แก้ไข

แชร์ผ่าน


Connect to Azure Cosmos DB for Apache Cassandra from Spark

APPLIES TO: Cassandra

This article is one among a series of articles on Azure Cosmos DB for Apache Cassandra integration from Spark. The articles cover connectivity, Data Definition Language(DDL) operations, basic Data Manipulation Language(DML) operations, and advanced Azure Cosmos DB for Apache Cassandra integration from Spark.

Prerequisites

Dependencies for connectivity

  • Spark connector for Cassandra: Spark connector is used to connect to Azure Cosmos DB for Apache Cassandra. Identify and use the version of the connector located in Maven central that is compatible with the Spark and Scala versions of your Spark environment. We recommend an environment that supports Spark 3.2.1 or higher, and the spark connector available at maven coordinates com.datastax.spark:spark-cassandra-connector-assembly_2.12:3.2.0. If using Spark 2.x, we recommend an environment with Spark version 2.4.5, using spark connector at maven coordinates com.datastax.spark:spark-cassandra-connector_2.11:2.4.3.

  • Azure Cosmos DB helper library for API for Cassandra: If you're using a version Spark 2.x, then in addition to the Spark connector, you need another library called azure-cosmos-cassandra-spark-helper with maven coordinates com.microsoft.azure.cosmosdb:azure-cosmos-cassandra-spark-helper:1.2.0 from Azure Cosmos DB in order to handle rate limiting. This library contains custom connection factory and retry policy classes.

    The retry policy in Azure Cosmos DB is configured to handle HTTP status code 429("Request Rate Large") exceptions. The Azure Cosmos DB for Apache Cassandra translates these exceptions into overloaded errors on the Cassandra native protocol, and you can retry with back-offs. Because Azure Cosmos DB uses provisioned throughput model, request rate limiting exceptions occur when the ingress/egress rates increase. The retry policy protects your spark jobs against data spikes that momentarily exceed the throughput allocated for your container. If using the Spark 3.x connector, implementing this library isn't required.

    Note

    The retry policy can protect your spark jobs against momentary spikes only. If you have not configured enough RUs required to run your workload, then the retry policy is not applicable and the retry policy class rethrows the exception.

  • Azure Cosmos DB account connection details: Your Azure API for Cassandra account name, account endpoint, and key.

Optimizing Spark connector throughput configuration

Listed in the next section are all the relevant parameters for controlling throughput using the Spark Connector for Cassandra. In order to optimize parameters to maximize throughput for spark jobs, the spark.cassandra.output.concurrent.writes, spark.cassandra.concurrent.reads, and spark.cassandra.input.reads_per_sec configs needs to be correctly configured, in order to avoid too much throttling and back-off (which in turn can lead to lower throughput).

The optimal value of these configurations depends on four factors:

  • The amount of throughput (Request Units) configured for the table that data is being ingested into.
  • The number of workers in your Spark cluster.
  • The number of executors configured for your spark job (which can be controlled using spark.cassandra.connection.connections_per_executor_max or spark.cassandra.connection.remoteConnectionsPerExecutor depending on Spark version)
  • The average latency of each request to Azure Cosmos DB, if you're collocated in the same Data Center. Assume this value to be 10 ms for writes and 3 ms for reads.

As an example, if we have five workers and a value of spark.cassandra.output.concurrent.writes= 1, and a value of spark.cassandra.connection.remoteConnectionsPerExecutor = 1, then we have five workers that are concurrently writing into the table, each with one thread. If it takes 10 ms to perform a single write, then we can send 100 requests (1000 milliseconds divided by 10) per second, per thread. With five workers, this would be 500 writes per second. At an average cost of five request units (RUs) per write, the target table would need a minimum 2500 request units provisioned (5 RUs x 500 writes per second).

Increasing the number of executors can increase the number of threads in a given job, which can in turn increase throughput. However, the exact impact of this can be variable depending on the job, while controlling throughput with number of workers is more deterministic. You can also determine the exact cost of a given request by profiling it to get the Request Unit (RU) charge. This will help you to be more accurate when provisioning throughput for your table or keyspace. Have a look at our article here to understand how to get request unit charges at a per request level.

Scaling throughput in the database

The Cassandra Spark connector will saturate throughput in Azure Cosmos DB efficiently. As a result, even with effective retries, you'll need to ensure you have sufficient throughput (RUs) provisioned at the table or keyspace level to prevent rate limiting related errors. The minimum setting of 400 RUs in a given table or keyspace won't be sufficient. Even at minimum throughput configuration settings, the Spark connector can write at a rate corresponding to around 6000 request units or more.

If the RU setting required for data movement using Spark is higher than what is required for your steady state workload, you can easily scale throughput up and down systematically in Azure Cosmos DB to meet the needs of your workload for a given time period. Read our article on elastic scale in API for Cassandra to understand the different options for scaling programmatically and dynamically.

Note

The guidance above assumes a reasonably uniform distribution of data. If you have a significant skew in the data (that is, an inordinately large number of reads/writes to the same partition key value), then you might still experience bottlenecks, even if you have a large number of request units provisioned in your table. Request units are divided equally among physical partitions, and heavy data skew can cause a bottleneck of requests to a single partition.

Spark connector throughput configuration parameters

The following table lists Azure Cosmos DB for Apache Cassandra-specific throughput configuration parameters provided by the connector. For a detailed list of all configuration parameters, see configuration reference page of the Spark Cassandra Connector GitHub repository.

Property Name Default value Description
spark.cassandra.output.batch.size.rows 1 Number of rows per single batch. Set this parameter to 1. This parameter is used to achieve higher throughput for heavy workloads.
spark.cassandra.connection.connections_per_executor_max (Spark 2.x) spark.cassandra.connection.remoteConnectionsPerExecutor (Spark 3.x) None Maximum number of connections per node per executor. 10*n is equivalent to 10 connections per node in an n-node Cassandra cluster. So, if you require five connections per node per executor for a five node Cassandra cluster, then you should set this configuration to 25. Modify this value based on the degree of parallelism or the number of executors that your spark jobs are configured for.
spark.cassandra.output.concurrent.writes 100 Defines the number of parallel writes that can occur per executor. Because you set "batch.size.rows" to 1, make sure to scale up this value accordingly. Modify this value based on the degree of parallelism or the throughput that you want to achieve for your workload.
spark.cassandra.concurrent.reads 512 Defines the number of parallel reads that can occur per executor. Modify this value based on the degree of parallelism or the throughput that you want to achieve for your workload
spark.cassandra.output.throughput_mb_per_sec None Defines the total write throughput per executor. This parameter can be used as an upper limit for your spark job throughput, and base it on the provisioned throughput of your Azure Cosmos DB container.
spark.cassandra.input.reads_per_sec None Defines the total read throughput per executor. This parameter can be used as an upper limit for your spark job throughput, and base it on the provisioned throughput of your Azure Cosmos DB container.
spark.cassandra.output.batch.grouping.buffer.size 1000 Defines the number of batches per single spark task that can be stored in memory before sending to API for Cassandra
spark.cassandra.connection.keep_alive_ms 60000 Defines the period of time until which unused connections are available.

Adjust the throughput and degree of parallelism of these parameters based on the workload you expect for your spark jobs, and the throughput you've provisioned for your Azure Cosmos DB account.

Connecting to Azure Cosmos DB for Apache Cassandra from Spark

cqlsh

The following commands detail how to connect to Azure Cosmos DB for Apache Cassandra from cqlsh. This is useful for validation as you run through the samples in Spark.
From Linux/Unix/Mac:

export SSL_VERSION=TLSv1_2
export SSL_VALIDATE=false
cqlsh.py YOUR-COSMOSDB-ACCOUNT-NAME.cassandra.cosmosdb.azure.com 10350 -u YOUR-COSMOSDB-ACCOUNT-NAME -p YOUR-COSMOSDB-ACCOUNT-KEY --ssl

1. Azure Databricks

The article below covers Azure Databricks cluster provisioning, cluster configuration for connecting to Azure Cosmos DB for Apache Cassandra, and several sample notebooks that cover DDL operations, DML operations and more.
Work with Azure Cosmos DB for Apache Cassandra from Azure Databricks

2. Azure HDInsight-Spark

The article below covers HDinsight-Spark service, provisioning, cluster configuration for connecting to Azure Cosmos DB for Apache Cassandra, and several sample notebooks that cover DDL operations, DML operations and more.
Work with Azure Cosmos DB for Apache Cassandra from Azure HDInsight-Spark

3. Spark environment in general

While the sections above were specific to Azure Spark-based PaaS services, this section covers any general Spark environment. Connector dependencies, imports, and Spark session configuration are detailed below. The "Next steps" section covers code samples for DDL operations, DML operations and more.

Connector dependencies:

  1. Add the maven coordinates to get the Cassandra connector for Spark
  2. Add the maven coordinates for the Azure Cosmos DB helper library for API for Cassandra

Imports:

import org.apache.spark.sql.cassandra._
//Spark connector
import com.datastax.spark.connector._
import com.datastax.spark.connector.cql.CassandraConnector

//CosmosDB library for multiple retry
import com.microsoft.azure.cosmosdb.cassandra

Spark session configuration:

 spark.cassandra.connection.host  YOUR_ACCOUNT_NAME.cassandra.cosmosdb.azure.com  
 spark.cassandra.connection.port  10350  
 spark.cassandra.connection.ssl.enabled  true  
 spark.cassandra.auth.username  YOUR_ACCOUNT_NAME  
 spark.cassandra.auth.password  YOUR_ACCOUNT_KEY  
// if using Spark 2.x
// spark.cassandra.connection.factory  com.microsoft.azure.cosmosdb.cassandra.CosmosDbConnectionFactory  

//Throughput-related...adjust as needed
 spark.cassandra.output.batch.size.rows  1  
// spark.cassandra.connection.connections_per_executor_max  10   // Spark 2.x
 spark.cassandra.connection.remoteConnectionsPerExecutor  10   // Spark 3.x
 spark.cassandra.output.concurrent.writes  1000  
 spark.cassandra.concurrent.reads  512  
 spark.cassandra.output.batch.grouping.buffer.size  1000  
 spark.cassandra.connection.keep_alive_ms  600000000 

Next steps

The following articles demonstrate Spark integration with Azure Cosmos DB for Apache Cassandra.