Skip to content

Latest commit

 

History

History
927 lines (715 loc) · 60.6 KB

sql-migration-guide.md

File metadata and controls

927 lines (715 loc) · 60.6 KB
Raw
layout title displayTitle license
global
Migration Guide: SQL, Datasets and DataFrame
Migration Guide: SQL, Datasets and DataFrame
Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to You under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
  • Table of contents {:toc}

Upgrading from Spark SQL 2.4 to 3.0

  • Since Spark 3.0, when inserting a value into a table column with a different data type, the type coercion is performed as per ANSI SQL standard. Certain unreasonable type conversions such as converting string to int and double to boolean are disallowed. A runtime exception will be thrown if the value is out-of-range for the data type of the column. In Spark version 2.4 and earlier, type conversions during table insertion are allowed as long as they are valid Cast. When inserting an out-of-range value to a integral field, the low-order bits of the value is inserted(the same as Java/Scala numeric type casting). For example, if 257 is inserted to a field of byte type, the result is 1. The behavior is controlled by the option spark.sql.storeAssignmentPolicy, with a default value as "ANSI". Setting the option as "Legacy" restores the previous behavior.

  • In Spark 3.0, the deprecated methods SQLContext.createExternalTable and SparkSession.createExternalTable have been removed in favor of its replacement, createTable.

  • In Spark 3.0, the deprecated HiveContext class has been removed. Use SparkSession.builder.enableHiveSupport() instead.

  • Since Spark 3.0, configuration spark.sql.crossJoin.enabled become internal configuration, and is true by default, so by default spark won't raise exception on sql with implicit cross join.

  • Since Spark 3.0, we reversed argument order of the trim function from TRIM(trimStr, str) to TRIM(str, trimStr) to be compatible with other databases.

  • In Spark version 2.4 and earlier, SQL queries such as FROM <table> or FROM <table> UNION ALL FROM <table> are supported by accident. In hive-style FROM <table> SELECT <expr>, the SELECT clause is not negligible. Neither Hive nor Presto support this syntax. Therefore we will treat these queries as invalid since Spark 3.0.

  • Since Spark 3.0, the Dataset and DataFrame API unionAll is not deprecated any more. It is an alias for union.

  • In Spark version 2.4 and earlier, the parser of JSON data source treats empty strings as null for some data types such as IntegerType. For FloatType and DoubleType, it fails on empty strings and throws exceptions. Since Spark 3.0, we disallow empty strings and will throw exceptions for data types except for StringType and BinaryType.

  • Since Spark 3.0, the from_json functions supports two modes - PERMISSIVE and FAILFAST. The modes can be set via the mode option. The default mode became PERMISSIVE. In previous versions, behavior of from_json did not conform to either PERMISSIVE nor FAILFAST, especially in processing of malformed JSON records. For example, the JSON string {"a" 1} with the schema a INT is converted to null by previous versions but Spark 3.0 converts it to Row(null).

  • The ADD JAR command previously returned a result set with the single value 0. It now returns an empty result set.

  • In Spark version 2.4 and earlier, users can create map values with map type key via built-in function like CreateMap, MapFromArrays, etc. Since Spark 3.0, it's not allowed to create map values with map type key with these built-in functions. Users can still read map values with map type key from data source or Java/Scala collections, though they are not very useful.

  • In Spark version 2.4 and earlier, Dataset.groupByKey results to a grouped dataset with key attribute wrongly named as "value", if the key is non-struct type, e.g. int, string, array, etc. This is counterintuitive and makes the schema of aggregation queries weird. For example, the schema of ds.groupByKey(...).count() is (value, count). Since Spark 3.0, we name the grouping attribute to "key". The old behaviour is preserved under a newly added configuration spark.sql.legacy.dataset.nameNonStructGroupingKeyAsValue with a default value of false.

  • In Spark version 2.4 and earlier, float/double -0.0 is semantically equal to 0.0, but -0.0 and 0.0 are considered as different values when used in aggregate grouping keys, window partition keys and join keys. Since Spark 3.0, this bug is fixed. For example, Seq(-0.0, 0.0).toDF("d").groupBy("d").count() returns [(0.0, 2)] in Spark 3.0, and [(0.0, 1), (-0.0, 1)] in Spark 2.4 and earlier.

  • In Spark version 2.4 and earlier, users can create a map with duplicated keys via built-in functions like CreateMap, StringToMap, etc. The behavior of map with duplicated keys is undefined, e.g. map look up respects the duplicated key appears first, Dataset.collect only keeps the duplicated key appears last, MapKeys returns duplicated keys, etc. Since Spark 3.0, these built-in functions will remove duplicated map keys with last wins policy. Users may still read map values with duplicated keys from data sources which do not enforce it (e.g. Parquet), the behavior will be undefined.

  • In Spark version 2.4 and earlier, partition column value is converted as null if it can't be casted to corresponding user provided schema. Since 3.0, partition column value is validated with user provided schema. An exception is thrown if the validation fails. You can disable such validation by setting spark.sql.sources.validatePartitionColumns to false.

  • In Spark version 2.4 and earlier, the SET command works without any warnings even if the specified key is for SparkConf entries and it has no effect because the command does not update SparkConf, but the behavior might confuse users. Since 3.0, the command fails if a SparkConf key is used. You can disable such a check by setting spark.sql.legacy.setCommandRejectsSparkCoreConfs to false.

  • In Spark version 2.4 and earlier, CSV datasource converts a malformed CSV string to a row with all nulls in the PERMISSIVE mode. Since Spark 3.0, the returned row can contain non-null fields if some of CSV column values were parsed and converted to desired types successfully.

  • In Spark version 2.4 and earlier, JSON datasource and JSON functions like from_json convert a bad JSON record to a row with all nulls in the PERMISSIVE mode when specified schema is StructType. Since Spark 3.0, the returned row can contain non-null fields if some of JSON column values were parsed and converted to desired types successfully.

  • Refreshing a cached table would trigger a table uncache operation and then a table cache (lazily) operation. In Spark version 2.4 and earlier, the cache name and storage level are not preserved before the uncache operation. Therefore, the cache name and storage level could be changed unexpectedly. Since Spark 3.0, cache name and storage level will be first preserved for cache recreation. It helps to maintain a consistent cache behavior upon table refreshing.

  • Since Spark 3.0, JSON datasource and JSON function schema_of_json infer TimestampType from string values if they match to the pattern defined by the JSON option timestampFormat. Set JSON option inferTimestamp to false to disable such type inferring.

  • In Spark version 2.4 and earlier, if org.apache.spark.sql.functions.udf(Any, DataType) gets a Scala closure with primitive-type argument, the returned UDF will return null if the input values is null. Since Spark 3.0, the UDF will return the default value of the Java type if the input value is null. For example, val f = udf((x: Int) => x, IntegerType), f($"x") will return null in Spark 2.4 and earlier if column x is null, and return 0 in Spark 3.0. This behavior change is introduced because Spark 3.0 is built with Scala 2.12 by default.

  • Since Spark 3.0, Proleptic Gregorian calendar is used in parsing, formatting, and converting dates and timestamps as well as in extracting sub-components like years, days and etc. Spark 3.0 uses Java 8 API classes from the java.time packages that based on ISO chronology (https://docs.oracle.com/javase/8/docs/api/java/time/chrono/IsoChronology.html). In Spark version 2.4 and earlier, those operations are performed by using the hybrid calendar (Julian + Gregorian, see https://docs.oracle.com/javase/7/docs/api/java/util/GregorianCalendar.html). The changes impact on the results for dates before October 15, 1582 (Gregorian) and affect on the following Spark 3.0 API:

    • CSV/JSON datasources use java.time API for parsing and generating CSV/JSON content. In Spark version 2.4 and earlier, java.text.SimpleDateFormat is used for the same purpose with fallbacks to the parsing mechanisms of Spark 2.0 and 1.x. For example, 2018-12-08 10:39:21.123 with the pattern yyyy-MM-dd'T'HH:mm:ss.SSS cannot be parsed since Spark 3.0 because the timestamp does not match to the pattern but it can be parsed by earlier Spark versions due to a fallback to Timestamp.valueOf. To parse the same timestamp since Spark 3.0, the pattern should be yyyy-MM-dd HH:mm:ss.SSS.

    • The unix_timestamp, date_format, to_unix_timestamp, from_unixtime, to_date, to_timestamp functions. New implementation supports pattern formats as described here https://docs.oracle.com/javase/8/docs/api/java/time/format/DateTimeFormatter.html and performs strict checking of its input. For example, the 2015-07-22 10:00:00 timestamp cannot be parse if pattern is yyyy-MM-dd because the parser does not consume whole input. Another example is the 31/01/2015 00:00 input cannot be parsed by the dd/MM/yyyy hh:mm pattern because hh supposes hours in the range 1-12.

    • The weekofyear, weekday, dayofweek, date_trunc, from_utc_timestamp, to_utc_timestamp, and unix_timestamp functions use java.time API for calculation week number of year, day number of week as well for conversion from/to TimestampType values in UTC time zone.

    • the JDBC options lowerBound and upperBound are converted to TimestampType/DateType values in the same way as casting strings to TimestampType/DateType values. The conversion is based on Proleptic Gregorian calendar, and time zone defined by the SQL config spark.sql.session.timeZone. In Spark version 2.4 and earlier, the conversion is based on the hybrid calendar (Julian + Gregorian) and on default system time zone.

    • Formatting of TIMESTAMP and DATE literals.

  • In Spark version 2.4 and earlier, invalid time zone ids are silently ignored and replaced by GMT time zone, for example, in the from_utc_timestamp function. Since Spark 3.0, such time zone ids are rejected, and Spark throws java.time.DateTimeException.

  • In Spark version 2.4 and earlier, the current_timestamp function returns a timestamp with millisecond resolution only. Since Spark 3.0, the function can return the result with microsecond resolution if the underlying clock available on the system offers such resolution.

  • In Spark version 2.4 and earlier, when reading a Hive Serde table with Spark native data sources(parquet/orc), Spark will infer the actual file schema and update the table schema in metastore. Since Spark 3.0, Spark doesn't infer the schema anymore. This should not cause any problems to end users, but if it does, please set spark.sql.hive.caseSensitiveInferenceMode to INFER_AND_SAVE.

  • Since Spark 3.0, TIMESTAMP literals are converted to strings using the SQL config spark.sql.session.timeZone. In Spark version 2.4 and earlier, the conversion uses the default time zone of the Java virtual machine.

  • In Spark version 2.4, when a spark session is created via cloneSession(), the newly created spark session inherits its configuration from its parent SparkContext even though the same configuration may exist with a different value in its parent spark session. Since Spark 3.0, the configurations of a parent SparkSession have a higher precedence over the parent SparkContext. The old behavior can be restored by setting spark.sql.legacy.sessionInitWithConfigDefaults to true.

  • Since Spark 3.0, parquet logical type TIMESTAMP_MICROS is used by default while saving TIMESTAMP columns. In Spark version 2.4 and earlier, TIMESTAMP columns are saved as INT96 in parquet files. To set INT96 to spark.sql.parquet.outputTimestampType restores the previous behavior.

  • Since Spark 3.0, if hive.default.fileformat is not found in Spark SQL configuration then it will fallback to hive-site.xml present in the Hadoop configuration of SparkContext.

  • Since Spark 3.0, Spark will cast String to Date/TimeStamp in binary comparisons with dates/timestamps. The previous behaviour of casting Date/Timestamp to String can be restored by setting spark.sql.legacy.typeCoercion.datetimeToString to true.

  • Since Spark 3.0, when Avro files are written with user provided schema, the fields will be matched by field names between catalyst schema and avro schema instead of positions.

  • Since Spark 3.0, when Avro files are written with user provided non-nullable schema, even the catalyst schema is nullable, Spark is still able to write the files. However, Spark will throw runtime NPE if any of the records contains null.

  • Since Spark 3.0, we use a new protocol for fetching shuffle blocks, for external shuffle service users, we need to upgrade the server correspondingly. Otherwise, we'll get the error message UnsupportedOperationException: Unexpected message: FetchShuffleBlocks. If it is hard to upgrade the shuffle service right now, you can still use the old protocol by setting spark.shuffle.useOldFetchProtocol to true.

  • Since Spark 3.0, a higher-order function exists follows the three-valued boolean logic, i.e., if the predicate returns any nulls and no true is obtained, then exists will return null instead of false. For example, exists(array(1, null, 3), x -> x % 2 == 0) will be null. The previous behaviour can be restored by setting spark.sql.legacy.arrayExistsFollowsThreeValuedLogic to false.

  • Since Spark 3.0, if files or subdirectories disappear during recursive directory listing (i.e. they appear in an intermediate listing but then cannot be read or listed during later phases of the recursive directory listing, due to either concurrent file deletions or object store consistency issues) then the listing will fail with an exception unless spark.sql.files.ignoreMissingFiles is true (default false). In previous versions, these missing files or subdirectories would be ignored. Note that this change of behavior only applies during initial table file listing (or during REFRESH TABLE), not during query execution: the net change is that spark.sql.files.ignoreMissingFiles is now obeyed during table file listing / query planning, not only at query execution time.

  • Since Spark 3.0, substitution order of nested WITH clauses is changed and an inner CTE definition takes precedence over an outer. In version 2.4 and earlier, WITH t AS (SELECT 1), t2 AS (WITH t AS (SELECT 2) SELECT * FROM t) SELECT * FROM t2 returns 1 while in version 3.0 it returns 2. The previous behaviour can be restored by setting spark.sql.legacy.ctePrecedence.enabled to true.

  • Since Spark 3.0, the add_months function does not adjust the resulting date to a last day of month if the original date is a last day of months. For example, select add_months(DATE'2019-02-28', 1) results 2019-03-28. In Spark version 2.4 and earlier, the resulting date is adjusted when the original date is a last day of months. For example, adding a month to 2019-02-28 results in 2019-03-31.

  • Since Spark 3.0, 0-argument Java UDF is executed in the executor side identically with other UDFs. In Spark version 2.4 and earlier, 0-argument Java UDF alone was executed in the driver side, and the result was propagated to executors, which might be more performant in some cases but caused inconsistency with a correctness issue in some cases.

  • The result of java.lang.Math's log, log1p, exp, expm1, and pow may vary across platforms. In Spark 3.0, the result of the equivalent SQL functions (including related SQL functions like LOG10) return values consistent with java.lang.StrictMath. In virtually all cases this makes no difference in the return value, and the difference is very small, but may not exactly match java.lang.Math on x86 platforms in cases like, for example, log(3.0), whose value varies between Math.log() and StrictMath.log().

  • Since Spark 3.0, Dataset query fails if it contains ambiguous column reference that is caused by self join. A typical example: val df1 = ...; val df2 = df1.filter(...);, then df1.join(df2, df1("a") > df2("a")) returns an empty result which is quite confusing. This is because Spark cannot resolve Dataset column references that point to tables being self joined, and df1("a") is exactly the same as df2("a") in Spark. To restore the behavior before Spark 3.0, you can set spark.sql.analyzer.failAmbiguousSelfJoin to false.

  • Since Spark 3.0, Cast function processes string literals such as 'Infinity', '+Infinity', '-Infinity', 'NaN', 'Inf', '+Inf', '-Inf' in case insensitive manner when casting the literals to Double or Float type to ensure greater compatibility with other database systems. This behaviour change is illustrated in the table below:

    Operation Result prior to Spark 3.0 Result starting Spark 3.0
    CAST('infinity' AS DOUBLE)
    CAST('+infinity' AS DOUBLE)
    CAST('inf' AS DOUBLE)
    CAST('+inf' AS DOUBLE)
    		 NULL Double.PositiveInfinity
    CAST('-infinity' AS DOUBLE)
    CAST('-inf' AS DOUBLE)
    		 NULL Double.NegativeInfinity
    CAST('infinity' AS FLOAT)
    CAST('+infinity' AS FLOAT)
    CAST('inf' AS FLOAT)
    CAST('+inf' AS FLOAT)
    		 NULL Float.PositiveInfinity
    CAST('-infinity' AS FLOAT)
    CAST('-inf' AS FLOAT)
    		 NULL Float.NegativeInfinity
    CAST('nan' AS DOUBLE) NULL Double.NaN
    CAST('nan' AS FLOAT) NULL Float.NaN

  • Since Spark 3.0, special values are supported in conversion from strings to dates and timestamps. Those values are simply notational shorthands that will be converted to ordinary date or timestamp values when read. The following string values are supported for dates:

    • epoch [zoneId] - 1970-01-01
    • today [zoneId] - the current date in the time zone specified by spark.sql.session.timeZone
    • yesterday [zoneId] - the current date - 1
    • tomorrow [zoneId] - the current date + 1
    • now - the date of running the current query. It has the same notion as today For example SELECT date 'tomorrow' - date 'yesterday'; should output 2. Here are special timestamp values:
    • epoch [zoneId] - 1970-01-01 00:00:00+00 (Unix system time zero)
    • today [zoneId] - midnight today
    • yesterday [zoneId] - midnight yesterday
    • tomorrow [zoneId] - midnight tomorrow
    • now - current query start time For example SELECT timestamp 'tomorrow';.

  • Since Spark 3.0, the size function returns NULL for the NULL input. In Spark version 2.4 and earlier, this function gives -1 for the same input. To restore the behavior before Spark 3.0, you can set spark.sql.legacy.sizeOfNull to true.

Upgrading from Spark SQL 2.4 to 2.4.1

  • The value of spark.executor.heartbeatInterval, when specified without units like "30" rather than "30s", was inconsistently interpreted as both seconds and milliseconds in Spark 2.4.0 in different parts of the code. Unitless values are now consistently interpreted as milliseconds. Applications that set values like "30" need to specify a value with units like "30s" now, to avoid being interpreted as milliseconds; otherwise, the extremely short interval that results will likely cause applications to fail.

  • When turning a Dataset to another Dataset, Spark will up cast the fields in the original Dataset to the type of corresponding fields in the target DataSet. In version 2.4 and earlier, this up cast is not very strict, e.g. Seq("str").toDS.as[Int] fails, but Seq("str").toDS.as[Boolean] works and throw NPE during execution. In Spark 3.0, the up cast is stricter and turning String into something else is not allowed, i.e. Seq("str").toDS.as[Boolean] will fail during analysis.

Upgrading from Spark SQL 2.3 to 2.4

  • In Spark version 2.3 and earlier, the second parameter to array_contains function is implicitly promoted to the element type of first array type parameter. This type promotion can be lossy and may cause array_contains function to return wrong result. This problem has been addressed in 2.4 by employing a safer type promotion mechanism. This can cause some change in behavior and are illustrated in the table below.

    Query Spark 2.3 or Prior Spark 2.4 Remarks
    SELECT array_contains(array(1), 1.34D); true false In Spark 2.4, left and right parameters are promoted to array type of double type and double type respectively.
    SELECT array_contains(array(1), '1'); true AnalysisException is thrown. Explicit cast can be used in arguments to avoid the exception. In Spark 2.4, AnalysisException is thrown since integer type can not be promoted to string type in a loss-less manner.
    SELECT array_contains(array(1), 'anystring'); null AnalysisException is thrown. Explicit cast can be used in arguments to avoid the exception. In Spark 2.4, AnalysisException is thrown since integer type can not be promoted to string type in a loss-less manner.

  • Since Spark 2.4, when there is a struct field in front of the IN operator before a subquery, the inner query must contain a struct field as well. In previous versions, instead, the fields of the struct were compared to the output of the inner query. Eg. if a is a struct(a string, b int), in Spark 2.4 a in (select (1 as a, 'a' as b) from range(1)) is a valid query, while a in (select 1, 'a' from range(1)) is not. In previous version it was the opposite.

  • In versions 2.2.1+ and 2.3, if spark.sql.caseSensitive is set to true, then the CURRENT_DATE and CURRENT_TIMESTAMP functions incorrectly became case-sensitive and would resolve to columns (unless typed in lower case). In Spark 2.4 this has been fixed and the functions are no longer case-sensitive.

  • Since Spark 2.4, Spark will evaluate the set operations referenced in a query by following a precedence rule as per the SQL standard. If the order is not specified by parentheses, set operations are performed from left to right with the exception that all INTERSECT operations are performed before any UNION, EXCEPT or MINUS operations. The old behaviour of giving equal precedence to all the set operations are preserved under a newly added configuration spark.sql.legacy.setopsPrecedence.enabled with a default value of false. When this property is set to true, spark will evaluate the set operators from left to right as they appear in the query given no explicit ordering is enforced by usage of parenthesis.

  • Since Spark 2.4, Spark will display table description column Last Access value as UNKNOWN when the value was Jan 01 1970.

  • Since Spark 2.4, Spark maximizes the usage of a vectorized ORC reader for ORC files by default. To do that, spark.sql.orc.impl and spark.sql.orc.filterPushdown change their default values to native and true respectively. ORC files created by native ORC writer cannot be read by some old Apache Hive releases. Use spark.sql.orc.impl=hive to create the files shared with Hive 2.1.1 and older.

  • Since Spark 2.4, writing an empty dataframe to a directory launches at least one write task, even if physically the dataframe has no partition. This introduces a small behavior change that for self-describing file formats like Parquet and Orc, Spark creates a metadata-only file in the target directory when writing a 0-partition dataframe, so that schema inference can still work if users read that directory later. The new behavior is more reasonable and more consistent regarding writing empty dataframe.

  • Since Spark 2.4, expression IDs in UDF arguments do not appear in column names. For example, a column name in Spark 2.4 is not UDF:f(col0 AS colA#28) but UDF:f(col0 AS `colA`).

  • Since Spark 2.4, writing a dataframe with an empty or nested empty schema using any file formats (parquet, orc, json, text, csv etc.) is not allowed. An exception is thrown when attempting to write dataframes with empty schema.

  • Since Spark 2.4, Spark compares a DATE type with a TIMESTAMP type after promotes both sides to TIMESTAMP. To set false to spark.sql.legacy.compareDateTimestampInTimestamp restores the previous behavior. This option will be removed in Spark 3.0.

  • Since Spark 2.4, creating a managed table with nonempty location is not allowed. An exception is thrown when attempting to create a managed table with nonempty location. To set true to spark.sql.legacy.allowCreatingManagedTableUsingNonemptyLocation restores the previous behavior. This option will be removed in Spark 3.0.

  • Since Spark 2.4, renaming a managed table to existing location is not allowed. An exception is thrown when attempting to rename a managed table to existing location.

  • Since Spark 2.4, the type coercion rules can automatically promote the argument types of the variadic SQL functions (e.g., IN/COALESCE) to the widest common type, no matter how the input arguments order. In prior Spark versions, the promotion could fail in some specific orders (e.g., TimestampType, IntegerType and StringType) and throw an exception.

  • Since Spark 2.4, Spark has enabled non-cascading SQL cache invalidation in addition to the traditional cache invalidation mechanism. The non-cascading cache invalidation mechanism allows users to remove a cache without impacting its dependent caches. This new cache invalidation mechanism is used in scenarios where the data of the cache to be removed is still valid, e.g., calling unpersist() on a Dataset, or dropping a temporary view. This allows users to free up memory and keep the desired caches valid at the same time.

  • In version 2.3 and earlier, Spark converts Parquet Hive tables by default but ignores table properties like TBLPROPERTIES (parquet.compression 'NONE'). This happens for ORC Hive table properties like TBLPROPERTIES (orc.compress 'NONE') in case of spark.sql.hive.convertMetastoreOrc=true, too. Since Spark 2.4, Spark respects Parquet/ORC specific table properties while converting Parquet/ORC Hive tables. As an example, CREATE TABLE t(id int) STORED AS PARQUET TBLPROPERTIES (parquet.compression 'NONE') would generate Snappy parquet files during insertion in Spark 2.3, and in Spark 2.4, the result would be uncompressed parquet files.

  • Since Spark 2.0, Spark converts Parquet Hive tables by default for better performance. Since Spark 2.4, Spark converts ORC Hive tables by default, too. It means Spark uses its own ORC support by default instead of Hive SerDe. As an example, CREATE TABLE t(id int) STORED AS ORC would be handled with Hive SerDe in Spark 2.3, and in Spark 2.4, it would be converted into Spark's ORC data source table and ORC vectorization would be applied. To set false to spark.sql.hive.convertMetastoreOrc restores the previous behavior.

  • In version 2.3 and earlier, CSV rows are considered as malformed if at least one column value in the row is malformed. CSV parser dropped such rows in the DROPMALFORMED mode or outputs an error in the FAILFAST mode. Since Spark 2.4, CSV row is considered as malformed only when it contains malformed column values requested from CSV datasource, other values can be ignored. As an example, CSV file contains the "id,name" header and one row "1234". In Spark 2.4, selection of the id column consists of a row with one column value 1234 but in Spark 2.3 and earlier it is empty in the DROPMALFORMED mode. To restore the previous behavior, set spark.sql.csv.parser.columnPruning.enabled to false.

  • Since Spark 2.4, File listing for compute statistics is done in parallel by default. This can be disabled by setting spark.sql.statistics.parallelFileListingInStatsComputation.enabled to False.

  • Since Spark 2.4, Metadata files (e.g. Parquet summary files) and temporary files are not counted as data files when calculating table size during Statistics computation.

  • Since Spark 2.4, empty strings are saved as quoted empty strings "". In version 2.3 and earlier, empty strings are equal to null values and do not reflect to any characters in saved CSV files. For example, the row of "a", null, "", 1 was written as a,,,1. Since Spark 2.4, the same row is saved as a,,"",1. To restore the previous behavior, set the CSV option emptyValue to empty (not quoted) string.

  • Since Spark 2.4, The LOAD DATA command supports wildcard ? and *, which match any one character, and zero or more characters, respectively. Example: LOAD DATA INPATH '/tmp/folder*/' or LOAD DATA INPATH '/tmp/part-?'. Special Characters like space also now work in paths. Example: LOAD DATA INPATH '/tmp/folder name/'.

  • In Spark version 2.3 and earlier, HAVING without GROUP BY is treated as WHERE. This means, SELECT 1 FROM range(10) HAVING true is executed as SELECT 1 FROM range(10) WHERE true and returns 10 rows. This violates SQL standard, and has been fixed in Spark 2.4. Since Spark 2.4, HAVING without GROUP BY is treated as a global aggregate, which means SELECT 1 FROM range(10) HAVING true will return only one row. To restore the previous behavior, set spark.sql.legacy.parser.havingWithoutGroupByAsWhere to true.

  • In version 2.3 and earlier, when reading from a Parquet data source table, Spark always returns null for any column whose column names in Hive metastore schema and Parquet schema are in different letter cases, no matter whether spark.sql.caseSensitive is set to true or false. Since 2.4, when spark.sql.caseSensitive is set to false, Spark does case insensitive column name resolution between Hive metastore schema and Parquet schema, so even column names are in different letter cases, Spark returns corresponding column values. An exception is thrown if there is ambiguity, i.e. more than one Parquet column is matched. This change also applies to Parquet Hive tables when spark.sql.hive.convertMetastoreParquet is set to true.

Upgrading from Spark SQL 2.2 to 2.3

  • Since Spark 2.3, the queries from raw JSON/CSV files are disallowed when the referenced columns only include the internal corrupt record column (named _corrupt_record by default). For example, spark.read.schema(schema).json(file).filter($"_corrupt_record".isNotNull).count() and spark.read.schema(schema).json(file).select("_corrupt_record").show(). Instead, you can cache or save the parsed results and then send the same query. For example, val df = spark.read.schema(schema).json(file).cache() and then df.filter($"_corrupt_record".isNotNull).count().

  • The percentile_approx function previously accepted numeric type input and output double type results. Now it supports date type, timestamp type and numeric types as input types. The result type is also changed to be the same as the input type, which is more reasonable for percentiles.

  • Since Spark 2.3, the Join/Filter's deterministic predicates that are after the first non-deterministic predicates are also pushed down/through the child operators, if possible. In prior Spark versions, these filters are not eligible for predicate pushdown.

  • Partition column inference previously found incorrect common type for different inferred types, for example, previously it ended up with double type as the common type for double type and date type. Now it finds the correct common type for such conflicts. The conflict resolution follows the table below:

    InputA \ InputB NullType IntegerType LongType DecimalType(38,0)* DoubleType DateType TimestampType StringType
    NullType NullType IntegerType LongType DecimalType(38,0) DoubleType DateType TimestampType StringType
    IntegerType IntegerType IntegerType LongType DecimalType(38,0) DoubleType StringType StringType StringType
    LongType LongType LongType LongType DecimalType(38,0) StringType StringType StringType StringType
    DecimalType(38,0)* DecimalType(38,0) DecimalType(38,0) DecimalType(38,0) DecimalType(38,0) StringType StringType StringType StringType
    DoubleType DoubleType DoubleType StringType StringType DoubleType StringType StringType StringType
    DateType DateType StringType StringType StringType StringType DateType TimestampType StringType
    TimestampType TimestampType StringType StringType StringType StringType TimestampType TimestampType StringType
    StringType StringType StringType StringType StringType StringType StringType StringType StringType

    Note that, for DecimalType(38,0)*, the table above intentionally does not cover all other combinations of scales and precisions because currently we only infer decimal type like BigInteger/BigInt. For example, 1.1 is inferred as double type.

  • Since Spark 2.3, when either broadcast hash join or broadcast nested loop join is applicable, we prefer to broadcasting the table that is explicitly specified in a broadcast hint. For details, see the section Broadcast Hint and SPARK-22489.

  • Since Spark 2.3, when all inputs are binary, functions.concat() returns an output as binary. Otherwise, it returns as a string. Until Spark 2.3, it always returns as a string despite of input types. To keep the old behavior, set spark.sql.function.concatBinaryAsString to true.

  • Since Spark 2.3, when all inputs are binary, SQL elt() returns an output as binary. Otherwise, it returns as a string. Until Spark 2.3, it always returns as a string despite of input types. To keep the old behavior, set spark.sql.function.eltOutputAsString to true.

  • Since Spark 2.3, by default arithmetic operations between decimals return a rounded value if an exact representation is not possible (instead of returning NULL). This is compliant with SQL ANSI 2011 specification and Hive's new behavior introduced in Hive 2.2 (HIVE-15331). This involves the following changes

    • The rules to determine the result type of an arithmetic operation have been updated. In particular, if the precision / scale needed are out of the range of available values, the scale is reduced up to 6, in order to prevent the truncation of the integer part of the decimals. All the arithmetic operations are affected by the change, ie. addition (+), subtraction (-), multiplication (*), division (/), remainder (%) and positive module (pmod).

    • Literal values used in SQL operations are converted to DECIMAL with the exact precision and scale needed by them.

    • The configuration spark.sql.decimalOperations.allowPrecisionLoss has been introduced. It defaults to true, which means the new behavior described here; if set to false, Spark uses previous rules, ie. it doesn't adjust the needed scale to represent the values and it returns NULL if an exact representation of the value is not possible.

  • Un-aliased subquery's semantic has not been well defined with confusing behaviors. Since Spark 2.3, we invalidate such confusing cases, for example: SELECT v.i from (SELECT i FROM v), Spark will throw an analysis exception in this case because users should not be able to use the qualifier inside a subquery. See SPARK-20690 and SPARK-21335 for more details.

  • When creating a SparkSession with SparkSession.builder.getOrCreate(), if there is an existing SparkContext, the builder was trying to update the SparkConf of the existing SparkContext with configurations specified to the builder, but the SparkContext is shared by all SparkSessions, so we should not update them. Since 2.3, the builder comes to not update the configurations. If you want to update them, you need to update them prior to creating a SparkSession.

Upgrading from Spark SQL 2.1 to 2.2

  • Spark 2.1.1 introduced a new configuration key: spark.sql.hive.caseSensitiveInferenceMode. It had a default setting of NEVER_INFER, which kept behavior identical to 2.1.0. However, Spark 2.2.0 changes this setting's default value to INFER_AND_SAVE to restore compatibility with reading Hive metastore tables whose underlying file schema have mixed-case column names. With the INFER_AND_SAVE configuration value, on first access Spark will perform schema inference on any Hive metastore table for which it has not already saved an inferred schema. Note that schema inference can be a very time-consuming operation for tables with thousands of partitions. If compatibility with mixed-case column names is not a concern, you can safely set spark.sql.hive.caseSensitiveInferenceMode to NEVER_INFER to avoid the initial overhead of schema inference. Note that with the new default INFER_AND_SAVE setting, the results of the schema inference are saved as a metastore key for future use. Therefore, the initial schema inference occurs only at a table's first access.

  • Since Spark 2.2.1 and 2.3.0, the schema is always inferred at runtime when the data source tables have the columns that exist in both partition schema and data schema. The inferred schema does not have the partitioned columns. When reading the table, Spark respects the partition values of these overlapping columns instead of the values stored in the data source files. In 2.2.0 and 2.1.x release, the inferred schema is partitioned but the data of the table is invisible to users (i.e., the result set is empty).

  • Since Spark 2.2, view definitions are stored in a different way from prior versions. This may cause Spark unable to read views created by prior versions. In such cases, you need to recreate the views using ALTER VIEW AS or CREATE OR REPLACE VIEW AS with newer Spark versions.

Upgrading from Spark SQL 2.0 to 2.1

  • Datasource tables now store partition metadata in the Hive metastore. This means that Hive DDLs such as ALTER TABLE PARTITION ... SET LOCATION are now available for tables created with the Datasource API.

    • Legacy datasource tables can be migrated to this format via the MSCK REPAIR TABLE command. Migrating legacy tables is recommended to take advantage of Hive DDL support and improved planning performance.

    • To determine if a table has been migrated, look for the PartitionProvider: Catalog attribute when issuing DESCRIBE FORMATTED on the table.

  • Changes to INSERT OVERWRITE TABLE ... PARTITION ... behavior for Datasource tables.

    • In prior Spark versions INSERT OVERWRITE overwrote the entire Datasource table, even when given a partition specification. Now only partitions matching the specification are overwritten.

    • Note that this still differs from the behavior of Hive tables, which is to overwrite only partitions overlapping with newly inserted data.

Upgrading from Spark SQL 1.6 to 2.0

  • SparkSession is now the new entry point of Spark that replaces the old SQLContext and

    HiveContext. Note that the old SQLContext and HiveContext are kept for backward compatibility. A new catalog interface is accessible from SparkSession - existing API on databases and tables access such as listTables, createExternalTable, dropTempView, cacheTable are moved here.

  • Dataset API and DataFrame API are unified. In Scala, DataFrame becomes a type alias for Dataset[Row], while Java API users must replace DataFrame with Dataset<Row>. Both the typed transformations (e.g., map, filter, and groupByKey) and untyped transformations (e.g., select and groupBy) are available on the Dataset class. Since compile-time type-safety in Python and R is not a language feature, the concept of Dataset does not apply to these languages’ APIs. Instead, DataFrame remains the primary programming abstraction, which is analogous to the single-node data frame notion in these languages.

  • Dataset and DataFrame API unionAll has been deprecated and replaced by union

  • Dataset and DataFrame API explode has been deprecated, alternatively, use functions.explode() with select or flatMap

  • Dataset and DataFrame API registerTempTable has been deprecated and replaced by createOrReplaceTempView

  • Changes to CREATE TABLE ... LOCATION behavior for Hive tables.

    • From Spark 2.0, CREATE TABLE ... LOCATION is equivalent to CREATE EXTERNAL TABLE ... LOCATION in order to prevent accidental dropping the existing data in the user-provided locations. That means, a Hive table created in Spark SQL with the user-specified location is always a Hive external table. Dropping external tables will not remove the data. Users are not allowed to specify the location for Hive managed tables. Note that this is different from the Hive behavior.

    • As a result, DROP TABLE statements on those tables will not remove the data.

  • spark.sql.parquet.cacheMetadata is no longer used. See SPARK-13664 for details.

Upgrading from Spark SQL 1.5 to 1.6

  • From Spark 1.6, by default, the Thrift server runs in multi-session mode. Which means each JDBC/ODBC connection owns a copy of their own SQL configuration and temporary function registry. Cached tables are still shared though. If you prefer to run the Thrift server in the old single-session mode, please set option spark.sql.hive.thriftServer.singleSession to true. You may either add this option to spark-defaults.conf, or pass it to start-thriftserver.sh via --conf:

    {% highlight bash %} ./sbin/start-thriftserver.sh
    --conf spark.sql.hive.thriftServer.singleSession=true
    ... {% endhighlight %}

  • From Spark 1.6, LongType casts to TimestampType expect seconds instead of microseconds. This change was made to match the behavior of Hive 1.2 for more consistent type casting to TimestampType from numeric types. See SPARK-11724 for details.

Upgrading from Spark SQL 1.4 to 1.5

  • Optimized execution using manually managed memory (Tungsten) is now enabled by default, along with code generation for expression evaluation. These features can both be disabled by setting spark.sql.tungsten.enabled to false.

  • Parquet schema merging is no longer enabled by default. It can be re-enabled by setting spark.sql.parquet.mergeSchema to true.

  • In-memory columnar storage partition pruning is on by default. It can be disabled by setting spark.sql.inMemoryColumnarStorage.partitionPruning to false.

  • Unlimited precision decimal columns are no longer supported, instead Spark SQL enforces a maximum precision of 38. When inferring schema from BigDecimal objects, a precision of (38, 18) is now used. When no precision is specified in DDL then the default remains Decimal(10, 0).

  • Timestamps are now stored at a precision of 1us, rather than 1ns

  • In the sql dialect, floating point numbers are now parsed as decimal. HiveQL parsing remains unchanged.

  • The canonical name of SQL/DataFrame functions are now lower case (e.g., sum vs SUM).

  • JSON data source will not automatically load new files that are created by other applications (i.e. files that are not inserted to the dataset through Spark SQL). For a JSON persistent table (i.e. the metadata of the table is stored in Hive Metastore), users can use REFRESH TABLE SQL command or HiveContext's refreshTable method to include those new files to the table. For a DataFrame representing a JSON dataset, users need to recreate the DataFrame and the new DataFrame will include new files.

Upgrading from Spark SQL 1.3 to 1.4

DataFrame data reader/writer interface

{:.no_toc}

Based on user feedback, we created a new, more fluid API for reading data in (SQLContext.read) and writing data out (DataFrame.write), and deprecated the old APIs (e.g., SQLContext.parquetFile, SQLContext.jsonFile).

See the API docs for SQLContext.read ( Scala, Java, Python ) and DataFrame.write ( Scala, Java, Python ) more information.

DataFrame.groupBy retains grouping columns

{:.no_toc}

Based on user feedback, we changed the default behavior of DataFrame.groupBy().agg() to retain the grouping columns in the resulting DataFrame. To keep the behavior in 1.3, set spark.sql.retainGroupColumns to false.

{% highlight scala %}

// In 1.3.x, in order for the grouping column "department" to show up, // it must be included explicitly as part of the agg function call. df.groupBy("department").agg($"department", max("age"), sum("expense"))

// In 1.4+, grouping column "department" is included automatically. df.groupBy("department").agg(max("age"), sum("expense"))

// Revert to 1.3 behavior (not retaining grouping column) by: sqlContext.setConf("spark.sql.retainGroupColumns", "false")

{% endhighlight %}

{% highlight java %}

// In 1.3.x, in order for the grouping column "department" to show up, // it must be included explicitly as part of the agg function call. df.groupBy("department").agg(col("department"), max("age"), sum("expense"));

// In 1.4+, grouping column "department" is included automatically. df.groupBy("department").agg(max("age"), sum("expense"));

// Revert to 1.3 behavior (not retaining grouping column) by: sqlContext.setConf("spark.sql.retainGroupColumns", "false");

{% endhighlight %}

{% highlight python %}

import pyspark.sql.functions as func

In 1.3.x, in order for the grouping column "department" to show up,

it must be included explicitly as part of the agg function call.

df.groupBy("department").agg(df["department"], func.max("age"), func.sum("expense"))

In 1.4+, grouping column "department" is included automatically.

df.groupBy("department").agg(func.max("age"), func.sum("expense"))

Revert to 1.3.x behavior (not retaining grouping column) by:

sqlContext.setConf("spark.sql.retainGroupColumns", "false")

{% endhighlight %}

Behavior change on DataFrame.withColumn

{:.no_toc}

Prior to 1.4, DataFrame.withColumn() supports adding a column only. The column will always be added as a new column with its specified name in the result DataFrame even if there may be any existing columns of the same name. Since 1.4, DataFrame.withColumn() supports adding a column of a different name from names of all existing columns or replacing existing columns of the same name.

Note that this change is only for Scala API, not for PySpark and SparkR.

Upgrading from Spark SQL 1.0-1.2 to 1.3

In Spark 1.3 we removed the "Alpha" label from Spark SQL and as part of this did a cleanup of the available APIs. From Spark 1.3 onwards, Spark SQL will provide binary compatibility with other releases in the 1.X series. This compatibility guarantee excludes APIs that are explicitly marked as unstable (i.e., DeveloperAPI or Experimental).

Rename of SchemaRDD to DataFrame

{:.no_toc}

The largest change that users will notice when upgrading to Spark SQL 1.3 is that SchemaRDD has been renamed to DataFrame. This is primarily because DataFrames no longer inherit from RDD directly, but instead provide most of the functionality that RDDs provide though their own implementation. DataFrames can still be converted to RDDs by calling the .rdd method.

In Scala, there is a type alias from SchemaRDD to DataFrame to provide source compatibility for some use cases. It is still recommended that users update their code to use DataFrame instead. Java and Python users will need to update their code.

Unification of the Java and Scala APIs

{:.no_toc}

Prior to Spark 1.3 there were separate Java compatible classes (JavaSQLContext and JavaSchemaRDD) that mirrored the Scala API. In Spark 1.3 the Java API and Scala API have been unified. Users of either language should use SQLContext and DataFrame. In general these classes try to use types that are usable from both languages (i.e. Array instead of language-specific collections). In some cases where no common type exists (e.g., for passing in closures or Maps) function overloading is used instead.

Additionally, the Java specific types API has been removed. Users of both Scala and Java should use the classes present in org.apache.spark.sql.types to describe schema programmatically.

Isolation of Implicit Conversions and Removal of dsl Package (Scala-only)

{:.no_toc}

Many of the code examples prior to Spark 1.3 started with import sqlContext._, which brought all of the functions from sqlContext into scope. In Spark 1.3 we have isolated the implicit conversions for converting RDDs into DataFrames into an object inside of the SQLContext. Users should now write import sqlContext.implicits._.

Additionally, the implicit conversions now only augment RDDs that are composed of Products (i.e., case classes or tuples) with a method toDF, instead of applying automatically.

When using function inside of the DSL (now replaced with the DataFrame API) users used to import org.apache.spark.sql.catalyst.dsl. Instead the public dataframe functions API should be used: import org.apache.spark.sql.functions._.

Removal of the type aliases in org.apache.spark.sql for DataType (Scala-only)

{:.no_toc}

Spark 1.3 removes the type aliases that were present in the base sql package for DataType. Users should instead import the classes in org.apache.spark.sql.types

UDF Registration Moved to sqlContext.udf (Java & Scala)

{:.no_toc}

Functions that are used to register UDFs, either for use in the DataFrame DSL or SQL, have been moved into the udf object in SQLContext.

{% highlight scala %}

sqlContext.udf.register("strLen", (s: String) => s.length())

{% endhighlight %}

{% highlight java %}

sqlContext.udf().register("strLen", (String s) -> s.length(), DataTypes.IntegerType);

{% endhighlight %}

Python UDF registration is unchanged.

Compatibility with Apache Hive

Spark SQL is designed to be compatible with the Hive Metastore, SerDes and UDFs. Currently, Hive SerDes and UDFs are based on Hive 1.2.1, and Spark SQL can be connected to different versions of Hive Metastore (from 0.12.0 to 2.3.6 and 3.0.0 to 3.1.2. Also see Interacting with Different Versions of Hive Metastore).

Deploying in Existing Hive Warehouses

{:.no_toc}

The Spark SQL Thrift JDBC server is designed to be "out of the box" compatible with existing Hive installations. You do not need to modify your existing Hive Metastore or change the data placement or partitioning of your tables.

Supported Hive Features

{:.no_toc}

Spark SQL supports the vast majority of Hive features, such as:

  • Hive query statements, including:

    • SELECT
    • GROUP BY
    • ORDER BY
    • CLUSTER BY
    • SORT BY

  • All Hive operators, including:

    • Relational operators (=, <=>, ==, <>, <, >, >=, <=, etc)
    • Arithmetic operators (+, -, *, /, %, etc)
    • Logical operators (AND, &&, OR, ||, etc)
    • Complex type constructors
    • Mathematical functions (sign, ln, cos, etc)
    • String functions (instr, length, printf, etc)

  • User defined functions (UDF)

  • User defined aggregation functions (UDAF)

  • User defined serialization formats (SerDes)

  • Window functions

  • Joins

    • JOIN
    • {LEFT|RIGHT|FULL} OUTER JOIN
    • LEFT SEMI JOIN
    • CROSS JOIN

  • Unions

  • Sub-queries

    • SELECT col FROM ( SELECT a + b AS col from t1) t2

  • Sampling

  • Explain

  • Partitioned tables including dynamic partition insertion

  • View

    • If column aliases are not specified in view definition queries, both Spark and Hive will generate alias names, but in different ways. In order for Spark to be able to read views created by Hive, users should explicitly specify column aliases in view definition queries. As an example, Spark cannot read v1 created as below by Hive.

      CREATE VIEW v1 AS SELECT * FROM (SELECT c + 1 FROM (SELECT 1 c) t1) t2;

      Instead, you should create v1 as below with column aliases explicitly specified.

      CREATE VIEW v1 AS SELECT * FROM (SELECT c + 1 AS inc_c FROM (SELECT 1 c) t1) t2;

  • All Hive DDL Functions, including:

    • CREATE TABLE
    • CREATE TABLE AS SELECT
    • ALTER TABLE

  • Most Hive Data types, including:

    • TINYINT
    • SMALLINT
    • INT
    • BIGINT
    • BOOLEAN
    • FLOAT
    • DOUBLE
    • STRING
    • BINARY
    • TIMESTAMP
    • DATE
    • ARRAY<>
    • MAP<>
    • STRUCT<>

Unsupported Hive Functionality

{:.no_toc}

Below is a list of Hive features that we don't support yet. Most of these features are rarely used in Hive deployments.

Major Hive Features

  • Tables with buckets: bucket is the hash partitioning within a Hive table partition. Spark SQL doesn't support buckets yet.

Esoteric Hive Features

  • UNION type
  • Unique join
  • Column statistics collecting: Spark SQL does not piggyback scans to collect column statistics at the moment and only supports populating the sizeInBytes field of the hive metastore.

Hive Input/Output Formats

  • File format for CLI: For results showing back to the CLI, Spark SQL only supports TextOutputFormat.
  • Hadoop archive

Hive Optimizations

A handful of Hive optimizations are not yet included in Spark. Some of these (such as indexes) are less important due to Spark SQL's in-memory computational model. Others are slotted for future releases of Spark SQL.

  • Block-level bitmap indexes and virtual columns (used to build indexes)
  • Automatically determine the number of reducers for joins and groupbys: Currently, in Spark SQL, you need to control the degree of parallelism post-shuffle using "SET spark.sql.shuffle.partitions=[num_tasks];".
  • Meta-data only query: For queries that can be answered by using only metadata, Spark SQL still launches tasks to compute the result.
  • Skew data flag: Spark SQL does not follow the skew data flags in Hive.
  • STREAMTABLE hint in join: Spark SQL does not follow the STREAMTABLE hint.
  • Merge multiple small files for query results: if the result output contains multiple small files, Hive can optionally merge the small files into fewer large files to avoid overflowing the HDFS metadata. Spark SQL does not support that.

Hive UDF/UDTF/UDAF

Not all the APIs of the Hive UDF/UDTF/UDAF are supported by Spark SQL. Below are the unsupported APIs:

  • getRequiredJars and getRequiredFiles (UDF and GenericUDF) are functions to automatically include additional resources required by this UDF.
  • initialize(StructObjectInspector) in GenericUDTF is not supported yet. Spark SQL currently uses a deprecated interface initialize(ObjectInspector[]) only.
  • configure (GenericUDF, GenericUDTF, and GenericUDAFEvaluator) is a function to initialize functions with MapredContext, which is inapplicable to Spark.
  • close (GenericUDF and GenericUDAFEvaluator) is a function to release associated resources. Spark SQL does not call this function when tasks finish.
  • reset (GenericUDAFEvaluator) is a function to re-initialize aggregation for reusing the same aggregation. Spark SQL currently does not support the reuse of aggregation.
  • getWindowingEvaluator (GenericUDAFEvaluator) is a function to optimize aggregation by evaluating an aggregate over a fixed window.

Incompatible Hive UDF

{:.no_toc}

Below are the scenarios in which Hive and Spark generate different results:

  • SQRT(n) If n < 0, Hive returns null, Spark SQL returns NaN.
  • ACOS(n) If n < -1 or n > 1, Hive returns null, Spark SQL returns NaN.
  • ASIN(n) If n < -1 or n > 1, Hive returns null, Spark SQL returns NaN.