compatibility.md

layout	title	nav_order
page	Compatibility	5

RAPIDS Accelerator for Apache Spark Compatibility with Apache Spark

The SQL plugin tries to produce results that are bit for bit identical with Apache Spark. There are a number of cases where there are some differences. In most cases operators that produce different results are off by default, and you can look at the configs for more information on how to enable them. In some cases we felt that enabling the incompatibility by default was worth the performance gain. All of those operators can be disabled through configs if it becomes a problem. Please also look at the current list of bugs which are typically incompatibilities that we have not yet addressed.

Ordering of Output

There are some operators where Spark does not guarantee the order of the output. These are typically things like aggregates and joins that may use a hash to distribute the work load among downstream tasks. In these cases the plugin does not guarantee that it will produce the same output order that Spark does. In cases such as an order by operation where the ordering is explicit the plugin will produce an ordering that is compatible with Spark's guarantee. It may not be 100% identical if the ordering is ambiguous.

In versions of Spark prior to 3.1.0 -0.0 is always < 0.0 but in 3.1.0 and above this is not true for sorting. For all versions of the plugin -0.0 == 0.0 for sorting.

Spark's sorting is typically a stable sort. Sort stability cannot be guaranteed in distributed work loads because the order in which upstream data arrives to a task is not guaranteed. Sort stability is only guaranteed in one situation which is reading and sorting data from a file using a single task/partition. The RAPIDS Accelerator does an unstable out of core sort by default. This simply means that the sort algorithm allows for spilling parts of the data if it is larger than can fit in the GPU's memory, but it does not guarantee ordering of rows when the ordering of the keys is ambiguous. If you do rely on a stable sort in your processing you can request this by setting spark.rapids.sql.stableSort.enabled to true and RAPIDS will try to sort all the data for a given task/partition at once on the GPU. This may change in the future to allow for a spillable stable sort.

Floating Point

For most basic floating-point operations like addition, subtraction, multiplication, and division the plugin will produce a bit for bit identical result as Spark does. For other functions like sin, cos, etc. the output may be different, but within the rounding error inherent in floating-point calculations. The ordering of operations to calculate the value may differ between the underlying JVM implementation used by the CPU and the C++ standard library implementation used by the GPU.

In the case of round and bround the results can be off by more because they can enlarge the difference. This happens in cases where a binary floating-point representation cannot exactly capture a decimal value. For example 1.025 cannot exactly be represented and ends up being closer to 1.02499. The Spark implementation of round converts it first to a decimal value with complex logic to make it 1.025 and then does the rounding. This results in round(1.025, 2) under pure Spark getting a value of 1.03 but under the RAPIDS accelerator it produces 1.02. As a side note Python will produce 1.02, Java does not have the ability to do a round like this built in, but if you do the simple operation of Math.round(1.025 * 100.0)/100.0 you also get 1.02.

For aggregations the underlying implementation is doing the aggregations in parallel and due to race conditions within the computation itself the result may not be the same each time the query is run. This is inherent in how the plugin speeds up the calculations and cannot be "fixed." If a query joins on a floating point value, which is not wise to do anyways, and the value is the result of a floating point aggregation then the join may fail to work properly with the plugin but would have worked with plain Spark. As of 22.06 this is behavior is enabled by default but can be disabled with the config spark.rapids.sql.variableFloatAgg.enabled.

Additionally, some aggregations on floating point columns that contain NaN can produce results different from Spark in versions prior to Spark 3.1.0. If it is known with certainty that the floating point columns do not contain NaN, set spark.rapids.sql.hasNans to false to run GPU enabled aggregations on them.

In the case of a distinct count on NaN values, prior to Spark 3.1.0, the issue only shows up if you have different NaN values. There are several different binary values that are all considered to be NaN by floating point. The plugin treats all of these as the same value, where as Spark treats them all as different values. Because this is considered to be rare we do not disable distinct count for floating point values even if spark.rapids.sql.hasNans is true.

0.0 vs -0.0

Floating point allows zero to be encoded as 0.0 and -0.0, but the IEEE standard says that they should be interpreted as the same. Most databases normalize these values to always be 0.0. Spark does this in some cases but not all as is documented here. The underlying implementation of this plugin treats them as the same for essentially all processing. This can result in some differences with Spark for operations, prior to Spark 3.1.0, like sorting, and distinct count. There are still differences with joins, and comparisons even after Spark 3.1.0.

We do not disable operations that produce different results due to -0.0 in the data because it is considered to be a rare occurrence.

Decimal Support

Apache Spark supports decimal values with a precision up to 38. This equates to 128-bits. However, when actually processing the data, in most cases, it is temporarily converted to Java's BigDecimal type which allows for effectively unlimited precision. This lets Spark do complicated calculations without the risk of missing an overflow and causing data corruption. It also lets Spark support some operations that require intermediate values that are larger than a 128-bit representation can support.

The RAPIDS Accelerator currently is limited to a maximum of 128-bits for storing or processing decimal values. This allows us to fully support the majority of decimal operations. But there are a few operations that we cannot support to the same degree as Spark can on the CPU.

Decimal Sum Aggregation

A number of fixes for overflow detection went into Spark 3.1.0. Please see SPARK-28067 and SPARK-32018 for more detailed information. Some of these fixes we were able to back port, but some of them require Spark 3.1.0 or above to fully be able to detect overflow in all cases. As such on versions of Spark older than 3.1.0 for large decimal values there is the possibility of data corruption in some corner cases. This is true for both the CPU and GPU implementations, but there are fewer of these cases for the GPU. If this concerns you, you should upgrade to Spark 3.1.0 or above.

When Apache Spark does a sum aggregation on decimal values it will store the result in a value with a precision that is the input precision + 10, but with a maximum precision of 38. For an input precision of 9 and above, Spark will do the aggregations as a Java BigDecimal value which is slow, but guarantees that any overflow can be detected because it can work with effectively unlimited precision. For inputs with a precision of 8 or below Spark will internally do the calculations as a long value, 64-bits. When the precision is 8, you would need at least 174,467,442,482 values/rows contributing to a single aggregation result before the overflow is no longer detected. Even then all the values would need to be either the largest or the smallest value possible to be stored in the type for the overflow to cause data corruption.

For the RAPIDS Accelerator we don't have direct access to unlimited precision for our calculations like the CPU does. For input values with a precision of 8 and below we follow Spark and process the data the same way, as a 64-bit value. For larger values we will do extra calculations looking at the higher order digits to be able to detect overflow in all cases. But because of this you may see some performance differences depending on the input precision used. The differences will show up when going from an input precision of 8 to 9 and again when going from an input precision of 28 to 29.

Decimal Average

Average is effectively doing a sum(input)/count(input), except the scale of the output type is the scale of the input + 4. As such it inherits some of the same issues that both sum and divide have. It also inherits some issues from Spark itself. See https://issues.apache.org/jira/browse/SPARK-37024 for a detailed description of some issues with average in Spark.

In order to be able to guarantee doing the divide with half up rounding at the end we only support average on input values with a precision of 23 or below. This is 38 - 10 for the sum guarantees and then 5 less to be able to shift the left-hand side of the divide enough to get a correct answer that can be rounded to the result that Spark would produce.

Divide and Multiply

Division and multiplication of decimal types is a little complicated in Apache Spark. For some arbitrary reason divide and multiply in Spark require that the precision and scale of the left-hand side and the right-hand side match. As such when planning a divide or multiply Spark will look at the original inputs to calculate the output precision and scale. Then it will cast the inputs to a common wider value where the scale is the max of the two input scales, and the precision is max of the two input non-scale portions (precision - scale) + the new scale. Then it will do the divide or multiply as a BigDecimal value, and return the result as a BigDecimal but lie about the precision and scale of the return type. Finally, Spark will insert a CheckOverflow expression that will round the scale of the BigDecimal value to that of the desired output type and check that the final precision will fit in the precision of the desired output type.

In order to match exactly with what Spark is doing the RAPIDS Accelerator would need at least 256-bit decimal values. We might implement that at some point, but until then we try to cover as much of division and multiplication as possible.

To combat this we look at the query plan and try to determine what is the smallest precision and scale for each parameter that would let us still produce the exact same answer as Apache Spark. We effectively try to undo what Spark did when widening the types to make them common.

Division

In Spark the output of a division operation is

val precision = p1 - s1 + s2 + max(6, s1 + p2 + 1)
val scale = max(6, s1 + p2 + 1)

Where p1 and s1 are the precision and scale of the left-hand side of the operation and p2 and s2 are the precision and scale of the right-hand side of the operation. But decimal divide inherently produces a result where the output scale is s1 - s2. In addition to this Spark will round the result to the given scale, and not just truncate it. This means that to produce the same result as Apache Spark we have to increase the scale of the left-hand side operation to be at least output_scale + s2 + 1. The + 1 is so the output is large enough that we can round it to the desired result. If this causes the precision of the left-hand side to go above 38, the maximum precision that 128-bits can hold, then we have to fall back to the CPU. Unfortunately the math is a bit complicated so there is no simple rule of thumb for this.

Multiplication

In Spark the output of a multiplication operation is

val precision = p1 + p2 + 1
val scale = s1 + s2

Where p1 and s1 are the precision and scale of the left-hand side of the operation and p2 and s2 are the precision and scale of the right-hand side of the operation. Fortunately, decimal multiply inherently produces the same scale, but Spark will round the result. As such, the RAPIDS Accelerator must add an extra decimal place to the scale and the precision, so we can round correctly. This means that if p1 + p2 > 36 we will fall back to the CPU to do processing.

How to get more decimal operations on the GPU?

Spark is very conservative in calculating the output types for decimal operations. It does this to avoid overflow in the worst case scenario, but generally will end up using a much larger type than is needed to store the final result. This means that over the course of a large query the precision and scale can grow to a size that would force the RAPIDS Accelerator to fall back to the CPU out of an abundance of caution. If you find yourself in this situation you can often cast the results to something smaller and still get the same answer. These casts should be done with some knowledge about the data being processed.

For example if we had a query like

SELECT SUM(cs_wholesale_cost * cs_quantity)/
       SUM(cs_sales_price * cs_quantity) cost_to_sale
  FROM catalog_sales
  GROUP BY cs_sold_date_sk
  ORDER BY cs_sold_date_sk

where cs_wholesale_cost and cs_sale_price are both decimal values with a precision of 7 and a scale of 2, Decimal(7, 2), and cs_quantity is a 32-bit integer. Only the first half of the query will be on the GPU. The following explanation is a bit complicated but tries to break down the processing into the distinct steps that Spark takes.

1. Multiplying a Decimal(7, 2) by an integer produces a Decimal(18, 2) value. This is the same for both multiply operations in the query.
2. The sum operation on the resulting Decimal(18, 2) column produces a Decimal(28, 2). This also is the same for both sum aggregations in the query.
3. The final divide operation is dividing a Decimal(28, 2) by another Decimal(28, 2) and produces a Decimal(38, 10).

We cannot guarantee that on the GPU the divide will produce the exact same result as the CPU for all possible inputs. But we know that we have at most 1,000,000 line items for each cs_sold_date_sk, and the average price/cost is no where close to the maximum value that Decimal(7, 2) can hold. So we can cast the result of the sums to a more reasonable Decimal(14, 2) and still produce an equivalent result, but totally on the GPU.

SELECT CAST(SUM(cs_wholesale_cost * cs_quantity) AS Decimal(14,2))/
       CAST(SUM(cs_sales_price * cs_quantity) AS Decimal(14,2)) cost_to_sale
  FROM catalog_sales
  GROUP BY cs_sold_date_sk
  ORDER BY cs_sold_date_sk

This should be done with some caution as it does reduce the range of values that the query could process before overflowing. It also can produce different result types. In this case instead of producing a Decimal(38, 10) the result is a Decimal(31, 17). If you really want the exact same result type you can cast the result back to a Decimal(38, 10), and the result will be identical to before. But, it can have a positive impact to performance.

If you have made it this far in the documentation then you probably know what you are doing and will use the following power only for good. It can often be difficult to determine if adding casts to put some processing on the GPU would improve performance or not. It can also be difficult to detect if a query might produce incorrect results because of a cast. To help answer some of these questions we provide spark.rapids.sql.decimalOverflowGuarantees that if set to false will disable guarantees for overflow checking and run all decimal operations on the GPU, even if it cannot guarantee that it will produce the exact same result as Spark. This should never be set to false in production because it disables all guarantees, and if your data does overflow, it might produce either a null value or worse an incorrect decimal value. But, it should give you more information about what the performance impact might be if you tuned it with casting. If you compare the results to GPU results with the guarantees still in place it should give you an idea if casting would still produce a correct answer. Even with this you should go through the query and your data and see what level of guarantees for outputs you are comfortable with.

Unicode

Spark delegates Unicode operations to the underlying JVM. Each version of Java complies with a specific version of the Unicode standard. The SQL plugin does not use the JVM for Unicode support and is compatible with Unicode version 12.1. Because of this there may be corner cases where Spark will produce a different result compared to the plugin.

CSV Reading

Spark allows for stripping leading and trailing white space using various options that are off by default. The plugin will strip leading and trailing space for all values except strings.

There are also discrepancies/issues with specific types that are detailed below.

CSV Strings

Writing strings to a CSV file in general for Spark can be problematic unless you can ensure that your data does not have any line deliminators in it. The GPU accelerated CSV parser handles quoted line deliminators similar to multiLine mode. But there are still a number of issues surrounding it and they should be avoided.

Escaped quote characters '\"' are not supported well as described by this issue.

CSV Dates

Only a limited set of formats are supported when parsing dates.

• "yyyy-MM-dd"
• "yyyy/MM/dd"
• "yyyy-MM"
• "yyyy/MM"
• "MM-yyyy"
• "MM/yyyy"
• "MM-dd-yyyy"
• "MM/dd/yyyy"
• "dd-MM-yyyy"
• "dd/MM/yyyy"

CSV Timestamps

The CSV parser does not support time zones. It will ignore any trailing time zone information, despite the format asking for a XXX or [XXX]. As such it is off by default and you can enable it by setting spark.rapids.sql.csvTimestamps.enabled to true.

The formats supported for timestamps are limited similar to dates. The first part of the format must be a supported date format. The second part must start with a 'T' to separate the time portion followed by one of the following formats:

• HH:mm:ss.SSSXXX
• HH:mm:ss[.SSS][XXX]
• HH:mm:ss[.SSSXXX]
• HH:mm
• HH:mm:ss
• HH:mm[:ss]
• HH:mm:ss.SSS
• HH:mm:ss[.SSS]

Just like with dates all timestamp formats are actually supported at the same time. The plugin will disable itself if it sees a format it cannot support.

Invalid timestamps in Spark, ones that have the correct format, but the numbers produce invalid dates or times, can result in an exception by default and how they are parsed can be controlled through a config. The RAPIDS Accelerator does not support any of this and will produce an incorrect date. Typically, one that overflowed.

CSV Floating Point

Parsing floating-point values has the same limitations as casting from string to float.

Also parsing of some values will not produce bit for bit identical results to what the CPU does. They are within round-off errors except when they are close enough to overflow to Inf or -Inf which then results in a number being returned when the CPU would have returned null.

CSV ANSI day time interval

This type was added in as a part of Spark 3.3.0, and it's not supported on Spark versions before 3.3.0. Apache Spark can overflow when reading ANSI day time interval values. The RAPIDS Accelerator does not overflow and as such is not bug for bug compatible with Spark in this case.

Interval string in csv	Spark reads to	The RAPIDS Accelerator reads to	Comments
interval '106751992' day	INTERVAL '-106751990' DAY	NULL	Spark issue
interval '2562047789' hour	INTERVAL '-2562047787' HOUR	NULL	Spark issue

There are two valid textual representations in CSV: the ANSI style and the HIVE style, e.g:

SQL Type	An instance of ANSI style	An instance of HIVE style
INTERVAL DAY	INTERVAL '100' DAY TO SECOND	100
INTERVAL DAY TO HOUR	INTERVAL '100 10' DAY TO HOUR	100 10
INTERVAL DAY TO MINUTE	INTERVAL '100 10:30' DAY TO MINUTE	100 10:30
INTERVAL DAY TO SECOND	INTERVAL '100 10:30:40.999999' DAY TO SECOND	100 10:30:40.999999
INTERVAL HOUR	INTERVAL '10' HOUR	10
INTERVAL HOUR TO MINUTE	INTERVAL '10:30' HOUR TO MINUTE	10:30
INTERVAL HOUR TO SECOND	INTERVAL '10:30:40.999999' HOUR TO SECOND	10:30:40.999999
INTERVAL MINUTE	INTERVAL '30' MINUTE	30
INTERVAL MINUTE TO SECOND	INTERVAL '30:40.999999' MINUTE TO SECOND	30:40.999999
INTERVAL SECOND	INTERVAL '40.999999' SECOND	40.999999
Currently, the RAPIDS Accelerator only supports the ANSI style.

ORC

The ORC format has fairly complete support for both reads and writes. There are only a few known issues. The first is for reading timestamps and dates around the transition between Julian and Gregorian calendars as described here. A similar issue exists for writing dates as described here. Writing timestamps, however only appears to work for dates after the epoch as described here.

The plugin supports reading uncompressed, snappy and zlib ORC files and writing uncompressed and snappy ORC files. At this point, the plugin does not have the ability to fall back to the CPU when reading an unsupported compression format, and will error out in that case.

Push Down Aggregates for ORC

Spark-3.3.0+ pushes down certain aggregations (MIN/MAX/COUNT) into ORC when the user-config spark.sql.orc.aggregatePushdown is set to true.
By enabling this feature, aggregate query performance will improve as it takes advantage of the statistics information.

Caution

Spark ORC reader/writer assumes that all ORC files must have valid column statistics. This assumption deviates from the ORC-specification which states that statistics are optional.
When a Spark-3.3.0+ job reads an ORC file with empty file-statistics, it fails while throwing the following runtime exception:

org.apache.spark.SparkException: Cannot read columns statistics in file: /PATH_TO_ORC_FILE
E    Caused by: java.util.NoSuchElementException
E        at java.util.LinkedList.removeFirst(LinkedList.java:270)
E        at java.util.LinkedList.remove(LinkedList.java:685)
E        at org.apache.spark.sql.execution.datasources.orc.OrcFooterReader.convertStatistics(OrcFooterReader.java:54)
E        at org.apache.spark.sql.execution.datasources.orc.OrcFooterReader.readStatistics(OrcFooterReader.java:45)
E        at org.apache.spark.sql.execution.datasources.orc.OrcUtils$.createAggInternalRowFromFooter(OrcUtils.scala:428)

The Spark community is planning to work on a runtime fallback to read from actual rows when ORC file-statistics are missing (see SPARK-34960 discussion).

Limitations With RAPIDS

RAPIDS does not support whole file statistics in ORC file in releases prior to release 22.06.

Writing ORC Files

If you are using release prior to release 22.06 where CUDF does not support writing file statistics, then the ORC files written by the GPU are incompatible with the optimization causing an ORC read-job to fail as described above.
In order to prevent job failures in releases prior to release 22.06, spark.sql.orc.aggregatePushdown should be disabled while reading ORC files that were written by the GPU.

Reading ORC Files

To take advantage of the aggregate optimization, the plugin falls back to the CPU as it is a meta data only query. As long as the ORC file has valid statistics (written by the CPU), then the pushing down aggregates to the ORC layer should be successful.
Otherwise, reading an ORC file written by the GPU requires aggregatePushdown to be disabled.

Parquet

The Parquet format has more configs because there are multiple versions with some compatibility issues between them. Dates and timestamps are where the known issues exist. For reads when spark.sql.legacy.parquet.datetimeRebaseModeInWrite is set to CORRECTED timestamps before the transition between the Julian and Gregorian calendars are wrong, but dates are fine. When spark.sql.legacy.parquet.datetimeRebaseModeInWrite is set to LEGACY, the read may fail for values occurring before the transition between the Julian and Gregorian calendars, i.e.: date <= 1582-10-04.

When writing spark.sql.legacy.parquet.datetimeRebaseModeInWrite is currently ignored as described here.

When spark.sql.parquet.outputTimestampType is set to INT96, the timestamps will overflow and result in an IllegalArgumentException thrown, if any value is before September 21, 1677 12:12:43 AM or it is after April 11, 2262 11:47:17 PM. To get around this issue, turn off the ParquetWriter acceleration for timestamp columns by either setting spark.rapids.sql.format.parquet.writer.int96.enabled to false or set spark.sql.parquet.outputTimestampType to TIMESTAMP_MICROS or TIMESTAMP_MILLIS to by -pass the issue entirely.

The plugin supports reading uncompressed, snappy and gzip Parquet files and writing uncompressed and snappy Parquet files. At this point, the plugin does not have the ability to fall back to the CPU when reading an unsupported compression format, and will error out in that case.

JSON

The JSON format read is a very experimental feature which is expected to have some issues, so we disable it by default. If you would like to test it, you need to enable spark.rapids.sql.format.json.enabled and spark.rapids.sql.format.json.read.enabled.

Currently, the GPU accelerated JSON reader doesn't support column pruning, which will likely make this difficult to use or even test. The user must specify the full schema or just let Spark infer the schema from the JSON file. eg,

We have a people.json file with below content

{"name":"Michael"}
{"name":"Andy", "age":30}
{"name":"Justin", "age":19}

Both below ways will work

• Inferring the schema

  val df = spark.read.json("people.json")

• Specifying the full schema

  val schema = StructType(Seq(StructField("name", StringType), StructField("age", IntegerType)))
  val df = spark.read.schema(schema).json("people.json")

While the below code will not work in the current version,

val schema = StructType(Seq(StructField("name", StringType)))
val df = spark.read.schema(schema).json("people.json")

JSON supporting types

The nested types(array, map and struct) are not supported yet in current version.

JSON Floating Point

Parsing floating-point values has the same limitations as casting from string to float.

Prior to Spark 3.3.0, reading JSON strings such as "+Infinity" when specifying that the data type is FloatType or DoubleType caused these values to be parsed even when allowNonNumericNumbers is set to false. Also, Spark versions prior to 3.3.0 only supported the "Infinity" and "-Infinity" representations of infinity and did not support "+INF", "-INF", or "+Infinity", which Spark considers valid when unquoted. The GPU JSON reader is consistent with the behavior in Spark 3.3.0 and later.

Another limitation of the GPU JSON reader is that it will parse strings containing non-string boolean or numeric values where Spark will treat them as invalid inputs and will just return null.

JSON Timestamps

There is currently no support for reading numeric values as timestamps and null values are returned instead (#4940). A workaround would be to read as longs and then cast to timestamp.

JSON Schema discovery

Spark SQL can automatically infer the schema of a JSON dataset if schema is not provided explicitly. The CPU handles schema discovery and there is no GPU acceleration of this. By default Spark will read/parse the entire dataset to determine the schema. This means that some options/errors which are ignored by the GPU may still result in an exception if used with schema discovery.

JSON options

Spark supports passing options to the JSON parser when reading a dataset. In most cases if the RAPIDS Accelerator sees one of these options that it does not support it will fall back to the CPU. In some cases we do not. The following options are documented below.

• allowNumericLeadingZeros - Allows leading zeros in numbers (e.g. 00012). By default this is set to false. When it is false Spark throws an exception if it encounters this type of number. The RAPIDS Accelerator strips off leading zeros from all numbers and this config has no impact on it.

• allowUnquotedControlChars - Allows JSON Strings to contain unquoted control characters (ASCII characters with value less than 32, including tab and line feed characters) or not. By default this is set to false. If the schema is provided while reading JSON file, then this flag has no impact on the RAPIDS Accelerator as it always allows unquoted control characters but Spark reads these entries incorrectly as null. However, if the schema is not provided and when the option is false, then RAPIDS Accelerator's behavior is same as Spark where an exception is thrown as discussed in JSON Schema discovery section.

• allowNonNumericNumbers - Allows NaN and Infinity values to be parsed (note that these are not valid numeric values in the JSON specification). Spark versions prior to 3.3.0 have inconsistent behavior and will parse some variants of NaN and Infinity even when this option is disabled (SPARK-38060). The RAPIDS Accelerator behavior is consistent with Spark version 3.3.0 and later.

Avro

The Avro format read is a very experimental feature which is expected to have some issues, so we disable it by default. If you would like to test it, you need to enable spark.rapids.sql.format.avro.enabled and spark.rapids.sql.format.avro.read.enabled.

Currently, the GPU accelerated Avro reader doesn't support reading the Avro version 1.2 files.

Supported types

The boolean, byte, short, int, long, float, double, string are supported in current version.

Regular Expressions

The following Apache Spark regular expression functions and expressions are supported on the GPU:

• RLIKE
• regexp
• regexp_extract
• regexp_like
• regexp_replace
• string_split
• str_to_map

Regular expression evaluation on the GPU is enabled by default. Execution will fall back to the CPU for regular expressions that are not yet supported on the GPU. However, there are some edge cases that will still execute on the GPU and produce different results to the CPU. To disable regular expressions on the GPU, set spark.rapids.sql.regexp.enabled=false.

These are the known edge cases where running on the GPU will produce different results to the CPU:

• Using regular expressions with Unicode data can produce incorrect results if the system LANG is not set to en_US.UTF-8 (#5549)
• Regular expressions that contain an end of line anchor '$' or end of string anchor '\Z' or '\z' immediately next to a newline or a repetition that produces zero or more results (#5610)`

The following regular expression patterns are not yet supported on the GPU and will fall back to the CPU.

• Line anchor ^ is not supported in some contexts, such as when combined with a choice (^|a).
• Line anchor $ is not supported by regexp_replace, and in some rare contexts.
• String anchor \Z is not supported by regexp_replace, and in some rare contexts.
• Line anchor $ and string anchors \z and \Z are not supported in patterns containing \W or \D
• Line and string anchors are not supported by string_split and str_to_map
• Word and non-word boundaries, \b and \B
• Lazy quantifiers, such as a*?
• Possessive quantifiers, such as a*+
• Character classes that use union, intersection, or subtraction semantics, such as [a-d[m-p]], [a-z&&[def]], or [a-z&&[^bc]]
• Empty groups: ()
• Regular expressions containing null characters (unless the pattern is a simple literal string)
• regexp_replace does not support back-references

Work is ongoing to increase the range of regular expressions that can run on the GPU.

Timestamps

Spark stores timestamps internally relative to the JVM time zone. Converting an arbitrary timestamp between time zones is not currently supported on the GPU. Therefore operations involving timestamps will only be GPU-accelerated if the time zone used by the JVM is UTC.

Windowing

Window Functions

Because of ordering differences between the CPU and the GPU window functions especially row based window functions like row_number, lead, and lag can produce different results if the ordering includes both -0.0 and 0.0, or if the ordering is ambiguous. Spark can produce different results from one run to another if the ordering is ambiguous on a window function too.

Range Window

When the order-by column of a range based window is numeric type like byte/short/int/long and the range boundary calculated for a value has overflow, CPU and GPU will get different results.

For example, consider the following dataset:

+------+---------+
| id   | dollars |
+------+---------+
|    1 |    NULL |
|    1 |      13 |
|    1 |      14 |
|    1 |      15 |
|    1 |      15 |
|    1 |      17 |
|    1 |      18 |
|    1 |      52 |
|    1 |      53 |
|    1 |      61 |
|    1 |      65 |
|    1 |      72 |
|    1 |      73 |
|    1 |      75 |
|    1 |      78 |
|    1 |      84 |
|    1 |      85 |
|    1 |      86 |
|    1 |      92 |
|    1 |      98 |
+------+---------+

After executing the SQL statement:

SELECT
 COUNT(dollars) over
    (PARTITION BY id
    ORDER BY CAST (dollars AS Byte) ASC
    RANGE BETWEEN 127 PRECEDING AND 127 FOLLOWING)
FROM table

The results will differ between the CPU and GPU due to overflow handling.

CPU: WrappedArray([0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0], [0])
GPU: WrappedArray([0], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19], [19])

To enable byte-range windowing on the GPU, set spark.rapids.sql.window.range.byte.enabled to true.

We also provide configurations for other integral range types:

• spark.rapids.sql.window.range.short.enabled
• spark.rapids.sql.window.range.int.enabled
• spark.rapids.sql.window.range.long.enabled

The reason why we default the configurations to false for byte/short and to true for int/long is that we think the most real-world queries are based on int or long.

Parsing strings as dates or timestamps

When converting strings to dates or timestamps using functions like to_date and unix_timestamp, the specified format string will fall into one of three categories:

• Supported on GPU and 100% compatible with Spark
• Supported on GPU but may produce different results to Spark
• Unsupported on GPU

The formats which are supported on GPU vary depending on the setting for timeParserPolicy.

CORRECTED and EXCEPTION timeParserPolicy

With timeParserPolicy set to CORRECTED or EXCEPTION (the default), the following formats are supported on the GPU without requiring any additional settings.

• dd/MM/yyyy
• yyyy/MM
• yyyy/MM/dd
• yyyy-MM
• yyyy-MM-dd
• yyyy-MM-dd HH:mm:ss
• MM-dd
• MM/dd
• dd-MM
• dd/MM

Valid Spark date/time formats that do not appear in the list above may also be supported but have not been extensively tested and may produce different results compared to the CPU. Known issues include:

• Valid dates and timestamps followed by trailing characters (including whitespace) may be parsed to non-null values on GPU where Spark would treat the data as invalid and return null

To attempt to use other formats on the GPU, set spark.rapids.sql.incompatibleDateFormats.enabled to true.

Formats that contain any of the following characters are unsupported and will fall back to CPU:

'k', 'K','z', 'V', 'c', 'F', 'W', 'Q', 'q', 'G', 'A', 'n', 'N',
'O', 'X', 'p', '\'', '[', ']', '#', '{', '}', 'Z', 'w', 'e', 'E', 'x', 'Z', 'Y'

Formats that contain any of the following words are unsupported and will fall back to CPU:

"u", "uu", "uuu", "uuuu", "uuuuu", "uuuuuu", "uuuuuuu", "uuuuuuuu", "uuuuuuuuu", "uuuuuuuuuu",
"y", "yy", yyy", "yyyyy", "yyyyyy", "yyyyyyy", "yyyyyyyy", "yyyyyyyyy", "yyyyyyyyyy",
"D", "DD", "DDD", "s", "m", "H", "h", "M", "MMM", "MMMM", "MMMMM", "L", "LLL", "LLLL", "LLLLL",
"d", "S", "SS", "SSS", "SSSS", "SSSSS", "SSSSSSSSS", "SSSSSSS", "SSSSSSSS"

LEGACY timeParserPolicy

With timeParserPolicy set to LEGACY and spark.rapids.sql.incompatibleDateFormats.enabled set to true, and spark.sql.ansi.enabled set to false, the following formats are supported but not guaranteed to produce the same results as the CPU:

• dd-MM-yyyy
• dd/MM/yyyy
• yyyy/MM/dd
• yyyy-MM-dd
• yyyy/MM/dd HH:mm:ss
• yyyy-MM-dd HH:mm:ss

LEGACY timeParserPolicy support has the following limitations when running on the GPU:

• Only 4 digit years are supported
• The proleptic Gregorian calendar is used instead of the hybrid Julian+Gregorian calendar that Spark uses in legacy mode

Formatting dates and timestamps as strings

When formatting dates and timestamps as strings using functions such as from_unixtime, only a subset of valid format strings are supported on the GPU.

Formats that contain any of the following characters are unsupported and will fall back to CPU:

'k', 'K','z', 'V', 'c', 'F', 'W', 'Q', 'q', 'G', 'A', 'n', 'N',
'O', 'X', 'p', '\'', '[', ']', '#', '{', '}', 'Z', 'w', 'e', 'E', 'x', 'Z', 'Y'

Formats that contain any of the following words are unsupported and will fall back to CPU:

"u", "uu", "uuu", "uuuu", "uuuuu", "uuuuuu", "uuuuuuu", "uuuuuuuu", "uuuuuuuuu", "uuuuuuuuuu",
"y", yyy", "yyyyy", "yyyyyy", "yyyyyyy", "yyyyyyyy", "yyyyyyyyy", "yyyyyyyyyy",
"D", "DD", "DDD", "s", "m", "H", "h", "M", "MMM", "MMMM", "MMMMM", "L", "LLL", "LLLL", "LLLLL",
"d", "S", "SS", "SSS", "SSSS", "SSSSS", "SSSSSSSSS", "SSSSSSS", "SSSSSSSS"

Note that this list differs very slightly from the list given in the previous section for parsing strings to dates because the two-digit year format "yy" is supported when formatting dates as strings but not when parsing strings to dates.

Casting between types

In general, performing cast and ansi_cast operations on the GPU is compatible with the same operations on the CPU. However, there are some exceptions. For this reason, certain casts are disabled on the GPU by default and require configuration options to be specified to enable them.

Float to Decimal

The GPU will use a different strategy from Java's BigDecimal to handle/store decimal values, which leads to restrictions:

• Float values cannot be larger than 1e18 or smaller than -1e18 after conversion.
• The results produced by GPU slightly differ from the default results of Spark.

As of 22.06 this conf is enabled, to disable this operation on the GPU when using Spark 3.1.0 or later, set spark.rapids.sql.castFloatToDecimal.enabled to false

Float to Integral Types

With both cast and ansi_cast, Spark uses the expression Math.floor(x) <= MAX && Math.ceil(x) >= MIN to determine whether a floating-point value can be converted to an integral type. Prior to Spark 3.1.0 the MIN and MAX values were floating-point values such as Int.MaxValue.toFloat but starting with 3.1.0 these are now integral types such as Int.MaxValue so this has slightly affected the valid range of values and now differs slightly from the behavior on GPU in some cases.

As of 22.06 this conf is enabled, to disable this operation on the GPU when using Spark 3.1.0 or later, set spark.rapids.sql.castFloatToIntegralTypes.enabled to false.

This configuration setting is ignored when using Spark versions prior to 3.1.0.

Float to String

The GPU will use different precision than Java's toString method when converting floating-point data types to strings. The GPU uses a lowercase e prefix for an exponent while Spark uses uppercase E. As a result the computed string can differ from the default behavior in Spark.

As of 22.06 this conf is enabled by default, to disable this operation on the GPU, set spark.rapids.sql.castFloatToString.enabled to false.

String to Float

Casting from string to floating-point types on the GPU returns incorrect results when the string represents any number in the following ranges. In both cases the GPU returns Double.MaxValue. The default behavior in Apache Spark is to return +Infinity and -Infinity, respectively.

• 1.7976931348623158E308 <= x < 1.7976931348623159E308
• -1.7976931348623159E308 < x <= -1.7976931348623158E308

Also, the GPU does not support casting from strings containing hex values.

As of 22.06 this conf is enabled by default, to enable this operation on the GPU, set spark.rapids.sql.castStringToFloat.enabled to false.

String to Date

The following formats/patterns are supported on the GPU. Timezone of UTC is assumed.

Format or Pattern	Supported on GPU?
"yyyy"	Yes
"yyyy-[M]M"	Yes
"yyyy-[M]M "	Yes
"yyyy-[M]M-[d]d"	Yes
"yyyy-[M]M-[d]d "	Yes
"yyyy-[M]M-[d]d *"	Yes
"yyyy-[M]M-[d]d T*"	Yes
"epoch"	Yes
"now"	Yes
"today"	Yes
"tomorrow"	Yes
"yesterday"	Yes

String to Timestamp

To allow casts from string to timestamp on the GPU, enable the configuration property spark.rapids.sql.castStringToTimestamp.enabled.

Casting from string to timestamp currently has the following limitations.

Format or Pattern	Supported on GPU?
"yyyy"	Yes
"yyyy-[M]M"	Yes
"yyyy-[M]M "	Yes
"yyyy-[M]M-[d]d"	Yes
"yyyy-[M]M-[d]d "	Yes
"yyyy-[M]M-[d]dT[h]h:[m]m:[s]s.[ms][ms][ms][us][us][us][zone_id]"	Partial [1]
"yyyy-[M]M-[d]d [h]h:[m]m:[s]s.[ms][ms][ms][us][us][us][zone_id]"	Partial [1]
"[h]h:[m]m:[s]s.[ms][ms][ms][us][us][us][zone_id]"	Partial [1]
"T[h]h:[m]m:[s]s.[ms][ms][ms][us][us][us][zone_id]"	Partial [1]
"epoch"	Yes
"now"	Yes
"today"	Yes
"tomorrow"	Yes
"yesterday"	Yes

• [1] The timestamp portion must have 6 digits for milliseconds. Only timezone 'Z' (UTC) is supported. Casting unsupported formats will result in null values.

Spark is very lenient when casting from string to timestamp because all date and time components are optional, meaning that input values such as T, T2, :, ::, 1:, :1, and ::1 are considered valid timestamps. The GPU will treat these values as invalid and cast them to null values.

Constant Folding

ConstantFolding is an operator optimization rule in Catalyst that replaces expressions that can be statically evaluated with their equivalent literal values. The RAPIDS Accelerator relies on constant folding and parts of the query will not be accelerated if org.apache.spark.sql.catalyst.optimizer.ConstantFolding is excluded as a rule.

long/double to Timestamp

Spark 330+ has an issue when casting a big enough long/double as timestamp, refer to https://issues.apache.org/jira/browse/SPARK-39209. Spark 330+ throws errors while the RAPIDS Accelerator can handle correctly when casting a big enough long/double as timestamp.

JSON string handling

The 0.5 release introduces the get_json_object operation. The JSON specification only allows double quotes around strings in JSON data, whereas Spark allows single quotes around strings in JSON data. The RAPIDS Spark get_json_object operation on the GPU will return None in PySpark or Null in Scala when trying to match a string surrounded by single quotes. This behavior will be updated in a future release to more closely match Spark.

Approximate Percentile

The GPU implementation of approximate_percentile uses t-Digests which have high accuracy, particularly near the tails of a distribution. Because the results are not bit-for-bit identical with the Apache Spark implementation of approximate_percentile, this feature is disabled by default and can be enabled by setting spark.rapids.sql.incompatibleOps.enabled=true.

Conditionals and operations with side effects (ANSI mode)

In Apache Spark condition operations like if, coalesce, and case/when lazily evaluate their parameters on a row by row basis. On the GPU it is generally more efficient to evaluate the parameters regardless of the condition and then select which result to return based on the condition. This is fine so long as there are no side effects caused by evaluating a parameter. For most expressions in Spark this is true, but in ANSI mode many expressions can throw exceptions, like for the Add expression if an overflow happens. This is also true of UDFs, because by their nature they are user defined and can have side effects like throwing exceptions.

Currently, the RAPIDS Accelerator assumes that there are no side effects. This can result it situations, specifically in ANSI mode, where the RAPIDS Accelerator will always throw an exception, but Spark on the CPU will not. For example:

spark.conf.set("spark.sql.ansi.enabled", "true")

Seq(0L, Long.MaxValue).toDF("val")
    .repartition(1) // The repartition makes Spark not optimize selectExpr away
    .selectExpr("IF(val > 1000, null, val + 1) as ret")
    .show()

If the above example is run on the CPU you will get a result like.

+----+
| ret|
+----+
|   1|
|null|
+----+

But if it is run on the GPU an overflow exception is thrown. As was explained before this is because the RAPIDS Accelerator will evaluate both val + 1 and null regardless of the result of the condition. In some cases you can work around this. The above example could be re-written so the if happens before the Add operation.

Seq(0L, Long.MaxValue).toDF("val")
    .repartition(1) // The repartition makes Spark not optimize selectExpr away
    .selectExpr("IF(val > 1000, null, val) + 1 as ret")
    .show()

But this is not something that can be done generically and requires inner knowledge about what can trigger a side effect.