By default, this parameter is set to false.

      • unmanaged-delta-table-location: The path where data for the unmanaged table is to be stored. It can be a Databricks DBFS path (for example FileStore/tables), or an S3 path (for example, s3://replicate-stage/unmanaged-table-data) where the S3 bucket is accessible to Databricks.

    If you want to operate in realtime mode, you can use the realtime section to specify your configuration. For example:

    realtime:
      threads: 4
    
      txn-size-rows: 100000
      _traceDBTasks: true
      skip-tables-on-failures : false
      enable-dependency-tracking: true
    
      reuse-temp-table: true
    

    Additional parameters #

    reuse-temp-table

    true or false.

    Enables reusing temporary tables instead of creating and dropping each time the Applier ingests CDC operations.

    The Applier creates temporary table using target table schema to ingest CDC operations to target table. Creating and dropping temporary tables on each CDC replay may slow down CDC ingestion. Enabling this parameter allows you to make CDC ingestion faster by telling the Applier to reuse the temporary tables.

    Default: true.

    parallel-transaction [v23.06.30.2]

    This configuration enables parallel batch processing. Parallel batch processing processes large transactions by splitting them into multiple batches and processing the batches concurrently. This speeds up real-time replication for large transactions and improves overall performance. This feature is available for both Legacy Databricks and Unity Catalog.

    The following configuration options are available:

    Option Value Details
    enable Boolean. {true|false}. Enables parallel batch processing.

    Default: true when txn-size-rows is greater than or equal to 2_000_000.

    txn-rows-threshold Integer Sets the threshold limit for a transaction to qualify for splitting. If transaction size hits this threshold limit, the Applier thread splits the transaction into multiple batches for parallel processing.

    Default: 1_000_000.

    max-chunks-per-table Integer Sets the number of batches to split a transaction into for parallel processing.

    Default: 5.

    threads Integer Sets the number of threads responsible for parallel batch processing. For example, if the maximum number of chunks for each table is 5 and threads is set to 10, then the Applier processes only two transactions concurrently at a time. If new transactions come simultaneously, then the main Applier threads process them serially in one batch.

    Default: The number of available processors to JVM.

    Enable Type-2 CDC #

    From version 22.07.19.3 onwards, Arcion supports Type-2 CDC for Databricks as the target. For more information, see Type-2 CDC and cdc-metadata-type.

For a detailed explanation of configuration parameters in the Applier file, see Applier Reference.

Databricks Unity Catalog support #

From version 22.08.31.3 onwards, Arcion supports Databricks Unity Catalog.

As of now, note the following about the state of Arcion’s Unity Catalog support:

  • Legacy Databricks only supports two-level namespace:

    • Schemas
    • Tables

    With introduction of Unity Catalog, Databricks now exposes a three-level namespace that organizes data.

    • Catalogs
    • Schemas
    • Tables

    Arcion adds support for Unity Catalog by introducing a new child storage type (DATABRICKS_LAKEHOUSE child of DATABRICKS_DELTALAKE).

  • If you’re using Unity Catalog, notice the following when configuring your Target Databricks with Arcion:

    • Set the connection type to DATABRICKS_LAKEHOUSE in the connection configuration file.
    • To avoid manual steps to configure staging, Databricks has introduced personal staging. To read the staging URL, we’ve added a new configuration parameter UNITY_CATALOG_PERSONAL_STAGE. The complete stage configuration is as follows:
      stage:
        type: UNITY_CATALOG_PERSONAL_STAGE
        staging-url: STAGING_URL
        file-format: DATA_FILE_FORMAT
      
      Replace the following:
      • STAGING_URL: the temporary staging URL—for example, stage://tmp/userName/rootDir.

      • DATA_FILE_FORMAT: the type of data file format. Supported formats are PARQUET and CSV.

        Default: PARQUET.

  • We use SparkJDBC42 driver for Legacy Databricks (DATABRICKS_DELTALAKE) and DatabricksJDBC42 for Unity catalog (DATABRICKS_LAKEHOUSE). For instructions on how to obtain these drivers, see Obtain the JDBC Driver for Databricks.

  • Replicant supports Unity Catalog on AWS and Azure platforms.

  • To configure Mapper file in Unity Catalog, see Mapping in Unity Catalog.