Databricks Runtime 5.5 LTS (EoS)

Article
09/30/2024

Note

Support for this Databricks Runtime version has ended. For the end-of-support date, see End-of-support history. For all supported Databricks Runtime versions, see Databricks Runtime release notes versions and compatibility.

Databricks released this version in July 2019. It was declared Long Term Support (LTS) in August 2019. Support ended on July 27, 2021. Databricks Runtime 5.5 Extended Support (EoS) was released on July 8, 2021 and extends 5.5 support through December 2021. It uses Ubuntu 18.04.5 LTS instead of the deprecated Ubuntu 16.04.6 LTS distribution used in the original Databricks Runtime 5.5 LTS. Ubuntu 16.04.6 LTS support ceased on April 1, 2021.

The following release notes provide information about Databricks Runtime 5.5, powered by Apache Spark.

New features

Delta Lake on Azure Databricks Auto Optimize

Today, when you write data into cloud storage, you need to compact files for optimal I/O performance. You have to worry about the right file size, how often to compact the files, how big a cluster to use, and so on. To address this class of problems, we are excited to announce the general availability of Auto Optimize with Delta Lake on Azure Databricks. During every write to Delta tables, we automatically figure out the right file sizes and compact files so that you don't have to worry about optimizing the storage layout. During writes, if the auto-optimize option is true, Azure Databricks automatically determines if an optimize is required and optimizes the small files. For details, see Configure Delta Lake to control data file size.

Delta Lake on Azure Databricks improved min, max, and count aggregation query performance

The performance of min, max, and count aggregation queries for Delta Lake on Azure Databricks has been significantly improved by reducing the amount of data that's read. These queries are now run using statistics and partition values in the metadata, rather than scanning the data.

Faster model inference pipelines with improved binary file data source and scalar iterator pandas UDF (Public Preview)

Machine learning tasks, especially in the image and video domain, often have to operate on a large number of files. In Databricks Runtime 5.4, we already made available the binary file data source to help ETL arbitrary files such as images, into Spark tables. In Databricks Runtime 5.5, we have added an option, recursiveFileLookup, to load files recursively from nested input directories. See Binary file.

The binary file data source enables you to run model inference tasks in parallel from Spark tables using a scalar pandas UDF. However, you might have to initialize the model for every record batch, which introduces overhead. In Databricks Runtime 5.5, we backported a new pandas UDF type called "scalar iterator" from Apache Spark master. With it you can initialize a model only once and apply the model to many input batches, which can result in a 2-3x speedup for models like ResNet50. See Series to scalar UDF.

Secrets API in R notebooks

The Secrets API lets you inject secrets into notebooks without hardcoding them. This API is now available in R notebooks in addition to existing support for Python and Scala notebooks. You can use the dbutils.secrets.get function to obtain secrets. Secrets are redacted before printing to a notebook cell.

Improvements

Support for running Delta Lake SQL operations in Python foreachBatch: We have fixed the known limitation of not being able to write to Delta tables from inside foreachBatch of a Structured Streaming query defined in Python. This is useful in a common Python streaming workloads; for example, Writing streaming aggregates in update mode using MERGE and foreachBatch.
Performance of Delta tables stored in Azure Data Lake Gen2: The check for the latest version of a Delta table on ADLS Gen2 now only checks the end of the transaction log, rather than listing all available versions. This optimization makes UPDATE a constant time operation, significantly improving the latency.
Scalability of Optimize ZORDER BY: Z-Ordering on very large Delta tables now uses smaller units of work that are governed by advanced admission control. This feature improves the stability of this operation without sacrificing cluster utilization.
Improved performance of DML commands on tables with large number of columns: We now perform better column pruning when scanning for matching data in UPDATE, DELETE, and MERGE commands.
Support for VNet + Service Endpoints setup in Spark - Synapse Analytics connector: We added ADL Gen2 paths to an allow list as temp data locations (.option("tempDir", "abfss://...")) and added a new option called useAzureMSI, to be used instead of forward_spark_azure_storage_credentials in case your Synapse Analytics is configured to authenticate via Managed Identities to your V2 storage account.
Automatic invalidation for disk caching: Disk caching now automatically detects files that have been modified or overwritten after being cached. Any stale entries are automatically invalidated and evicted from the cache. See Optimize performance with caching on Azure Databricks.
Upgraded Python library wheel from 0.33.3 to 0.33.4.
Upgraded R library nlme from 3.1-139 to 3.1-140.

Bug fixes

Fixed cancellation for R commands that do not run Spark jobs. Previously, R commands that do not run Spark jobs could be cancelled, but the notebooks' state would be lost; commands can now be cancelled without losing notebook state.
Dropping or moving a managed table now invalidates the cached Delta Lake log.
Fixed a bug that writing Delta Lake checkpoint may fail due to FileAlreadyExistsException.
Scala REPL now sets the proper -target:jvm-1.8 flag to support calling Java methods that use Java 8 features.

Apache Spark

Databricks Runtime 5.5 includes Apache Spark 2.4.3. This release includes all Spark fixes and improvements included in Databricks Runtime 5.4 (EoS), as well as the following additional bug fixes and improvements made to Spark:

[SPARK-24695] Allow UDFs to return CalendarInterval
[SPARK-28056] add docstring/doctest for SCALAR_ITER pandas UDF
[SPARK-28185] Closes the generator when Python UDFs stop early
[SPARK-24703] Support for interval multiplication
[SPARK-27018][CORE] Fix incorrect removal of checkpointed file in PeriodicCheckpointer
[SPARK-28127][SQL] Micro optimization on TreeNode's mapChildren method
[SPARK-26038] Decimal toScalaBigInt/toJavaBigInteger for decimals not fitting in long
[SPARK-26555][SQL] make ScalaReflection subtype checking thread safe
[SPARK-28081][ML] Handle large vocab counts in word2vec
[SPARK-21882][CORE] OutputMetrics doesn't count written bytes correctly in the saveAsHadoopDataset function
[SPARK-28030] convert filePath to URI in binary file data source
[SPARK-27803][SQL][PYTHON] Fix column pruning for Python UDF
[SPARK-27917][SQL] canonical form of CaseWhen object is incorrect
[SPARK-27798][SQL] from_avro shouldn't produces same value when converted to local relation
[SPARK-27873][SQL] columnNameOfCorruptRecord should not be checked with column names in CSV header when disabling enforceSchema
[SPARK-27907][SQL] HiveUDAF should return NULL in case of 0 rows
[SPARK-27699][SQL] Partially push down disjunctive predicated in Parquet/ORC
[SPARK-27868][CORE] Better default value and documentation for socket server backlog.
[SPARK-27869][CORE] Redact sensitive information in System Properties from UI
[SPARK-27863][SQL][BACKPORT-2.4] Metadata files and temporary files should not be counted as data files
[SPARK-27657][ML] Fix the log format of ml.util.Instrumentation.logFai…
[SPARK-27858][SQL] Fix for Avro deserialization on union types with multiple non-null types
[SPARK-27711][CORE] Unset InputFileBlockHolder at the end of tasks
[SPARK-27351][SQL] Wrong outputRows estimation after AggregateEstimation with only null value column
[SPARK-27539][SQL] Fix inaccurate aggregate outputRows estimation with column containing null values"
[SPARK-27800][SQL] Fix wrong answer on BitwiseXor test cases
[SPARK-27639][SQL] InMemoryTableScan shows the table name on UI if possible
[SPARK-27726][CORE] Fix performance of ElementTrackingStore deletes when using InMemoryStore under high loads
[SPARK-27771][SQL] Add SQL description for grouping functions (cube, rollup, grouping and grouping_id)
[SPARK-27735][SS] Parsing interval string should be case-insensitive in SS
[SPARK-26856][PYSPARK] Python support for from_avro and to_avro APIs
[SPARK-26870][SQL] Move to_avro/from_avro into functions object due to Java compatibility
[SPARK-26812][SQL] Report correct nullability for complex datatypes in Union
[SPARK-27671][SQL] Fix error when casting from a nested null in a struct
[SPARK-27673][SQL] Add since info to random, regex, null expressions
[SPARK-27672][SQL] Add since info to string expressions
[SPARK-25139][SPARK-18406][CORE] Avoid NonFatals to kill the Executor in PythonRunner
[SPARK-27624][CORE] Fix CalenderInterval to show an empty interval correctly
[SPARK-27577][MLLIB] Correct thresholds downsampled in BinaryClassificationMetrics
[SPARK-27621][ML] Linear Regression - validate training related params such as loss only during fitting phase
[SPARK-26048][SPARK-24530] Cherrypick all the missing commits to 2.4 release script
[SPARK-24935][SQL] support INIT -> UPDATE -> MERGE -> FINISH in Hive UDAF adapter

Maintenance updates

See Databricks Runtime 5.5 maintenance updates.

System environment

Operating System: Ubuntu 16.04.6 LTS
Java: 1.8.0_252
Scala: 2.11.12
Python: 2.7.12 for Python 2 clusters and 3.5.2 for Python 3 clusters.
R: R version 3.6.0 (2019-04-26)
GPU clusters: The following NVIDIA GPU libraries are installed:
- Tesla driver 375.66
- CUDA 9.0
- cuDNN 7.0