Databricks Runtime 4.3 (EoS)

Databricks released this version in August 2018.

The following release notes provide information about Databricks Runtime 4.3, powered by Apache Spark.

New features

• Delta Lake
  • TRUNCATE TABLE command: Delete all rows from a table. Unlike its counterpart for Spark tables, Delta tables do not support deleting specific partitions.
  • ALTER TABLE REPLACE COLUMNS command: Replace columns in a Delta table. It supports changing the comment of a column and reordering multiple columns.
  • FSCK REPAIR TABLE command: Remove the file entries from the transaction log of a Delta table that can no longer be found in the underlying file system. This can happen when these files have been manually deleted.
  • Support for queries on stale Delta tables to improve the interactive query experience: Queries on Delta tables can now run on a stale version of the table when up to date results are not necessary. This feature reduces the latency of queries especially when underlying Delta tables are updated continuously through streams.
• Structured Streaming
  • Scalable streaming write support for Azure Synapse Analytics Connector.
  • Support for foreachBatch() in Python (already available in Scala). See foreach and foreachBatch documentation for more details.
  • Support for choosing either the min or max watermark when there are multiple input streams in a query. Previously the minimum timestamp was always used. See the multiple watermark policy for more details.
  • Support for the LIMIT operator for streams in Append and Complete output modes. To minimize OOM errors on the driver, LIMIT is automatically applied when you use display() on unbounded streams.

Improvements

• Delta Lake

  • Private preview of new scalable implementation of MERGE INTO command that does not have the 10000 row insert limit. Contact support if you'd like to try this out.
  • Better performance and scalability of the OPTIMIZE command, especially on larger clusters.
  • The OPTIMIZE command now commits to the table incrementally, meaning that if the command fails, a retry will not need to process the entire data set.
  • Reduced the number of file system RPCs required to discover new data when using Delta Lake as a streaming source.
  • Added support for df.writeStream.table(table-name) in Python to create a Delta table from a stream.

• Improved performance for queries with multiple joins, aggregations, or windows.

• Improved efficiency for partition-level pruning in queries with broadcast hash joins.

• Improvements to whole stage code generation to detect duplicate expressions, reduce the amount of code generated, and improve performance for certain expression types.

• High concurrency clusters now support running %fs in notebooks.

• Upgraded Py4J used by PySpark to 0.10.7.

• Improved performance of disk caching on Azure Ls series instances. The cache is now enabled by default on these instances, accelerating workloads that repeatedly read Parquet files.

Deprecated features

• Data Skipping outside of Databricks Delta is deprecated. An enhanced version of data skipping will continue to be available as part of Delta Lake. We recommend that you switch to using Delta Lake to continue to take advantage of this feature. See Databricks Delta Data Skipping for details.

Bug fixes

• Fixed incorrect predicate pushdown MERGE INTO statement for Delta when the ON condition had predicates that referenced only the target table.

• Fixed bug in mapGroupsWithState and flatMapGroupsWithState that prevented setting timeouts when state has been removed (SPARK-22187).

• Fixed bug the prevented watermarking to work correctly with Trigger.Once (SPARK-24699).

• Update command now validates the columns in the SET clause to make sure all columns actually exist and no column is set more than once.

• Fixed a potential race condition that could cause deadlocks for directory commit.

• Fixed a bug causing a deprecated version of the DBFS client to be used when refreshing mounts.

Known issues

• Delta Lake configuration options for a table take effect only in the first notebook that loads the table.

Apache Spark

Databricks Runtime 4.3 includes Apache Spark 2.3.1. This release includes all fixes and improvements included in Databricks Runtime 4.2 (EoS), as well as the following additional bug fixes and improvements made to Spark:

• [SPARK-24934][SQL] Explicitly allow supported types in upper/lower bounds for in-memory partition pruning
  • When complex data types are used in query filters against cached data, Spark always returns an empty result set. The in-memory stats-based pruning generates incorrect results, because null is set for upper/lower bounds for complex types. The fix is to not use in-memory stats-based pruning for complex types.
• [SPARK-24957][SQL] Average with decimal followed by aggregation returns wrong result
  • The incorrect results of AVERAGE might be returned. The CAST added in the Average operator will be bypassed if the result of Divide is the same type which it is casted to.
• [SPARK-24867][SQL] Add AnalysisBarrier to DataFrameWriter
  • SQL cache is not being used when using DataFrameWriter to write a DataFrame with UDF. This is a regression caused by the changes we made in AnalysisBarrier, since not all the Analyzer rules are idempotent.
• [SPARK-24790][SQL] Allow complex aggregate expressions in Pivot
  • Relax the check to allow complex aggregate expressions, like ceil(sum(col1)) or sum(col1) + 1, which roughly means any aggregate expression that could appear in an Aggregate plan except pandas UDF.
• [SPARK-24870][SQL] Cache can't work normally if there are case letters in SQL
  • Fixes a plan canonicalization issue.
• [SPARK-24852]Have spark.ml training use updated Instrumentation APIs.
• [SPARK-24891][SQL] Fix HandleNullInputsForUDF rule
  • Make HandleNullInputsForUDF rule idempotent, to avoid plan mismatch in cache manager when a plan is analyzed more than once.
• [SPARK-24878][SQL] Fix reverse function for array type of primitive type containing null.
• [SPARK-24871][SQL] Refactor Concat and MapConcat to avoid creating concatenator object for each row.
• [SPARK-24802][SQL] Add a new config for Optimization Rule Exclusion
  • Provides a config to users to exclude some optimizer rules.
• [SPARK-24879][SQL] Fix NPE in Hive partition pruning filter pushdown
  • When the partition predicate is something like col IN (1, null), an NPE will be thrown. This patch fixes it.
• [SPARK-23731][SQL] Make FileSourceScanExec canonicalizable after being (de)serialized
• [SPARK-24755][CORE] Executor loss can cause task to not be resubmitted
  • Fixes a bug that Spark may not resubmit tasks failed by executor loss. This bug was introduced in Spark 2.3.
• [SPARK-24677][CORE] Avoid NoSuchElementException from MedianHeap
  • Fixes a speculative tasks related bug when collecting task duration metrics.
• [SPARK-24868][PYTHON] add sequence function in Python
• [SPARK-21811][SPARK-24012][SPARK-24737][SPARK-24165][SPARK-24734][SPARK-24840][SQL] Fix type coercions and nullabilities.
• [SPARK-24699][SS] Make watermarks work with Trigger.Once by saving updated watermark to commit log
• [SPARK-24537][R] Add array_remove / array_zip / map_from_arrays / array_distinct
• [SPARK-22187][SS] Update unsaferow format for saved state in flatMapGroupsWithState to allow timeouts with deleted state (4.x)
• [SPARK-24681][SQL] Verify nested column names in Hive metastore
  • Make sure nested column names do not include ',', ':', and ';' in Hive metastore
• [SPARK-23486]cache the function name from the external catalog for lookupFunctions
  • To speed up function lookups.
• [SPARK-24781][SQL] Using a reference from Dataset in Filter/Sort might not work
• [SPARK-24208][SQL] Fix attribute deduplication for FlatMapGroupsInPandas
  • Fix self-join failure on a dataset which contains a FlatMapGroupsInPandas because of duplicate attributes
• [SPARK-24530][PYTHON] Add a control to force Python version in Sphinx via environment variable, SPHINXPYTHON
• [SPARK-24250]support accessing SQLConf inside tasks
  • Save all the SQL configs to job properties when an SQL execution is triggered. At executor side we rebuild the SQLConf from job properties.
• [SPARK-23936][SQL] Implement map_concat
• [SPARK-23914][SQL] Add array_union function
• [SPARK-24732][SQL] Type coercion between MapTypes.
  • Adds support for type coercion between MapTypes where both the key types and the value types are compatible. For example, types MapType(IntegerType, FloatType) and MapType(LongType, DoubleType) can be coerced to type MapType(LongType, DoubleType)
• [SPARK-24662][SQL][SS] Support limit in structured streaming
• [SPARK-24730][SS] Add policy to choose max as global watermark when streaming query has multiple watermarks (branch-4.x)
• [SPARK-24596][SQL] Non-cascading Cache Invalidation
  • When uncache or drop temp view, it's unnecessary to cascadingly uncache all the plans that depend on the view, as the underlying data are not changed.
• [SPARK-23927][SQL] Add "sequence" expression
• [SPARK-24636][SQL] Type coercion of arrays for array_join function
• [SPARK-22384][SQL] Refine partition pruning when attribute is wrapped in Cast
  • Improve partition pruning, able to push down partition predicates with safe type cast(int to long, not long to int).
• [SPARK-24385][SQL] Resolve self-join condition ambiguity for EqualNullSafe
  • Implements EqualNullSafe for self-join condition ambiguity resolving.
• [SPARK-24696][SQL] ColumnPruning rule fails to remove extra Project
  • Fixes a bug in the ColumnPruning rule that caused an infinite loop error in the Optimizer.
• [SPARK-24603][SQL] Fix findTightestCommonType reference in comments
• [SPARK-24613][SQL] Cache with UDF could not be matched with subsequent dependent caches
  • Wraps the logical plan with a AnalysisBarrier for execution plan compilation in CacheManager, in order to avoid the plan being analyzed again. This is also a regression of Spark 2.3.
• [SPARK-24017][SQL] Refactor ExternalCatalog to be an interface
• [SPARK-24324][PYTHON] pandas Grouped Map UDF should assign result columns by name
  • Assigns result columns by schema name if user labeled with strings, otherwise using position.
• [SPARK-23778][CORE] Avoid unneeded shuffle when union gets an empty RDD
  • Ignores incoming empty RDDs in the union method to avoid an unneeded extra-shuffle when all the other RDDs have the same partitioning.
• [SPARK-24552][CORE][SQL] Use unique id instead of attempt number for writes.
  • Passes the unique task attempt id instead of attempt number to v2 data sources because attempt number is reused when stages are retried. This affects the data source V1 and V2 APIs, but the file format APIs will not be affected because DBR using different commit protocol.
• [SPARK-24588][SS] streaming join should require HashClusteredPartitioning from children.
• [SPARK-24589][CORE] Correctly identify tasks in output commit coordinator.
  • Adds more information to the stage state tracked by the coordinator, so that only one task is allowed to commit the output. This fix also removes the useless code changes introduced by SPARK-18113.
• [SPARK-23933][SQL] Add map_from_arrays function
• [SPARK-24583][SQL] Wrong schema type in InsertIntoDataSourceCommand
  • When creating a Delta table with NOT NULL constraints, we could drop nullability and insert the NULL values without checking the violation.
• [SPARK-24542][SQL] UDF series UDFXPathXXXX allow users to pass carefully crafted XML to access arbitrary files
  • This is a security patch reported from the community. UDF series UDFXPathXXXX allow users to pass carefully crafted XML to access arbitrary files. When users use the external access control library, users might bypass them and access the file contents.
• [SPARK-23934][SQL] Adding map_from_entries function
• [SPARK-23912][SQL] Add array_distinct
• [SPARK-24574][SQL] array_contains, array_position, array_remove and element_at functions deal with Column type

Maintenance updates

See Databricks Runtime 4.3 maintenance updates.

System environment

• Operating System: Ubuntu 16.04.4 LTS
• Java: 1.8.0_162
• Scala: 2.11.8
• Python: 2.7.12 for Python 2 clusters and 3.5.2 for Python 3 clusters.
• R: R version 3.4.4 (2018-03-15)
• GPU clusters: The following NVIDIA GPU libraries are installed:
  • Tesla driver 375.66
  • CUDA 9.0
  • cuDNN 7.0

Installed Python libraries

Installed R Libraries

Installed Java and Scala libraries (Scala 2.11 cluster version)