- In PySpark, 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 `SparkSession`s, so we should not update them. Since 3.0, the builder comes to not update the configurations. This is the same behavior as Java/Scala API in 2.3 and above. If you want to update them, you need to update them prior to creating a `SparkSession`.
- 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)`.
- In Spark version 2.4 and earlier, the `from_json` function produces `null`s for JSON strings and JSON datasource skips the same independetly of its mode if there is no valid root JSON token in its input (` ` for example). Since Spark 3.0, such input is treated as a bad record and handled according to specified mode. For example, in the `PERMISSIVE` mode the ` ` input is converted to `Row(null, null)` if specified schema is `key STRING, value INT`.
- 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 users can still distinguish them via `Dataset.show`, `Dataset.collect` etc. Since Spark 3.0, float/double -0.0 is replaced by 0.0 internally, and users can't distinguish them any more.
- 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 udefined.
- 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.execution.setCommandRejectsSparkConfs` to `false`.
- Spark applications which are built with Spark version 2.4 and prior, and call methods of `UserDefinedFunction`, need to be re-compiled with Spark 3.0, as they are not binary compatible with Spark 3.0.
- 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.
<b>AnalysisException is thrown since integer type can not be promoted to string type in a loss-less manner.</b>
</th>
<th>
<b>Users can use explicit cast</b>
</th>
</tr>
</table>
- 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 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.
- In PySpark, when Arrow optimization is enabled, previously `toPandas` just failed when Arrow optimization is unable to be used whereas `createDataFrame` from Pandas DataFrame allowed the fallback to non-optimization. Now, both `toPandas` and `createDataFrame` from Pandas DataFrame allow the fallback by default, which can be switched off by `spark.sql.execution.arrow.fallback.enabled`.
- 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`.
## Upgrading From Spark SQL 2.3.0 to 2.3.1 and above
- As of version 2.3.1 Arrow functionality, including `pandas_udf` and `toPandas()`/`createDataFrame()` with `spark.sql.execution.arrow.enabled` set to `True`, has been marked as experimental. These are still evolving and not currently recommended for use in production.
## 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:
<tableclass="table">
<tr>
<th>
<b>InputA \ InputB</b>
</th>
<th>
<b>NullType</b>
</th>
<th>
<b>IntegerType</b>
</th>
<th>
<b>LongType</b>
</th>
<th>
<b>DecimalType(38,0)*</b>
</th>
<th>
<b>DoubleType</b>
</th>
<th>
<b>DateType</b>
</th>
<th>
<b>TimestampType</b>
</th>
<th>
<b>StringType</b>
</th>
</tr>
<tr>
<td>
<b>NullType</b>
</td>
<td>NullType</td>
<td>IntegerType</td>
<td>LongType</td>
<td>DecimalType(38,0)</td>
<td>DoubleType</td>
<td>DateType</td>
<td>TimestampType</td>
<td>StringType</td>
</tr>
<tr>
<td>
<b>IntegerType</b>
</td>
<td>IntegerType</td>
<td>IntegerType</td>
<td>LongType</td>
<td>DecimalType(38,0)</td>
<td>DoubleType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
</tr>
<tr>
<td>
<b>LongType</b>
</td>
<td>LongType</td>
<td>LongType</td>
<td>LongType</td>
<td>DecimalType(38,0)</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
</tr>
<tr>
<td>
<b>DecimalType(38,0)*</b>
</td>
<td>DecimalType(38,0)</td>
<td>DecimalType(38,0)</td>
<td>DecimalType(38,0)</td>
<td>DecimalType(38,0)</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
</tr>
<tr>
<td>
<b>DoubleType</b>
</td>
<td>DoubleType</td>
<td>DoubleType</td>
<td>StringType</td>
<td>StringType</td>
<td>DoubleType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
</tr>
<tr>
<td>
<b>DateType</b>
</td>
<td>DateType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>DateType</td>
<td>TimestampType</td>
<td>StringType</td>
</tr>
<tr>
<td>
<b>TimestampType</b>
</td>
<td>TimestampType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>TimestampType</td>
<td>TimestampType</td>
<td>StringType</td>
</tr>
<tr>
<td>
<b>StringType</b>
</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
<td>StringType</td>
</tr>
</table>
Note that, for <b>DecimalType(38,0)*</b>, 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.
- In PySpark, now we need Pandas 0.19.2 or upper if you want to use Pandas related functionalities, such as `toPandas`, `createDataFrame` from Pandas DataFrame, etc.
- In PySpark, the behavior of timestamp values for Pandas related functionalities was changed to respect session timezone. If you want to use the old behavior, you need to set a configuration `spark.sql.execution.pandas.respectSessionTimeZone` to `False`. See [SPARK-22395](https://issues.apache.org/jira/browse/SPARK-22395) for details.
- In PySpark, `na.fill()` or `fillna` also accepts boolean and replaces nulls with booleans. In prior Spark versions, PySpark just ignores it and returns the original Dataset/DataFrame.
- 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](sql-performance-tuning.html#broadcast-hint-for-sql-queries) and [SPARK-22489](https://issues.apache.org/jira/browse/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`).
- 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.
- In PySpark, `df.replace` does not allow to omit `value` when `to_replace` is not a dictionary. Previously, `value` could be omitted in the other cases and had `None` by default, which is counterintuitive and error-prone.
- 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](https://issues.apache.org/jira/browse/SPARK-20690) and [SPARK-21335](https://issues.apache.org/jira/browse/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 `SparkSession`s, 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.
- 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.
- 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.
`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`