Databricks Runtime 5.3 (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 April 2019.

The following release notes provide information about Databricks Runtime 5.3, powered by Apache Spark.

New Features

Databricks Delta

Time travel GA

Delta time travel has graduated to general availability. It adds the ability to query a snapshot of a table using a timestamp string or a version, using SQL syntax as well as DataFrameReader options for timestamp expressions.

SELECT count(*) FROM events TIMESTAMP AS OF timestamp_expression
SELECT count(*) FROM events VERSION AS OF version

Time travel has many use cases, including:

Re-creating analyses, reports, or outputs (for example, the output of a machine learning model), which is useful for debugging or auditing, especially in regulated industries.
Writing complex temporal queries.
Fixing mistakes in your data.
Providing snapshot isolation for a set of queries for fast changing tables.

Delta time travel

See Work with Delta Lake table history.

MySQL table replication to Delta Public Preview

Stream data from a MySQL table directly into Delta for downstream consumption in Spark analytics or data science workflows. Leveraging the same strategy that MySQL uses for replication to other instances, the binlog is used to identify updates that are then processed and streamed to Databricks as follows:

Reads change events from the database log.
Streams the events to Databricks.
Writes in the same order to a Delta table.
Maintains state in case of disconnects from the source.

Private preview features

This release contains the following private preview features for Delta Lake:

Added an experimental command that allows Presto to directly query Delta tables.
Auto Optimize optimizes the file layout during individual writes to a Delta table and prevents small files from accumulating in your Delta tables to maintain lightning-speed query performance. In addition, writes and upserts into partitioned Delta tables may enjoy large speedups thanks to an adaptive shuffle introduced into writes.

If you want to participate in either preview, contact your Databricks account team.

Optimized DBFS FUSE folder for deep learning workloads

Azure Databricks now offers an optimized FUSE mount. You can have high-performance data access during training and inference without applying init scripts. Data stored under dbfs:/ml and accessible locally at file:/dbfs/ml is now backed by this optimized FUSE mount. See Load data for machine learning and deep learning.

Improvements

Notebook-scoped libraries:
- Support for Extras in dbutils.library.installPyPI.
- Support for notebook-scoped library installation for libraries in the Azure file system, including wasbs, adls, and abfss.
- Improved error messages for notebook-scoped libraries on unsupported targets. For example: "Library utilities are not available on Databricks Runtime for Machine Learning."
Connect to Azure Data Lake Storage Gen2 (ADLS Gen2) without having to grant an account-wide "Storage Blob Data Contributor" RBAC role to the Service Principal.
Databricks Advisor has additional hints to improve the performance of queries:
- When there are many small files per partition -> Convert table to Delta and run OPTIMIZE.
- When there is a highly selective filter -> Convert to Delta and use ZORDER BY.
- When metadata fetches from Hive become a bottleneck -> Convert your table to Delta.
Upgraded Ubuntu from 16.04.5 LTS to 16.04.6 LTS.
Upgraded Scala 2.11 from 2.11.8 to 2.11.12.
Upgraded some installed Python libraries:
- pip: 18.1 to 19.0.3.
- setuptools: 40.6.3 to 40.8.0
- wheel: 0.32.3 to 0.33.1
Upgraded R from 3.4.4 to 3.5.2.
Upgraded several installed R libraries. See Installed R Libraries.

Deprecation

Apache Kafka 0.8 connector for Structured Streaming is no longer supported. It is included in Databricks Runtime 5.3 but will be removed from a future Databricks Runtime release.
Databricks ML Model Export is deprecated and will be removed in Databricks Runtime 6.0. Use MLeap for importing and exporting models instead.
Removed spark.databricks.pyspark.enableExecutorFsPerms as a configuration option.

Bug fixes

Fixed cluster-wide Python Egg library installation for clusters enabled for table ACLs.
Improved Python command cancellation by fixing the situation in which cancellation is called before the command is executed.
Fixed a bug in table ACLs: now when you list objects in a database or catalog, you see only the objects that you have permission to see.

Apache Spark

Databricks Runtime 5.3 includes Apache Spark 2.4.0. This release includes all Spark fixes and improvements included in Databricks Runtime 5.2 (EoS), as well as the following additional bug fixes and improvements made to Spark:

[SPARK-27134][SQL] array_distinct function does not work correctly with columns containing array of array
[SPARK-24669][SQL] Invalidate tables in case of DROP DATABASE CASCADE
[SPARK-26572][SQL] fix aggregate codegen result evaluation
[SPARK-27046][DSTREAMS] Remove SPARK-19185 related references from documentation
[SPARK-26449][PYTHON] Add transform method to DataFrame API
[SPARK-26740][SQL] Read timestamp/date column stats written by Spark 3.0
[SPARK-26909][SQL] use unsafeRow.hashCode() as hash value in HashAggregate
[SPARK-26990][SQL] FileIndex: use user specified field names if possible
[SPARK-26851][SQL] Fix double-checked locking in CachedRDDBuilder
[SPARK-26864][SQL] Query may return incorrect result when Python UDF is used as a left-semi join condition
[SPARK-26887][SQL][PYTHON] Create datetime.date directly instead of creating datetime64 as intermediate data.
[SPARK-26859][SQL] Fix field writer index bug in non-vectorized ORC deserializer
[SPARK-26864][SQL] Query may return incorrect result when Python UDF is used as a join condition and the udf uses attributes from both legs of left semi join
[SPARK-24360][SPARK-26091][BACKPORT][SQL] Support Hive 3.1 metastore
[SPARK-26873][SQL] Use a consistent timestamp to build Hadoop Job IDs.
[SPARK-26734][STREAMING] Fix StackOverflowError with large block queue
[SPARK-26758][CORE] Idle Executors are not getting killed after spark.dynamiAllocation.executorIdleTimeout value
[SPARK-26751][SQL] Fix memory leak when statement run in background and throw exception which is not HiveSQLException
[SPARK-26806][SS] EventTimeStats.merge should handle zeros correctly
[SPARK-26745][SPARK-24959][SQL] Revert count optimization in JSON datasource by
[SPARK-26757][GRAPHX] Return 0 for count on empty Edge/Vertex RDDs
[SPARK-26726] Synchronize the amount of memory used by the broadcast variable to the UI display
[SPARK-26718][SS] Fixed integer overflow in SS kafka rateLimit calculation
[SPARK-26708][SQL] Incorrect result caused by inconsistency between a SQL cache's cached RDD and its physical plan
[SPARK-26735][SQL] Verify plan integrity for special expressions
[SPARK-26619][SQL] Prune the unused serializers from SerializeFromObject
[SPARK-26379][SS] Use dummy TimeZoneId to avoid UnresolvedException in CurrentBatchTimestamp
[SPARK-26495][SQL] Simplify the SelectedField extractor
[SPARK-26379][SS] Fix issue on adding current_timestamp/current_date to streaming query
[SPARK-26709][SQL] OptimizeMetadataOnlyQuery does not handle empty records correctly
[SPARK-26680][SQL] Eagerly create inputVars while conditions are appropriate
[SPARK-26682][SQL] Use taskAttemptID instead of attemptNumber for Hadoop.
[SPARK-26706][SQL] Fix illegalNumericPrecedence for ByteType
[SPARK-26228][MLLIB] OOM issue encountered when computing Gramian matrix
[SPARK-26665][CORE] Fix a bug that BlockTransferService.fetchBlockSync may hang forever
[SPARK-26549][PYSPARK] Fix for python worker reuse take no effect for parallelize lazy iterable range
[SPARK-26351][MLLIB] Update doc and minor correction in the mllib evaluation metrics
[SPARK-26450]Avoid rebuilding map of schema for every column in a projection
[SPARK-26638][PYSPARK][ML] Pyspark vector classes always return error for unary negation
[SPARK-26633][REPL] Add ExecutorClassLoader.getResourceAsStream
[SPARK-26629][SS] Fixed error with multiple file stream in a query + restart on a batch that has no data for one file stream
[SPARK-25992][PYTHON] Document SparkContext cannot be shared for multiprocessing
[SPARK-26615][CORE] Fixing transport server/client resource leaks in the core unittests
[SPARK-26350][SS] Allow to override group id of the Kafka consumer
[SPARK-26538][SQL] Set default precision and scale for elements of postgres numeric array
[SPARK-26586][SS] Fix race condition that causes streams to run with unexpected confs
[SPARK-26551][SQL] Fix schema pruning error when selecting one complex field and having is not null predicate on another one
[SPARK-26576][SQL] Broadcast hint not applied to partitioned table
[SPARK-26571][SQL] Update Hive Serde mapping with canonical name of Parquet and Orc FileFormat
[SPARK-26267][SS] Retry when detecting incorrect offsets from Kafka (2.4)
[SPARK-26559][ML][PYSPARK] ML image can't work with numpy versions prior to 1.9
[SPARK-26078][SQL] Dedup self-join attributes on IN subqueries

Maintenance updates

See Databricks Runtime 5.3 maintenance updates.

System environment

Operating System: Ubuntu 16.04.6 LTS
Java: 1.8.0_191
Scala: 2.11.12
Python: 2.7.12 for Python 2 clusters and 3.5.2 for Python 3 clusters.
R: R version 3.5.2 (2018-12-20)
GPU clusters: The following NVIDIA GPU libraries are installed:
- Tesla driver 375.66
- CUDA 9.0
- cuDNN 7.0