aidp-jdbc-custom — Generic JDBC escape hatch

The catch-all skill for any DB with a JDBC driver. Skips the AIDP aidataplatform format and uses native Spark JDBC. Useful for DBs like SQLite, ClickHouse, DuckDB, IBM DB2, SAP HANA, or any niche driver the user has uploaded.

When to use

• The DB doesn't have a dedicated aidp-* skill in this plugin.
• User has a .jar JDBC driver they want to use.
• Mentioned: "custom JDBC", "JDBC driver", "any JDBC".

When NOT to use

• For Postgres / MySQL / SQL Server / Oracle → use the dedicated skill. The aidataplatform format gives the connector pushdown and connection pooling that this skill doesn't.
• For Snowflake → aidp-snowflake. The Spark connector is much better than raw JDBC.

Two ways to load a non-bundled JDBC driver

Option A — Runtime-load (recommended; no cluster restart)

The plugin ships a helper that loads a JDBC JAR into a running Spark session via Java's URLClassLoader + DriverManager. It works without admin access and without restarting the kernel.

import os
from oracle_ai_data_platform_connectors.jdbc import (
    add_jdbc_jar_at_runtime, download_jdbc_jar,
)

# Download once (Maven Central is reachable from AIDP clusters)
jar = download_jdbc_jar(
    maven_url="https://repo1.maven.org/maven2/org/xerial/sqlite-jdbc/3.46.0.0/sqlite-jdbc-3.46.0.0.jar",
    target_path="/tmp/sqlite-jdbc-3.46.0.0.jar",
)

# Register with the running Spark session
add_jdbc_jar_at_runtime(spark, jar_path=jar, driver_class="org.sqlite.JDBC")

# Now standard Spark JDBC works
df = (spark.read.format("jdbc")
      .option("url",      "jdbc:sqlite::memory:")
      .option("driver",   "org.sqlite.JDBC")
      .option("dbtable",  "(SELECT 1 AS c1, 2 AS c2, 3 AS c3)")
      .option("fetchsize","1000")
      .load())
df.show()

The helper is implemented at scripts/oracle_ai_data_platform_connectors/jdbc/runtime_load.py — internally it builds a URLClassLoader rooted at the existing thread context loader, calls DriverManager.registerDriver, and sets the JVM thread context class loader so Spark's Utils.classForName resolves the driver class.

Option B — Cluster Library tab (durable, requires admin)

For frequently-used drivers, upload the JAR to a Volume and attach via the cluster Library tab. This persists across cluster restarts. Requires cluster admin access. After attach + restart:

# Driver class is now on the system classpath; no runtime trick needed.
df = (spark.read.format("jdbc")
      .option("url", JDBC_URL).option("driver", DRIVER)
      .option("dbtable", TABLE).load())

Generic template

df = (spark.read.format("jdbc")
      .option("url",      "jdbc:<vendor>://<host>:<port>/<db>")
      .option("driver",   "<full.class.Name>")
      .option("user",     os.environ["CUST_DB_USER"])
      .option("password", os.environ["CUST_DB_PASSWORD"])
      .option("dbtable",  os.environ["CUST_DB_TABLE"])
      .option("fetchsize", "10000")
      .load())

Common driver classes

Gotchas

• No predicate pushdown beyond what Spark JDBC infers. This skill is the escape hatch, not the optimized path.
• dbtable accepts a subquery — wrap in parens to filter at the source: option("dbtable", "(SELECT * FROM big_table WHERE date > '2025-01-01') t").
• fetchsize=10000 is a good default; smaller values create driver chatter, larger values risk OOM on the executor.
• Partitioning — for parallel reads, use option("partitionColumn", ...).option("lowerBound", ...).option("upperBound", ...).option("numPartitions", N). Without these the read is single-partition and serial.
• Driver JAR mismatch — symptom is ClassNotFoundException: <driver class>. Re-check that the JAR is attached to the running cluster (not just uploaded to a Volume).

References

• Official sample: oracle-samples/oracle-aidp-samples → data-engineering/ingestion/Connect_Using_Custom_JDBC_Driver.ipynb
• Spark JDBC docs: https://spark.apache.org/docs/latest/sql-data-sources-jdbc.html