This article is contributed. See the original author and article here.

Many organisations have established their data-lake on Azure to manage their end to end data analytics estate. In some cases, organisations’ customers/partners leverage other cloud providers and we want to meet them wherever they are, after all Azure is an open and versatile platform.

 

If the partner or customer is already using Azure, there are a myriad of options to move data into their estate. Azure Data Share however stands out as its geared towards these cross-tenant data sharing use-cases. It allows users to create invitations, define T&Cs, define snapshot frequency and type (incremental/full), and revoke shares.

 

Pulling data into Azure from other clouds is also rather straight-forward using one of Azure Data Factory’s 90+ copy-activity connectors, including AWS, GCP, Salesforce, Oracle and many more.

Some of these connectors support being used as a source (read) and sink (write). Azure native services, Oracle, SAP, and some others can be used as source and sink. However, not all connectors support this, in which case developers can default to the generic connectors such as ODBC, filesystem, and SFTP connectors.

 

In this blog I want to outline another approach using spark to read and write selected datasets to other clouds such as GCS or S3. However, this methodology applies to really any service that has a spark or Hadoop driver. This gives us bidirectional on-demand access to any cloud storage. As data is read into memory we can join, filter, aggregate data as needed from multiple environments.

Caveat emptor, as data egresses you may be subject to network costs.

 

Pre-Reqs

  1. Azure subscription, Azure Databricks (ADB) workspace, and Azure Data Factory
  2. Google Cloud Platform subscription

 

Google Cloud

  1. Create a service account https://console.cloud.google.com/iam-admin/serviceaccounts/ > ad/nt > create key with type json > keep the json doc safe
  2. Go to https://console.cloud.google.com/storage/ > create bucket (single region, standard storage) > add “Storage Object Admin” permission to service account created in step 1. > upload a test csv file

 

Azure

  1. Navigate to your Azure Databricks workspace (or create one via Quickstart Guide)

     

  2. Upload your GCS service account json to DBFS storage, you can use the Databricks CLI
    databricks fs cp ./myspecialkey.json "dbfs:/data"

     

  3. Create a cluster using the 6.5 Databricks runtime (includes Apache Spark 2.4.5, Scala 2.11) with the following Spark Config:
    spark.hadoop.fs.gs.auth.service.account.json.keyfile /dbfs/data/myspecialkey.json 
    spark.hadoop.fs.gs.impl com.google.cloud.hadoop.fs.gcs.GoogleHadoopFileSystem
    spark.hadoop.fs.gs.project.id {Your-GCP-Project-ID} 
    spark.hadoop.fs.gs.auth.service.account.enable true
    spark.databricks.pyspark.trustedFilesystems com.google.cloud.hadoop.fs.gcs.GoogleHadoopFileSystem,com.databricks.adl.AdlFileSystem,com.databricks.s3a.S3AFileSystem,shaded.databricks.org.apache.hadoop.fs.azure.NativeAzureFileSystem,shaded.databricks.v20180920_b33d810.org.apache.hadoop.fs.azurebfs.SecureAzureBlobFileSystem

    Note: the spark.databricks.pyspark.trustedFilesystems is needed to work around org.apache.spark.api.python.PythonSecurityException exception.

     

  4. Once the cluster is created add Google’s GCS connector as a library
    • Clusters > {your cluster} > Libraries > Install New > Maven > Coordinates: com.google.cloud.bigdataoss:gcs-connector:1.5.2-hadoop2
    • Make note of the version, other versions cause inflection and IOException errors as Databricks uses Hadoop 2.x. Later versions of Databricks runtimes (7.1+) may move to Hadoop 3.

     

  5. At this point we can simply use ls, cp, mv, rm to move data across between ADLS storage and GCS. Note: I set up credential pass-through in my cluster which authenticated me to my data-lake using my SSO/Active Directory credentials.
    dbutils.fs.cp("abfss://{filesystem}@{yourADLSaccount}.dfs.core.windows.net/movies/","gs://{yourGCSbucket}/",True)​

    ls_gs.png 

  6. We can also read individual files/folders into a data-frame.
    df = spark.read.csv("gs://{yourGCSbucket}/{somefile} ", header=True)​

    This will read data directly from your GCS bucket, note this may incur GCP egress costs.

     2_read_gs_df.png 

  7. From here the cloud is the limit, we can define spark tables across multiple environments and query across them as if it were all co-located. Don’t forget however, that spark has to import all the data into memory which may impact performance and egress costs.
    %sql
    CREATE TABLE adlsstest (movie long, title string, genres string, year long, rating long, RottonTomato string) USING CSV LOCATION "abfss://{filesystem}@{yourADLSaccount}.dfs.core.windows.net/movies/moviesDB.csv";
    CREATE TABLE gstest (movie long, title string, genres string, year long, rating long, RottonTomato string) USING CSV LOCATION "gs://{yourGCSbucket}/moviesDB.csv";
    CREATE VIEW myview AS
      SELECT *,'adls' AS src FROM adlsstest
      UNION ALL
      SELECT *,'gcs' as src FROM gstest

jomallin_2-1596027074532.png

 

Automation

From here it is a piece of cake to parameterise and automate movement of data. We can set up an Azure Data Factory pipeline to pass parameters into the Azure Databricks notebooks to do stuff. In this example I copy all files from a specified ADLS directory into a GCS target directory.

 

  1. Create a new notebook in Databricks using the code at the end
  2. Navigate to your Azure Data Factory (or create one via Quickstart Guide)
  3. Create a pipeline with a Databricks activity (here’s a guide)
  4. In the Data Factory pipeline create 3 parameters: sourcedir, targetdir, and myfile. Define some default values using the respective ABFSS and GS formats: 4_adf_integration.png

     

  5. Add the notebook path and map the pipeline parameters to the notebook parameters under Azure Databricks activity > Settings5_adf_parameters.png

     

     

  6. Click Debug, you can optionally modify the parameter values > Click Ok > As its running click on details (spectacles icon) and view notebook output 6_adf_run_debug.png

     

  7. Head back to your GCP portal, and you will see the files appear in GCS7_gcp_gs_directory.png

As we are accessing ADLS from an automated job we cannot use credential passthrough. My colleague Nicholas Hurt wrote a great piece discussing different approaches to authenticating to ADLS from ADB.

I am using a service principal for this demo. Azure documentation outlines how to set this up. Also, I set up a secret scope using the Databricks CLI and stored the service principal key there.

 

 

 

 

 

# authenticate using a service principal and OAuth 2.0
spark.conf.set("fs.azure.account.auth.type", "OAuth")
spark.conf.set("fs.azure.account.oauth.provider.type", "org.apache.hadoop.fs.azurebfs.oauth2.ClientCredsTokenProvider")
spark.conf.set("fs.azure.account.oauth2.client.id", "YOUR-APPLICATION-ID")
spark.conf.set("fs.azure.account.oauth2.client.secret", dbutils.secrets.get(scope = "mykeys", key = "mysp"))
spark.conf.set("fs.azure.account.oauth2.client.endpoint", "https://login.microsoftonline.com/YOUR-TENANT-ID/oauth2/token")

# readdf=spark.read.format("csv").option("header", "true").load("abfss://fs1@####.dfs.core.windows.net/movies/moviesDB.csv")
dbutils.widgets.text("sourcedir", "","")
sourcedir = dbutils.widgets.get("sourcedir")
print ("Param -'sourcedir':", sourcedir)

dbutils.widgets.text("targetdir", "","")
targetdir = dbutils.widgets.get("targetdir")
print ("Param -'targetdir':", targetdir)

dbutils.widgets.text("myfile", "","")
myfile = dbutils.widgets.get("myfile")
print ("Param -'myfile':", myfile)

##dbutils.fs.cp("abfss://fs1@####.dfs.core.windows.net/movies/","gs://####/",True)
dbutils.fs.cp(sourcedir,targetdir,True)
##df = spark.read.csv("gs://####/moviesDB.csv",header=True)
df = spark.read.csv(myfile,header=True)

 

 

 

 

 

Conclusion

Using this approach, we can move data between different storage providers if they provide a compatible jar. In fact, this will work across S3, GCS, BigQuery and many more.

Further we discussed how to automate the process to tie in with broader Azure data orchestration. This approach augments multi-cloud, on-prem data integration capabilities available out of the box with Azure Data Factory.

I should call out that this approach does not support mounts, however that is a minor limitation.

 

Brought to you by Dr. Ware, Microsoft Office 365 Silver Partner, Charleston SC.

%d bloggers like this: