Here’s a simple, line-by-line summary of the Serverless compute limitations (09/29/2025):

General limitations

• No Scala or R support → only Python + SQL.
• Only Spark Connect APIs work → no Spark RDD APIs.
• No JAR libraries allowed.
• All workspace users can use serverless compute.
• Notebook tags not supported → use budget policies instead.
• ANSI SQL is default → turn off with spark.sql.ansi.enabled=false.
• No Databricks Container Services.
• Query timeout: max 9000 sec (~2.5 hrs) per query in notebooks (configurable). Jobs don’t have this limit.
• Must use Unity Catalog to connect to external data.
• UDFs cannot access the internet → no CREATE FUNCTION (External).
• Max row size = 128 MB when creating DataFrame from local data.
• No Spark UI → use Query Profile instead.
• No Spark logs → only client-side logs available.
• Cross-workspace access allowed only in same region, and if no IP ACL/PrivateLink.
• No global temp views → use session temp views or tables.
• No Maven coordinates.

Streaming limitations

• Only Trigger.AvailableNow supported → no default/time-based triggers.
• All other Standard access mode streaming limits apply.

Machine learning limitations

• No Spark MLlib.
• No GPU support.

Notebook limitations

• Notebook-scoped libraries not cached between sessions.
• Sharing TEMP tables/views across users not supported.
• No autocomplete / Variable Explorer for DataFrames.
• Default save format = .ipynb → if saved in source format, metadata may break.

Job limitations

• Driver size is fixed (cannot resize).
• Task logs not isolated → logs from multiple tasks mixed together.
• No task libraries for notebook tasks → use notebook-scoped libraries.

Compute-specific limitations

• Not supported: compute policies, init scripts, compute-scoped libraries, instance pools, event logs.
• Most Spark configs blocked → only supported list allowed.
• No environment variables → use widgets for parameters.

Caching limitations

• No cache/un-cache APIs:
  • df.cache(), df.persist(), df.unpersist()
  • spark.catalog.cacheTable(), uncacheTable(), clearCache()
  • SQL: CACHE TABLE, UNCACHE TABLE, REFRESH TABLE, CLEAR CACHE

Hive limitations

• No Hive SerDe tables and LOAD DATA into them.
• Supported file types: Avro, CSV, Delta, JSON, Kafka, ORC, Parquet, Text, XML, BinaryFile.
• No Hive variables (${env:var}, ${configName}, etc.).
  • Instead use: DECLARE VARIABLE, SET VARIABLE, session variables, or IDENTIFIER.

Supported data sources

Write/Update/Delete: CSV, JSON, AVRO, DELTA, KAFKA, PARQUET, ORC, TEXT, UNITY_CATALOG, BINARYFILE, XML, SIMPLESCAN, ICEBERG.

Read: All above plus: MySQL, PostgreSQL, SQL Server, Redshift, Snowflake, Synapse (SQLDW), BigQuery, Oracle, Salesforce, Salesforce Data Cloud, Teradata, Workday RaaS, MongoDB, Databricks.

👉 In short: Serverless is Python + SQL only, works best with Unity Catalog + Delta/Parquet data, but has no JARs, no caching, no Spark UI, no MLlib/GPUs, and strict config limits.

Mohammad Gufran Jahangir