Query databases using JDBC

Databricks supports connecting to external databases using JDBC. This article provides the basic syntax for configuring and using these connections with examples in Python, SQL, and Scala.

Experimental

The configurations described in this article are Experimental. Experimental features are provided as-is and are not supported by Databricks through customer technical support. To get full query federation support, you should instead use Lakehouse Federation, which enables your Databricks users to take advantage of Unity Catalog syntax and data governance tools.

Partner Connect provides optimized integrations for syncing data with many external external data sources. See What is Databricks Partner Connect?.

Important

The examples in this article do not include usernames and passwords in JDBC URLs. Databricks recommends using secrets to store your database credentials. For example:

username = dbutils.secrets.get(scope = "jdbc", key = "username")
password = dbutils.secrets.get(scope = "jdbc", key = "password")

val username = dbutils.secrets.get(scope = "jdbc", key = "username")
val password = dbutils.secrets.get(scope = "jdbc", key = "password")

To reference Databricks secrets with SQL, you must configure a Spark configuration property during cluster initilization.

For a full example of secret management, see Secret workflow example.

Note

The Spark UI support is unavailable on Databricks on Google Cloud as of this release.

Read data with JDBC

You must configure a number of settings to read data using JDBC. Note that each database uses a different format for the <jdbc-url>.

employees_table = (spark.read
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .load()
)

CREATE TEMPORARY VIEW employees_table_vw
USING JDBC
OPTIONS (
  url "<jdbc-url>",
  dbtable "<table-name>",
  user '<username>',
  password '<password>'
)

val employees_table = spark.read
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .load()

Spark automatically reads the schema from the database table and maps its types back to Spark SQL types.

employees_table.printSchema

DESCRIBE employees_table_vw

employees_table.printSchema

You can run queries against this JDBC table:

display(employees_table.select("age", "salary").groupBy("age").avg("salary"))

SELECT age, avg(salary) as salary
FROM employees_table_vw
GROUP BY age

display(employees_table.select("age", "salary").groupBy("age").avg("salary"))

Write data with JDBC

Saving data to tables with JDBC uses similar configurations to reading. See the following example:

(employees_table.write
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<new-table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .save()
)

CREATE TABLE new_employees_table
  USING JDBC
OPTIONS (
  url "<jdbc-url>",
  dbtable "<table-name>",
  user '<username>',
  password '<password>'
) AS
SELECT * FROM employees_table_vw

employees_table.write
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<new-table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .save()

The default behavior attempts to create a new table and throws an error if a table with that name already exists.

You can append data to an existing table using the following syntax:

(employees_table.write
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<new-table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .mode("append")
  .save()
)

CREATE TABLE IF NOT EXISTS new_employees_table
  USING JDBC
OPTIONS (
  url "<jdbc-url>",
  dbtable "<table-name>",
  user '<username>',
  password '<password>'
);

INSERT INTO new_employees_table
SELECT * FROM employees_table_vw;

employees_table.write
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<new-table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .mode("append")
  .save()

You can overwrite an existing table using the following syntax:

(employees_table.write
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<new-table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .mode("overwrite")
  .save()
)

CREATE OR REPLACE TABLE new_employees_table
  USING JDBC
OPTIONS (
  url "<jdbc-url>",
  dbtable "<table-name>",
  user '<username>',
  password '<password>'
) AS
SELECT * FROM employees_table_vw;

employees_table.write
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<new-table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .mode("overwrite")
  .save()

Control parallelism for JDBC queries

By default, the JDBC driver queries the source database with only a single thread. To improve performance for reads, you need to specify a number of options to control how many simultaneous queries Databricks makes to your database. For small clusters, setting the numPartitions option equal to the number of executor cores in your cluster ensures that all nodes query data in parallel.

Warning

Setting numPartitions to a high value on a large cluster can result in negative performance for the remote database, as too many simultaneous queries might overwhelm the service. This is especially troublesome for application databases. Be wary of setting this value above 50.

Note

Speed up queries by selecting a column with an index calculated in the source database for the partitionColumn.

The following code example demonstrates configuring parallelism for a cluster with eight cores:

employees_table = (spark.read
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  # a column that can be used that has a uniformly distributed range of values that can be used for parallelization
  .option("partitionColumn", "<partition-key>")
  # lowest value to pull data for with the partitionColumn
  .option("lowerBound", "<min-value>")
  # max value to pull data for with the partitionColumn
  .option("upperBound", "<max-value>")
  # number of partitions to distribute the data into. Do not set this very large (~hundreds)
  .option("numPartitions", 8)
  .load()
)

CREATE TEMPORARY VIEW employees_table_vw
USING JDBC
OPTIONS (
  url "<jdbc-url>",
  dbtable "<table-name>",
  user '<username>',
  password '<password>',
  partitionColumn "<partition-key>",
  lowerBound "<min-value>",
  upperBound "<max-value>",
  numPartitions 8
)

val employees_table = spark.read
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  // a column that can be used that has a uniformly distributed range of values that can be used for parallelization
  .option("partitionColumn", "<partition-key>")
  // lowest value to pull data for with the partitionColumn
  .option("lowerBound", "<min-value>")
  // max value to pull data for with the partitionColumn
  .option("upperBound", "<max-value>")
  // number of partitions to distribute the data into. Do not set this very large (~hundreds)
  .option("numPartitions", 8)
  .load()

Note

Databricks supports all Apache Spark options for configuring JDBC.

When writing to databases using JDBC, Apache Spark uses the number of partitions in memory to control parallelism. You can repartition data before writing to control parallelism. Avoid high number of partitions on large clusters to avoid overwhelming your remote database. The following example demonstrates repartitioning to eight partitions before writing:

(employees_table.repartition(8)
  .write
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<new-table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .save()
)

CREATE TABLE new_employees_table
  USING JDBC
OPTIONS (
  url "<jdbc-url>",
  dbtable "<table-name>",
  user '<username>',
  password '<password>'
) AS
SELECT /*+ REPARTITION(8) */ * FROM employees_table_vw

employees_table.repartition(8)
  .write
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<new-table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .save()

Push down a query to the database engine

You can push down an entire query to the database and return just the result. The table parameter identifies the JDBC table to read. You can use anything that is valid in a SQL query FROM clause.

pushdown_query = "(select * from employees where emp_no < 10008) as emp_alias"

employees_table = (spark.read
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", pushdown_query)
  .option("user", "<username>")
  .option("password", "<password>")
  .load()
)

CREATE TEMPORARY VIEW employees_table_vw
USING JDBC
OPTIONS (
  url "<jdbc-url>",
  dbtable "(select * from employees where emp_no < 10008) as emp_alias",
  user '<username>',
  password '<password>'
)

val pushdown_query = "(select * from employees where emp_no < 10008) as emp_alias"

val employees_table = spark.read
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", pushdown_query)
  .option("user", "<username>")
  .option("password", "<password>")
  .load()

Control number of rows fetched per query

JDBC drivers have a fetchSize parameter that controls the number of rows fetched at a time from the remote database.

Setting

Result

Too low

High latency due to many roundtrips (few rows returned per query)

Too high

Out of memory error (too much data returned in one query)

The optimal value is workload dependent. Considerations include:

  • How many columns are returned by the query?

  • What data types are returned?

  • How long are the strings in each column returned?

Systems might have very small default and benefit from tuning. For example: Oracle’s default fetchSize is 10. Increasing it to 100 reduces the number of total queries that need to be executed by a factor of 10. JDBC results are network traffic, so avoid very large numbers, but optimal values might be in the thousands for many datasets.

Use the fetchSize option, as in the following example:

employees_table = (spark.read
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .option("fetchSize", "100")
  .load()
)

CREATE TEMPORARY VIEW employees_table_vw
USING JDBC
OPTIONS (
  url "<jdbc-url>",
  dbtable "<table-name>",
  user '<username>',
  password '<password>'.
  fetchSize 100
)

val employees_table = spark.read
  .format("jdbc")
  .option("url", "<jdbc-url>")
  .option("dbtable", "<table-name>")
  .option("user", "<username>")
  .option("password", "<password>")
  .option("fetchSize", "100")
  .load()