Enabling Cross-Platform Query Optimization via Expressive Markup Language

Abstract

A database system receives a request from a user. The request invokes a data set function (DSF) and uses a property to be provided by the DSF. The database system determines that a function descriptor is available for the DSF. The function descriptor is expressed as markup language instructions. The function descriptor defines the property of the DSF. The database system uses the function descriptor to define a property for the DSF.

Description

BACKGROUND

Data processing systems may execute a workflow composed of systems and/or user defined functions that may execute on a single or multiple execution engines. The typical execution scenario, which is facilitated by the Foreign Function Execution (FFE) and TERADATA®'s QUERYGRID™ (TERADATA's parallel data transport framework) infrastructures) where the TERADATA and QUERYGRID trademarks are owned by Teradata US, Inc, involves the following steps: (1) Users submit their SQL queries (involving remote analytical function invocations) to a database system (such as, but not limited to, a TERADATA database system), (2) the database system sends a function invocation request along with the needed data to the remote system, (3) The remote system executes the function and generates the results, and (4) the results are sent back to the database system (within the context of the caller SQL query).

FIG. 1 illustrates such a query. As can be seen, the query includes a “SESSIONIZE” data set function (DSF), which is defined to be a function that takes one or more data sets (or relational tables) as input and generates one or more data sets (or relational tables) as output. DSFs can execute locally at the caller system or remotely on a coprocessor engine. A coprocessor engine (or coprocessor) is a second data processing engine executing any part of the query data processing workflow. A coprocessor may reside at the same integrated platform or at a separate platform on premises or at the cloud. A remote analytical function (“RAF”), where an RAF is defined to be a function that may be called by a database system but that is executed by a foreign system, is a special case of a DSF.

The challenge is that given such a query, e.g., the query invoking a DSF such as the SESSIONIZE function in the FIG. 1, the DSF may be a “blackbox” to the database system in that the database system does not know the properties of the invoked DSF. For example, the database system may not know (1) the output schema to be returned from the DSF (i.e., Q1 in FIG. 1), (2) whether all columns of the input table (e.g., web_click in the example shown in FIG. 1) are needed by the DSF or only small subset of columns (i.e., Q2 in FIG. 1), and (3) whether or not a post-function predicate (“user_id=o” in the example shown in FIG. 1) can be pushed inside the ON clause to be evaluated on the web_click table before executing the DSF (i.e., Q3 in FIG. 1).

Determining these properties could enable optimizations that yield significant performance gains. However, capturing these properties is not straightforward because the properties may be dynamic and some or all of them may only be resolved at query time, where “query time” is defined to be the time that the query is being executed by the database system, given a specific invocation.

FIGS. 2 and 3 illustrate an example. FIG. 2 shows the invocation format and input schema specifications for the Unpivot( ) DSF provided by a coprocessor (specifically, an Aster coprocessor). The output schema can be interpreted as shown in FIG. 3.

The wide diversity of the DSFs and the high complexity of their property specifications, as illustrated in FIGS. 2 and 3, show the challenge of using DSFs.

SUMMARY

In one aspect, a method includes a database system receiving a request from a user. The request invokes a data set function (DSF) and uses a property to be provided by the DSF. The database system determines that a function descriptor is available for the DSF. The function descriptor is expressed as markup language instructions. The function descriptor defines the property of the DSF. The database system uses the function descriptor to define a property for the DSF.

Implementations may include one or more of the following. The method may include a developer creating the DSF to execute on a remote system, the developer writing the descriptor for the DSF, and the database system receiving and storing the descriptor for the DSF. The method may include a developer creating the DSF to execute on the database system, the developer writing the descriptor for the DSF, and the database system receiving and storing the descriptor for the DSF. The markup language may be an instruction-based language. The method may include using the function descriptor to define an output schema for the DSF. The method may include using the function descriptor to define an input schema for the DSF. The method may include using the function descriptor to determine to push a predicate in the request from the DSF's output to the input of the DSF. The method may include using the function descriptor to determine to push a projection in the request from the input of the DSF to the output of the DSF. The method may include using the function descriptor to estimate a cardinality of the property. The method may include using the function descriptor to determine if the DSF inherits or obeys specific ordering or partitioning schemes.

In another aspect, a non-transitory computer-readable tangible medium records a computer program. The computer program includes executable instructions, that, when executed, perform a method. The method includes a database system receiving a request from a user. The request invokes a data set function (DSF) and uses a property to be provided by the DSF. The database system determines that a function descriptor is available for the DSF. The function descriptor is expressed as markup language instructions. The function descriptor defines the property of the DSF. The database system uses the function descriptor to define a property for the DSF.

In another aspect, a method includes a database system receiving a request from a user, wherein the request invokes a data set function (DSF) and uses a property to be provided by the DSF. The database system determines that a function descriptor is not available for the DSF. The database system determines that a contract function is available for the DSF. The database system using the contract function to optimize the request.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is an illustration of a request that invokes a DSF.

FIG. 2 shows the invocation format and output schema specifications for the Unpivot( ) DSF provided by a coprocessor.

FIG. 3 illustrates a description of the output schema for the Unpivot( ) DSF of FIG. 2.

FIG. 4 illustrates the lifecycle of a newly introduced function descriptor from creation to enabling optimizations.

FIGS. 5A and 5B illustrate the co-existence of the function descriptor mechanism and the contract function mechanism.

FIG. 6 shows the structure of the ADD instruction.

FIG. 7 shows the structure of the CASE instruction.

FIG. 8 shows the specifications of a Pivot DSF invocation.

FIG. 9 shows the output schema specifications for the Pivot DSF.

FIG. 10 shows an example Pivot DSF invocation and output schema.

FIG. 11 shows the specifications of a PCAPlot DSF.

FIG. 12 shows the output schema for the PCAPlot DSF.

FIG. 13 shows an example of the invocation of PCAPlot DSF and the resulting output schema.

FIG. 14 shows the specifications of a NaiveBayesPredict DSF.

FIG. 15 shows the output schema for the NaiveBayesPredict DSF.

FIG. 16 shows an example of the invocation of NaiveBayesPredict DSF and the resulting output schema.

DETAILED DESCRIPTION

The following detailed description illustrates embodiments of the present disclosure. These embodiments are described in sufficient detail to enable a person of ordinary skill in the art to practice these embodiments without undue experimentation. It should be understood, however, that the embodiments and examples described herein are given by way of illustration only, and not by way of limitation. Various substitutions, modifications, additions, and rearrangements may be made that remain potential applications of the disclosed techniques. Therefore, the description that follows is not to be taken as limiting on the scope of the appended claims. In particular, an element associated with a particular embodiment should not be limited to association with that particular embodiment but should be assumed to be capable of association with any embodiment discussed herein.

A technique for addressing the complexity of using DSFs captures properties of the analytical functions that can enable efficient compilation and execution of the function from within the database system. The technique focusses on three specific properties of a DSF (although it will be understood that the technique is not limited to these specific properties and may be used to define other properties of a DSF):

Output Schema: Inferring the schema specifications, i.e., column names and data types, of the DSF's output. The output schema is helpful in compilation of the SQL query (see Q1 in FIG. 1).

Input Schema: Inferring the minimal (i.e., mandatory) set of columns in the input table(s) that the DSF needs for execution. This property is used for enabling the projection-push optimization, i.e., eliminating any unneeded (or non-beneficial) columns before transferring the data to the coprocessor. Projection-push can significantly reduce the data transfer overhead if the base table has many columns while the function only needs few of these columns for its internal processing (refer to Q2 in FIG. 1).

Predicate Push: Inferring the possibilities of pushing a post-function predicate, i.e., a predicate on the function's output, to evaluate on the function's input instead. This property is useful for enabling the predicate-push optimization, i.e., eliminating any unneeded records before transferring the data to the coprocessor. Predicate-push can significantly reduce the data transfer overhead, and it can also enable the generation of more efficient query plans by possibly leveraging available access paths (refer to Q3 in FIG. 1).

In a standard un-optimized setup, a database system may employ a “contract function” to communicate with the coprocessor at query time to retrieve DSF properties. This contract function mechanism has several limitations, including:

Communication Overhead: The communication between the database system and the coprocessor goes through the network and, in the TERADATA system, the QUERYGRID, which introduces some unnecessary overhead.

Limited Property Inference and No Optimizations: The mechanism is used to infer only the output schema of the DSF, which is needed for query compilation. No other properties are inferred. As a result, the predicate push and projection push optimizations highlighted above are difficult to detect or apply.

Mock Function Execution: In order to infer the output schema of a DSF, the contract mechanism performs a mock execution of the DSF over empty inputs. The schema of the inputs as well as the values of any parameters extracted from the DSF's invocation are used in the mock execution to get the corresponding output schema. The mock execution introduces some overhead.

Some database systems allow for plan directives to aid in selecting a specific plan for a given query (or at least disabling some of the possible alternatives). This mechanism is used either for debugging purposes or is leveraged by expert users to enforce a specific execution plan. Plan directives have fundamental differences to the technique described herein including: (1) plan directives typically target the standard relational operators, e.g., joins, sorting, selection, and aggregation, while the technique described herein targets the DSFs integrated with the SQL engine; (2) plan directives are usually at the physical plan level, e.g., disabling or enabling a physical operation such as index scan, hash join, nested-loop join, sort-based aggregation, etc., while the technique described herein operates at the logical plan level; (3) plan directives are static, while the technique described herein is dynamic, i.e., the same DSF in two different queries may have different properties; and (4) plan directives usually narrow down the optimizer choices of the possible plans to generate, while the technique described herein opens up additional opportunities for generating more equivalent plans.

The technique described herein employs a “function descriptor” as a possible way of communicating DSF properties between a coprocessor and a database system.

FIG. 4 illustrates the lifecycle of a newly introduced function descriptor from creation to enabling optimizations. In a first phase (labeled “1” in FIG. 4), the developer creates a DSF (or alternatively, the developer integrates the remote coprocessor with the database system) and writes a function descriptor in a markup language, described below. Note that, while FIG. 4 and the discussion of FIG. 4 that follows assumes the DSF is developed to be executed on a “Remote System” and thus could be described as an RAF, the same concepts discussed here apply to DSFs that execute locally at the caller system.

In a second phase (i.e., the “function installation” phase, labeled “2” in FIG. 4) the function descriptor is uploaded to the database system during installation of the DSF (or, alternatively, during the registration of the remote system with the database system) and stored inside the database system for future access.

In a third phase (labeled “3” in FIG. 4), given a query invoking the DSF, the database system retrieves the corresponding function descriptor and applies it over the given function invocation to infer the desired properties.

In a fourth phase (labeled “4” in FIG. 4), the database system uses the inferred properties to annotate the query plan with metadata information.

In a fifth phase (labeled “5” in FIG. 4), the database system's query optimizer uses the metadata added in the fourth phase to trigger appropriate rewriting and optimization, e.g., predicate push and projection push.

The technique uses a markup language for function descriptors. The markup language may be based on Java Script Object Notation (JSON) or any similar method or language for expressing the properties of a DSF. The markup language is (1) simple to allow broad adoptability, (2) highly expressive for wide coverage to most, if not all, DSFs, and (3) extensible for future extensions. The markup language is designed to capture several properties of interest to the query optimizer, including the output schema, input schema, predicate push, cardinality estimation, and interesting ordering of a DSF. The markup language is applicable to describe both DSFs and functions that are native to the database system. The markup language may be verified for syntactic and semantic accuracy and then interpreted at query time or cached and reused for fulfilling subsequent query processing requirements.

The technique uses function descriptors, which are valid instances of the markup language that act as the driver for triggering appropriate optimizations based on the DSF's properties and the query context.

The technique endeavors to enable runtime optimizations that minimize the data transfer back and forth between the database system and the coprocessor. These optimizations may also reduce the execution resource consumption and improve the performance at the coprocessor engine. This is demonstrated with two optimizations mentioned above, namely projection push and predicate push. Note that these optimizations are also applicable for DSFs executing locally on the database system by minimizing the data flow into the relatively expensive function execution. Hence, benefits extend to both remote and local function processing.

Function Descriptor Overview

A function descriptor is a JSON-based document that captures specific properties of a given DSF. These properties are of interest to the query optimizer to help enable better execution plans. In some cases, the properties' values are static, i.e., the values depend solely on the DSF's logic independent of any invocation details. In this case, the function descriptor includes the property name along with its value.

However, in many cases, the properties' values are dynamic, i.e., the values depend on the content of the DSF's invocation, and hence are determined at query time. In this case, the function descriptor includes the property name along with instructions on how to infer the value given specific invocation details. Given the polymorphic nature of most DSFs, dynamic properties are very common.

The dynamic nature of DSF properties is illustrated in FIGS. 2 and 3 for the Unpivot DSF.

The function descriptor is designed with the following principles in mind:

Extensibility: Function descriptors are designed to be extensible, which is one of the reasons JSON-based format is used. The function descriptors may initially be designed to capture a specific set of DSF properties of key interest to the database system's query optimizer. However, the extensibility feature enables extending the descriptors to capture more properties in the future.

Lightweight: Since the function descriptors are retrieved and parsed during query compilation and optimization, the function descriptors are lightweight with respect to storage, retrieval, and parsing.

Usability: Function descriptors are written either by a team developing the corresponding DSFs or by a team registering the DSFs with the database system. The function descriptors are designed as a high-level expressive language for better usability by these teams as well as the ease of comprehending the properties of a given DSF.

Portability: While the function descriptor system may be developed for a specific foreign processor (such as the Aster coprocessor system), the function descriptors should be a backbone for descripting DSFs from other coprocessors, e.g., Spark, TensorFlow, and Fuzzy Logix, as well as local database functions.

The function descriptors capture the following set of DSF properties:

Output Schema: The property captures the schema of the DSF's output, which is consumed within the caller SQL query in the database system. The output schema captures the column names, the data types, and possibly nullability specifications. Usage (Why it is captured): This property is essential for the compilation of the query and the generation of the query plan.

Input Schema: This property captures the minimal (mandatory) set of columns that the DSF needs for execution. Usage (Why it is captured): This property is used for enabling the projection-push optimization, which involves the elimination of any unneeded (or non-beneficial) columns before transferring the data to the coprocessor. Projection-push can significantly reduce the data transfer overhead if the base table has many columns while the DSF only needs few of these columns for its internal processing.

Predicate Push: This property captures the possibility of pushing a post-DSF predicate, i.e., a predicate on the DSF's output, to evaluate on the DSF's input instead. Usage (Why it is captured): This property is used for enabling the predicate-push optimization, which involves the elimination of any unneeded records before transferring the data to the coprocessor. Predicate-push can significantly reduce the data transfer overhead, and it can also enable the generation of more efficient query plans by possibly leveraging available access paths.

Cardinality Estimation: This property captures some estimations on the output cardinality (the number of rows) of the DSF. Usage (Why it is captured): This property is used for enabling better query planning and possibly avoiding Incremental Plan Execution (IPE) overhead.

Interesting Ordering: This property captures whether the output of the DSF inherits or obeys specific ordering or partitioning scheme. Usage (Why it is captured): This property is used for enabling better query planning by possibly avoiding unnecessary re-ordering or re-distribution.

These properties are captured in a JSON document with the following structure:

{ “functionName” : “<FunctionName>”,
“outputSchema”: [<array of Output-Schema docs>],
“inputSchema”: [<array of Input-Schema docs>],
“predicatePush”: [<array of Predicate-Push docs>],
“cardinalityEstimation”: [<array of Cardinality-Estimation docs>],
“interestingOrdering”: [<array of Interesting-Ordering docs>]
}

The elements of the above structure may have the attributes described in Tables 1 and 2.

TABLE 1
	Value
Element Name	Type	Value Domain	Description
“functionName”	String	Arbitrary string	This property specifies the name of
			the corresponding DSF.
“outputSchema”	Array	JSON	This property holds one document
	of	documents of	of type OS-D1 for each output that
	JSON	type OS-D1	is candidate to be the primary
	documents		output. A primary output is an
			output that is returned to the caller
			SQL query. The document contains
			instructions for inferring the output
			schema, e.g., column names and
			data types, at query time.
“inputSchema”	Array	JSON	This property holds one document
	of	documents of	of type IS-D1 for each input to the
	JSON	type IS-D1	DSF. The document contains
	documents		instructions for inferring the
			minimal input schema needed by
			the DSF.
“predicatePush”	Array	JSON	This property holds one document
	of	documents of	for each output that is candidate to
	JSON	type PP-D1	be the primary output. The
	documents		document specifies the types of
			post-DSF predicates that can be
			pushed as pre-DSF predicates.
“cardinalityEstimation”	Array	JSON	This property holds one document
	of	documents of	for each output that is candidate to
	JSON	type CE-D1	be the primary output. The
	documents		document specifies instructions for
			inferring some information on the
			cardinality estimation (number of
			rows) of the DSF's output.
“interestingOrdering”	Array	JSON	This property holds one document
	of	documents of	for each output that is candidate to
	JSON	type ORD-D1	be the primary output. The
	documents		document specifies several
			interesting-ordering properties of
			the DSF's output.

TABLE 2
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“functionName”	Mandatory	No	N/A
“outputSchema”	Mandatory	Yes	Empty array ([ ])
		(Indicates that
		the descriptor does
		not cover the output
		schema specifications,
		and the contract function
		fall back mechanism
		is in effect).
“inputSchema”	Mandatory	Yes	Empty array ([ ])
		(Indicates that
		the descriptor does
		not cover the input
		schema specifications,
		and projection-
		push optimization
		is disabled for
		all inputs).
“predicatePush”	Mandatory	Yes	Empty array ([ ])
		(Indicates that
		the predicate-push
		optimization is
		disabled, i.e., no
		post-DSF predicates can
		be pushed to be
		pre-DSF).
“cardinalityEstimation”	Optional	Yes	Empty array ([ ])
		(Indicates that
		the descriptor does
		provide information on
		the cardinalities of
		the DSF's output).
“interestingOrdering”	Optional	Yes	Empty array ([ ])
		(Indicates that
		the descriptor does
		provide information
		on any interesting
		ordering that the
		DSF's output
		may follow).

The function descriptor mechanism is introduced to overcome the limitations of the contract function described above. However, the function descriptors are optional, i.e., a DSF may not have a function descriptor. This may be because either the DSF's properties are too complex to be expressed using the markup language, the DSF is not of a high-priority and the developer is not willing to spend some time developing the function descriptor, or there is not enough expertise to develop the function descriptor. Therefore, the function descriptor mechanism is designed to co-exist with the contract function mechanism.

FIGS. 5A and 5B illustrate the co-existence of the two mechanisms. When a DSF is invoked, the database system 502 checks the function descriptor 504 first. If a function descriptor 504 for the DSF exists, is available, and can provide the value of the requested property, then the contract function 506 is not accessed, as shown in FIG. 5A. Otherwise, as shown in FIG. 5B, the contract function 506 is executed (this is only for the output schema inference because the other properties may not be supported by the contract functions). If neither a function descriptor 504 nor a contract function 506 exists or provides the value of the requested property for the DSF, optimization of the DSF is not implemented.

Markup Language: Building Block Instructions

The markup language is the language in which the function descriptors are expressed. It is a high-level JSON-based language consisting of a set of instructions.

The language may have two building block instructions that are used to express the values of the different properties. These instructions are:

ADD Instruction: The ADD instruction is used to add column information to the output or input schema lists. The added columns may come from one of the DSF's input tables, one of the DSF's invocation parameters, or they may be predefined.

CASE Instruction: The CASE instruction is a control (branching) instruction, very similar to the CASE statement in programming languages. It is used because in many cases the content of a DSF's output schema may depend on the presence or absence of some values or parameters. Such conditional construction of the property value requires a CASE instruction.

ADD Instruction

The ADD instruction is used to add column information to the output schema list or input schema list. The instruction includes specifications to define: (1) The location within the list to which the added column(s) are augmented, (2) the source that provides the column information to be added, which can be one of the input tables or one of the DSF parameters, (3) the data types of the columns to be added, and (4) any manipulation operations to be applied on the column names before adding them to the list.

FIG. 6 shows the structure of the ADD instruction. The ADD instruction consists of three nested document types, referred to as ADD-D1, ADD-D2, and ADD-D3 as indicated in FIG. 6.

The ADD-D1 document is the first (parent) level of an ADD instructions. The elements of document type ADD-D1 are described in Tables 3 and 4.

TABLE 3
	Value
Element Name	Type	Value Domain	Description
“instruction”	String	Fixed value = “ADD”	This property indicates that the
			document is an ADD instruction.
“placement”	String	The value is one of the	This property specifies the
		following values:	position to which the newly
		“begin”,	added columns are inserted in the
		“end”,	generated output schema list.
		“before	“begin”: The new columns are
		<ColName>”,	added to the begin of the list (e.g.,
		“after <ColName>”	the left-most position)
			“end”: The new columns are added
			to the end of the list (e.g., the right-
			most position)
			“before <ColName>”: The new
			columns are added just before the
			column name specified in variable
			<ColName>.
			Constraint: <ColName> must be
			already added to the output
			schema list by aprevious ADD
			instruction. Otherwise, an error is
			generated.
			“after <ColName>”: The new
			columns are added just after the
			column name specified in variable
			<ColName>.
			Constraint: <ColName> must be
			already added to the output schema
			list by a previous ADD instruction.
			Otherwise, an error is generated.
“arguments”	Array of	JSON documents of type	Each document must define one or
	documents	ADD- D2	more columns to be added to the
			output schema list

TABLE 4
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“instruction”	Mandatory	No	N/A
“placement”	Mandatory	No	N/A
“arguments”	Mandatory	No	N/A

Document type ADD-D2 is the second level document in an ADD instruction, as shown in FIG. 6. Document ADD-D2 describes the details of the columns to be added to the output schema list. The elements of document type ADD-D2 are described in Tables 5 and 6.

TABLE 5
	Value
Element Name	Type	Value Domain	Description
“source”	String	The value must be one of	This property specifies the source of the
		the following:	columns to be added (the lineage).
		“inputTable”,	“inputTable”: Indicates that the column
		“parameter”,	names to be added come from one of the
		“predefined”	DSF's input tables (which are defined in
			the ON clause(s) in the DSF invocation).
			“parameter”: Indicates that the column
			names to be added come from one of the
			DSF's parameters (which are defined in
			the USING clause in the DSF
			invocation).
			“predefined”: Indicates that the column
			names to be added are predefined, i.e., the
			names depend on neither the inputs nor
			the parameters in the DSF invocation.
			These are usually additional columns that
			the DSF adds to the output, and the
			names, data types, etc. of these columns
			are predefined.
“name”	String	The allowed value depends	This property specifies which columns
		on the value assigned to	are inherited from the “source” and
		the “source” property, and	augmented to the output schema list.
		it must follow specific
		expressions as specified in
		Table 7.
“dataType”	String	The value must be either a	This property specifies the data type of
		constant string =	the column(s) added to the output schema
		“unchanged”, or string	list.
		defining the data type and	“unchanged”: Indicates that the columns
		the null-ability status	will inherit the same data types they
			already have.
			Constraint: This setting is allowed only if
			the “source” property is set to
			“inputTable” or “parameter”.
			Implicit Assumption: It is implicitly
			implied that these columns allow Null (a
			conservative assumption). The markup
			language could be extended to provide
			more detailed specifications at the
			expense of adding complexity.
			“<dataType> [Not Null]”: Indicates that
			the column(s) conforms to the data type
			specified in <dataType>.
			The “Not Null” specification is optional.
			If it is known that the column(s) will not
			have Null values, then “Not Null” can be
			defined along with the data type.
			Constraint: This setting must be used if
			the “source” property is set to
			“predefined”. Otherwise, this setting can
			be still used if the “source” property is set
			to “inputTable” or “parameter”,
			especially if some manipulation
			operations are applied on the column(s).
“manipulations”	Array of	JSON documents of type	In some cases, some manipulation
	JSON	ADD-D3	operations need to be applied on the
	documents		column names before adding them to the
			output schema list, e.g., prefixing the
			names with some fixed string. The
			“manipulations” element defines these
			operations (if any).

TABLE 6
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“source”	Mandatory	No	N/A
“name”	Mandatory	No	N/A
“dataType”	Mandatory	No	N/A
“manipulations”	Optional	Yes (Indicates	Either the entire
		that there are	element is not
		no manipulation	present, or an
		operations to be	empty array ([ ]).
		applied on the
		column names).

The allowed expressions in the “name” property in document type ADD-D2 are listed in Table 7.

TABLE 7
Value in	Allowed Expression
“source”	in “name”	Description	# of Added Columns
“inputTable”	inputId.*	Adding all columns from	Equal to the number of
		inputId to the output schema list	columns in inputId
	inputId.ColName	Adding the column named ColName	One
		from inputId to the output schema
		list
	inputId.[i, j]	Adding the columns starting from	j − i + 1
		position i (inclusive) to position j
		(inclusive) from inputId to the output
		schema list. The position index starts
		from 1.
	inputId.PartitionBy.*	Adding the columns in the Partition	Equal to the number of
		By clause of inputId to the output	columns in the
		schema list	Partition By clause of
			inputId
	inputId.OrderBy.*	Adding the columns in the Order By	Equal to the number of
		clause of inputId to the output	columns in the Order
		schema list	By clause of inputId
“parameter”	parameterName.*	Adding all the columns from the	Equal to the number
		DSF's parameter parameterName to	of columns referenced
		the output schema list	in parameterName
	parameterName.ColName	Adding the column ColName from	One
		the DSF's parameter parameterName
		to the output schema list
“predefined”	ColName	String specifying the predefined	One
		column name to be added to the
		output schema list

Referencing a specific input within a DSF's invocation (which is denoted by the “inputId” variable in the expressions listed in Table 7) follows the following rules:

Position-Based Referencing: In this referencing scheme, a reference to a specific input is achieved by the position of its ON clause relative to the other ON clauses. For example, the input in the first ON clause is referenced as “input1”, the input in the second ON clause is referenced as “input2”, etc.

Applicability: This referencing scheme is applicable only if the order of the inputs in the DSF's invocation is fixed and there are no optional ON clauses.

Alias-Based Referencing: In this referencing scheme, a reference to a specific input is achieved by its alias. For example, the input in the first ON clause has alias “model”, and thus this input can be referenced in the “name” property of document type ADD-D2 by that alias.

Applicability: This referencing scheme is applicable only if the DSF's manual mandates specific aliases to be given to the inputs. In this case, the relative order among the ON clauses in not important, and hence the Relative-Order referencing must not be used.

Referencing a specific parameter from a DSF's invocation (which is denoted by the “parameterName” variable in the expressions listed in Table 7) is straightforward because parameters can be only referenced by their names.

Document type ADD-D3 is the third level document in an ADD instruction. If there are some manipulation operations to be applied over the column names before adding them to the output schema list, then these operations are defined according to the structure in document type ADD-D3, as shown in FIG. 6.

The elements of document type ADD-D3 are described in Tables 8 and 9:

TABLE 8
	Value
Element Name	Type	Value Domain	Description
“operation”	String	The value must be one of	This property specifies the type
		the following values:	of the manipulation operation
		“concat”,	to be applied.
		“prefix”,
		“postfix”
“argument”	String	The value is either a constant	The property provides the string,
		string or string retrieved from	which is an argument to the
		a DSF's parameter.	operation.
		Therefore, the value is either:	“<constant>”: The argument string
		<constant>,	is a fixed string.
		valueFromParameter(<paramName>)	valueFromParameter(<paramName>):
			The argument string is dynamic and
			it is retrieved at query time from
			the DSF's parameter paramName.

TABLE 9
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“operation”	Mandatory	No	N/A
“argument”	Mandatory	No	N/A

As an example of use of the ADD instruction, the output schema specifications of the Unpivot( ) function shown in FIGS. 2 and 3 are shown below. For simplicity, in this example only first three entries from FIG. 2, which do not require checking to determine whether the InputTypes parameter is set to “true,” are included. Note that the list is order sensitive. Note also that the names written in the “Accumulate” field depend on the parameter names in the query itself. For example, in the Unpivot( ) function shown in FIG. 2, the parameters attribute_column and value_column are accumulate columns.

{“instruction”: “ADD”,
“placement”: “end”,
“arguments”: {
{“source”: “parameter”,
“name”: “Accumulate.*”,
“datatype”: “unchanged”,
“manipulations”: [ ]
},
{“source”: “parameter”,
“name”: “AttributeColumn.*”,
“datatype”: “VARCHAR”,
“manipulations”: [ ]
},
{“source”: “parameter”,
“name”: “ValueColumn.*”,
“datatype”: “VARCHAR”,
“manipulations”: [ ]
},
}

CASE Instruction

The CASE instruction is used when there are different alternatives for the output schema based on some conditions, e.g., the presence of absence of a specific parameter, or a parameter is set to specific value. In such scenarios, the CASE instruction is used to create the branching.

As in most programming languages, at most one CASE branch can evaluate to True to be executed. Once the condition(s) of one branch evaluate to True, no further branches are evaluated. It is possible that none of the branches evaluates to True, and as a result none of them is executed.

The CASE instruction consists of three levels of documents, referred to as CASE-D1, CASE-D2, CASE-D3.

FIG. 7 shows the structure of the CASE instruction.

Document type CASE-D1 is the parent document of the CASE instruction, as shown in FIG. 7. The elements of document type CASE-D1 are described in Tables 10 and 11.

TABLE 10
	Value
Element Name	Type	Value Domain	Description
“instruction”	String	Fixed	This property indicates
		value =	that the document is a
		“CASE”	CASE instruction.
“arguments”	Array of	JSON	Each document defines one
	documents	documents	condition-action branch
		of type	in the case statement.
		CASE-D2

TABLE 11
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“instruction”	Mandatory	No	N/A
“arguments”	Mandatory	No	N/A

Document type CASE-D2 is the second-level document of the CASE instruction, as shown in FIG. 7. Each document contains a list of one or more conjunctive predicates, and a list of one or more actions to be taken if the predicates evaluate to True. The elements of document type CASE-D2 are described in Tables 12 and 13.

TABLE 12
	Value
Element Name	Type	Value Domain	Description
“condition”	Array of	JSON	This is an array of
	documents	documents	documents representing
		of type	a list of conjunctive
		CASE-D3	predicates. That is, if
			all the predicates evaluate
			to True, then the actions
			defined in the
			“action” element
			are executed.
“action”	Array of	JSON	Each document is an ADD
	documents	documents	instruction to add columns
		of type	to the output schema list.
		ADD-D1	The “action”
			documents execute only
			if the “condition”
			documents evaluate
			to True.

TABLE 13
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“condition”	Mandatory	Yes (Indicates	Empty array ([ ])
		that the condition
		evaluates to True.
		Typically, as in
		most programming
		languages, this is
		the last branch
		of the CASE
		instruction.)
“action”	Mandatory	Yes (Indicates	Empty array ([ ])
		that no action
		is taken.)

Document type CASE-D3 is the third-level document of the CASE instruction, as shown in FIG. 7. Each document represents a single condition. The elements of document type CASE-D3 are described in Tables 14 and 15.

TABLE 14
	Value
Element Name	Type	Value Domain	Description
“leftOperand”	String	The value must be:	This field specifies the
		“valueFromParameter	parameter from which the left
		<paraName>”	operand value is retrieved.
“operator”	String	The value must be one of the	Each document is an ADD
		following operators:	instruction to add columns to
		“Exists”,	the output schema list. The
		“Not Exists”,	“action” documents execute
		“In”,	only if the “condition”
		“Not In”,	documents evaluate to True.
		“Contains”,	All operators except “Exists”
		“Like”,	and “Not Exists” mandate the
		“Between”,	presence of both the left and
		“=”,	right operands, and the
		“!=”,	operators' semantics follows
		“>”,	that of programming languages.
		“>=”,	“Exists” and “Not Exists”
		“<”,	require only the left operand to
		“<=”,	be present, and they have the
			following semantics:
			“Exists”: Returns True iff (if-
			and-only-if) the parameter
			specified in the “leftOperand”
			exists in the DSF invocation.
			Otherwise, returns False.
			“Not Exists”: Returns True iff
			the parameter specified in the
			“leftOperand” does not exist in
			the DSF invocation.
			Otherwise, returns False.
“valueType”	String	The value is one of the	This property defines the data
		following values:	type based on which the
		“int”,	comparison operator is
		“double”,	evaluated.
		“string”
			Constraints:
			For operators “Exists”
			and “Not Exists”, the
			property can be omitted,
			otherwise it is ignored.
			For all other operators, it
			is mandatory.
			For operators “Contains”
			and “Like”, the
			“valueType” must be set
			to “string”.
“rightOperand”	String	The value is one of the	The value of the right operand
		following:	is either provided as a constant
		<constant>, or	or retrieved from the specified
		valueFromParameter(<paraName>)	parameter.
			The format of the
			“rightOperand” field depends
			on the comparison operator
			being used.
			Constraints:
			For operators “Exists”
			and “Not Exists”, the property
			can be omitted, otherwise it is
			ignored. For all other operators,
			it is mandatory.
			“Exists”: This property can be
			omitted, otherwise it is ignored.
			“Not Exists”: This property can
			be omitted, otherwise it is
			ignored.
			“In”: the value in the right
			operand is a comma-separated
			string. The string is parsed into
			a list of string values (the
			separator is the comma
			character). Finally, each string
			is casted to the type defined in
			“valueType” for evaluation.
			“Not In”: the value in the right
			operand is a comma-separated
			string. The string is parsed into
			a list of string values (the
			separator is the comma
			character). Finally, each string
			is casted to the type defined in
			“valueType” for evaluation.
			“Contains”: the value in the
			right operand is a string that is
			checked as to whether it is
			contained within the left
			operand value.
			“Like”: the value in the right
			operand is a regular expression
			string. The expression is
			evaluated using the standard
			C++ regular expression library
			(regex).
			“Between”: the value in the
			right operand is a string
			consisting of two comma-
			separated values. The first value
			is the lower bound, and the 2nd
			value is the upper bound. The
			values are casted to the type
			defined in “valueType” for
			evaluation.
			“=”: the value in the right
			operand is a string, which is
			casted to the type defined in
			“valueType” for evaluation.
			“!=”: the value in the right
			operand is a string, which is
			casted to the type defined in
			“valueType” for evaluation.
			“>”: the value in the right
			operand is a string, which is
			casted to the type defined in
			“valueType” for evaluation.
			“>=”: the value in the right
			operand is a string, which is
			casted to the type defined in
			“valueType” for evaluation.
			“<”: the value in the right
			operand is a string, which is
			casted to the type defined in
			“valueType” for evaluation.
			“<=”: the value in the right
			operand is a
			string, which is casted to the
			type defined in “valueType” for
			evaluation.

TABLE 15
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“leftOperand”	Mandatory	No	N/A
“operator”	Mandatory	No	N/A
“valueType”	Optional	No	N/A
	(Only if the
	operator is set
	to “Exits” or
	“Not Exists”)
“rightOperand”	Optional	No	N/A
	(Only if the
	operator is set
	to “Exits” or
	“Not Exists”)

As an example, the CASE instruction may be used to add the last three entries from the example illustrated in FIGS. 2 and 3 as shown below.

(“instruction”: “CASE”,
“argument”:[
{“condition”:[{“leftOperand”:“valuefromParameter(InputTupe
s)”,
“operator”:“in”,
“valueType”:“string”,
“rightOperand”:“true,yes,t,y,1”
}]
“action”:[
{“instruction”: “ADD”,
“placement”: “end”,
“arguments”: {
{“source”: “parameter”,
“name”: “ValueColumn.*”,
“datatype”: “DOUBLE”,
“manipulations”:[{“operation”:”postfix”,
“argument”: “_double”]
}
]
},
{“source”: “parameter”,
“name”: “ValueColumn.*”,
“datatype”: “LONG”,
“manipulations”:[{“operation”:”postfix”,
“argument”: “_long”]
}
]
},
{“source”: “parameter”,
“name”: “ValueColumn.*”,
“datatype”: “VARCHAR”,
“manipulations”:[{“operation”:”postfix”,
“argument”: “_str”]
}
]
},
}

Markup Language: DSF Property Specifications

As mentioned previously, the markup language is designed to capture the following properties:

Output Schema

Input Schema

Predicate Push

Cardinality Estimation

Interesting Ordering

In the following sections, the instructions and the language specifications for each of the targeted properties is presented.

Output Schema Property

The Output Schema property captures the schema of the DSF's output, which is consumed within the caller SQL query in the database system. The output schema is an ordered list and it captures the column names, the data types, and possibly nullability specifications.

This property is used for the compilation of the query and the generation of the query plan.

A DSF may generate multiple outputs. One of the outputs, referred to as the “primary output,” may be returned to the caller SQL query. The other outputs are sent from the coprocessor system to the database system and are stored in relational tables. The function descriptor for the DSF maintains the output schema for each output that is candidate to be a primary output.

The DSF developer may decide which outputs are candidates to be a primary output. For example, assume a DSF, F( ), generates 5 outputs, namely O1, O2, O3, O4, and O5. Assume also that the

DSF developer (who is assumed to be a domain expert) decided that only O1, O2, and O3 are candidates to be a primary output. In this case, each of these three outputs is given a unique Id (name), and the DSF descriptor includes the output schema specifications for each O1, O2, and O3 output. At query time, the user can pass a parameter to the DSF specifying which among the candidate outputs is the primary output.

Each candidate output may have a JSON document describing its output schema under the “$.outputSchema” element, where “$” is the root document (the document of type DESC-D1). In the case in which the output does not have a corresponding document, the database system uses the contract function mechanism to figure out the output schema, as discussed above in connection with FIGS. 5A and 5B.

The output schema document (of type OS-D1) has the following structure:

{ “property” : “outputSchema”
“outputId”: “<Output-unique-Id>”
“instructions”: <array of documents, each describes one
instruction>
}

An output schema document may use both the ADD and CASE instructions in the instructions array. The elements in document type OS-D1 are described in Tables 16 and 17.

TABLE 16
	Value
Element Name	Type	Value Domain	Description
“property”	String	Fixed value =	This property indicates that the
		“outputSchema”	document is describing an output
			schema.
“outputId”	String	The value is one of the	This property specifies the output
		unique names assigned	whose schema is defined by this
		to the candidate	document.
		outputs.
		If there is only one
		candidate output, then
		the value is set to
		“standard”.
“instructions”	Array of	JSON documents of	Each document in the array is a
	documents	type:	single instruction document. The
		ADD-D1 (ADD	allowed instructions include:
		instruction), or	ADD,
		CASE-D1 (CASE	CASE
		instruction)

TABLE 17
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“property”	Mandatory	No	N/A
“outputId”	Mandatory	No	N/A
“instructions”	Mandatory	Yes (Indicates	Empty array ([ ])
		that the descriptor
		does not provide
		output schema
		specifications
		for the output
		identified by
		“outputId”. As a
		result, for that
		output, the contract
		function mechanism
		is used).

Input Schema Property

The input schema property captures the minimal (mandatory) set of columns that the DSF needs for execution.

The input schema property is used for enabling the projection-push optimization, which involves the elimination of any un-needed (or non-beneficial) columns before transferring the data to the coprocessor. Projection-push can significantly reduce the data transfer overhead if the base table has many columns while the DSF only needs few of these columns for its internal processing.

A DSF may have multiple inputs (multiple ON clauses), and each of these inputs may have a document under the “$.inputSchema” array element, where “$” is the root document (The document of type DESC-D1). If a given input does not have a corresponding document, then this implies that the projection-push optimization is not enabled for that input, and whatever the user specifies is sent to the coprocessor.

The input schema document (of type IS-D1) has the following structure:

{ “property” : “inputSchema” ,
  “inputId”: “<inputPosition>” | “<alias>”,
  “surplus”: “unknown” | “notAllowed” | “ignored” |
“propagatedBack”,
“nullElimination”: “Yes” | “No”,
“instructions”: <array of documents, each describes one
instruction>
}

The elements in document type IS-D1 are described in Tables 18 and 19:

TABLE 18
	Value
Element Name	Type	Value Domain	Description
“property”	String	Fixed value =	This property indicates that the
		“inputSchema”	document is describing an input
			schema.
“inputId”	String	The string is either an	This property specifies the input
		integer number	whose minimal schema is being
		indicating the input's	defined by this document.
		position among the	<inputPosition>: An integer
		other inputs (if any), or	number indicating the input's
		an arbitrary alias string	position among the other inputs
			(if any). It is used for DSFs
			using Position-Based
			Referencing.
			<alias>: Arbitrary string
			referring to one of the inputs'
			alias. It is used for DSFs using
			Alias-Based Referencing.
“surplus”	String	The value must be one	This property specifies whether
		of:	or not the DSF can accept
		“unknown”,	additional columns beyond the
		“notAllowed”,	required ones (called “surplus”)
		“ignored”,	and specifies the behavior of the
		“propagatedBack”	DSF with respect to these
			additional columns. This
			behavior affects the types of
			optimizations that can be
			applied.
			“unknown”: Indicates that the
			behavior is unknown, and the
			projection-push optimization is
			disabled in this case.
			“notAllowed”: Indicates that the
			DSF does not accept any
			additional columns beyond
			those needed by the DSF. If
			additional columns are sent, the
			execution fails.
			For DSFs exhibiting this
			behavior, projection push is not
			required. Therefore, the
			“instructions” field can be
			simply set to [ ].
			“ignored”: Indicates that the
			DSF accepts additional columns
			(usually the required ones are
			defined in some parameters).
			These additional columns are
			ignored (dropped) by the DSF.
			For DSFs exhibiting this
			behavior, projection push is
			critical, and these additional
			columns should be eliminated.
			“propagatedBack”: Indicates
			that the DSF accepts additional
			columns (usually the required
			ones are defined in some
			parameters). These additional
			columns are ignored (skipped)
			during the evaluation, but they
			are returned along with the
			results, e.g., pass though
			columns.
			For DSFs exhibiting this
			behavior, projection push is a
			possibility. If the additional
			columns are not referenced in
			the caller SQL query, then they
			can be eliminated before the
			DSF's execution. Otherwise, the
			decision to apply the projection
			push (to eliminate the additional
			columns) becomes a cost-based
			optimization (either send them
			and get them back as pass
			though columns, or not to send
			them but execute a local join
			with the DSF's output to
			augment them back to the result
			set).
“nullElimination”	String	The value must be one	This property specifies whether
		of: “Yes”, “No”	or not the DSF ignores the
			records containing a null value in
			any of the required (mandatory)
			columns.
			“Yes”: Indicates that the DSF
			ignores the records containing a
			null value in any of the required
			(mandatory) columns (which are
			specified by the “instructions”
			field (See below). If that is the
			case, the database system can
			eliminate these records before
			sending them to the coprocessor
			(it is a cost-based optimization
			from database system).
			“No”: Indicates that the
			optimization of null-based record
			elimination is disabled for the
			input identified by the “inputId”
			field.
“instructions”	Array of	JSON documents of	Each document in the array is a
	documents	type ADD-D1 (ADD	single instruction document. The
		instruction)	allowed instructions include:
			ADD
			CASE

TABLE 19
		Empty Content	Empty Content
Element Name	Presence	Allowed	Representation
“property”	Mandatory	No	N/A
“inputId”	Mandatory	No	N/A
“surplus”	Mandatory	No	N/A
“nullElimination”	Optional	No	N/A
“instructions”	Mandatory	Yes (Indicates	Empty array ([ ])
		that for this
		input, just send
		whatever the user
		specifies in the
		query).

It is possible that for a given output, the “instructions” array in the OS-D1 document is not empty. However, when these instructions are applied over a given DSF invocation, they may return an empty output schema list. As an example, the “instructions” array may contain a CASE instruction, and at query time, none of the CASE branches evaluates to true. Since the output schema specifications are mandatory to have for the compilation of the query, an empty output schema list means that the function descriptor fails to provide the desired value, and the contract function fall back mechanism is executed, as described above in connection with FIGS. 5A and 5B.

In general, the specifications of the input schema are less important in some scenarios such as the following ones:

Small Dimensions Table as Input: If the input to be sent is very small, then projection push is not effective anyway.

Automatically-Generated Input: If the input table is generated from another function (e.g., from a nested function), and that generated input has the columns required by the outer function and there is no need for projection push on that input for the outer function.

Surplus=“NotAllowed”: In this case, the execution fails anyway if any extra columns are sent to the DSF, and thus projection push is not critical (the end-used should have taken care of that).

Surplus=“Unknown”: In this case, the projection-push optimization is disabled, and the “instructions” field is ignored.

In these scenarios, where the specifications of the input schema are less important, the recommendation is to simply set the “instructions” array for that input to empty array ([ ]).

Predicate Push Property

The predicate push property captures the possibility of pushing a post-DSF predicate, i.e., a predicate on the DSF's output, to be evaluated on the DSF's input instead. This property is used for enabling the predicate-push optimization, which involves the elimination of any unneeded records before transferring the data to the coprocessor. Predicate-push can significantly reduce the data transfer overhead, and it can also enable the generation of more efficient query plans by possibly leveraging available access paths.

A function descriptor may contain a predicate-push specification document for each candidate output under the “$.predicatePush” element, where “$” is the root document (i.e., the document of type DESC-D1). If for a given candidate output there is no corresponding document, then this means that the predicate push optimization is not enabled for that output. The predicate push document has the following structure:

{ “property” : “predicatePush” ,
“outputId”: “<output-unique-Id>”,
“targetInputId”: “<inputPosition>” | “<alias>”,
“rowIndependence”: “Yes” | “No” ,
“partitionIndependence”: “Yes” | “No”
}

The elements in document type PP-D 1 are described in Tables 20 and 21.

TABLE 20
	Value
Element Name	Type	Value Domain	Description
“property”	String	Fixed value =	This property indicates that the
		“predicatePush”	document is describing the
			predicate push specifications.
“outputId”	String	The value is one of	This property specifies the
		the unique names	output whose predicate-push
		assigned to the	properties are defined by this
		candidate outputs.	document.
		If there is only one
		candidate output,
		then the value is set to
		“standard”.
“targetInputId”	String	The string is either an	This property specifies the
		integer number	DSF's input that is targeted for
		indicating the input's	the predicate push (if a
		position among the	predicate is pushed). This
		other inputs (if any),	property is needed because a
		or an arbitrary alias	DSF may have multiple inputs
		string	and it is possible to have
			ambiguity with respect to
			which input should receive the
			pushed predicate (if any).
			<inputPosition>: An integer
			number indicating the input's
			position among the other
			inputs (if any). It is used for
			DSFs using Position-Based
			Referencing
			<alias>: Arbitrary string
			referring to one of the inputs'
			alias. It is used for DSFs using
			Alias-Based Referencing
“rowIndependence”	String	The value must be	This property is set to “Yes”
		one of the following:	iff the DSF satisfies the
		“Yes”, “No”	following two conditions:
			1. The DSF operates on each
			row independently of the
			others.
			2. For a single input row, the
			DSF may produce:
			zero output rows, or
			one or more rows all as
			the input row values
			(or subset of its fields)
			plus possibly
			additional columns
			If this property is set to “Yes”,
			then a post-DSF predicate on
			any of the input columns, that
			pass through to the output, can
			be pushed as a pre-DSF
			predicate and is evaluated on
			the DSF's input.
“partitionIndependence”	String	The value must be	This property is set to “Yes” iff
		one of the following:	the DSF satisfies the following
		“Yes”, “No”	two conditions:
			The DSF operates on each
			partition independent of the
			others
			For a single input partition, the
			DSF may produce:
			zero output rows, or
			one or more rows all
			have the same values in
			the partitioning
			columns as those of the
			input partition (or
			subset of its fields) plus
			possibly additional columns
			If this property is set to “Yes”,
			then a post-DSF predicate on
			any of the partition by columns,
			that pass through to the output,
			can be pushed as a pre-DSF
			predicate and gets evaluated on
			the DSF's input.
			Constraint:
			“partitionIndependence” is
			applicable only to DSFs whose
			base input is partitioned, i.e.,
			has PARTITION BY clause)
			involving non-trivial
			expressions, i.e., excluding
			expressions such as
			“PARTITION BY 1,” that do
			not actually result in
			partitioning.

TABLE 21
		Empty
		Content	Empty Content
Element Name	Presence	Allowed	Representation
“property”	Mandatory	No	N/A
“outputId”	Mandatory	No	N/A
“targetInputId”	Optional	No	N/A
	(If the DSF has one
	input, or both
	“rowIndependence” and
	“partitionIndependence”
	are set to “No”)
“rowIndependence”	Mandatory	No	N/A
“partitionIndependence”	Mandatory	No	N/A

Cardinality Estimation Property

The cardinality estimation property captures some estimates of the output cardinality (e.g., number of rows) of the DSF. This property is used for enabling better query planning, e.g., join planning and possibly avoiding IPE overhead.

A function descriptor may contain a cardinality estimation specification document for each candidate output under the “$.cardinalityEstimation” element, where “$” is the root document (The document of type DESC-D1). If for a given output there is no corresponding document, then this means that there is no cardinality estimation information available for that output. The document has the following structure:

{ “property” : “cardinalityEstimation”,
“outputId”: “<Output-unique-Id>”,
“minCardinality”: “unknown” | “<expr>” ,
“maxCardinality”: “unknown” | “<expr>” ,
“expectedCardinality”: “unknown” | “<expr>” ,
“expConfidence” : “low” | “high”
}

The elements in document type CE-D1 are described in Tables 22, 23, and 24.

TABLE 22
	Value
Element Name	Type	Value Domain	Description
“property”	String	Fixed value =	This property indicates that the
		“cardinalityEstimation”	document is describing the
			cardinality estimation
			specifications.
“outputId”	String	The value is one of	This property specifies the output
		the unique names	whose cardinality, i.e., the number
		assigned to the	of rows, is estimated by this
		candidate outputs.	document.
		If there is only one
		candidate output,
		then the value is
		set to “standard”.
“minCardinality”	String	The value is either	This property defines the minimum
		“unknown”, or a	number of rows that can be
		string expression	returned from the DSF.
		over some pre-	“unknown”: Indicates that the
		defined operands	minimum cardinality is unknown.
			<expr>: A simple algebraic
			expression over operands of the
			format defined in Table
			Cardinality-Expr.
			The allowed algebraic operators in
			the expression are:
			“+”, “−“, “*”, “/”
“maxCardinality”	String	The value is either	This property defines the maximum
		“unknown”, or a	number of rows that can be
		string expression	returned from the DSF.
		over some pre-	“unknown”: Indicates that the
		defined operands	maximum
			cardinality is unknown.
			<expr>: A simple algebraic
			expression over operands of the
			format defined in Table
			Cardinality-Expr.
			The allowed algebraic operators in
			the expression are:
			“+”, “−“, “*”, “/”
“expectedCardinality”	String	The value is either	This property defines the expected
		“unknown”, or a	(average) number of rows that can
		string expression	be returned from the DSF.
		over some pre-	“unknown”: Indicates that the
		defined operands	average cardinality is unknown.
			<expr>: A simple algebraic
			expression over operands of the
			format defined in Table
			Cardinality-Expr.
			The allowed algebraic operators in
			the expression are:
			“+”, “−“, “*”, “/”
“expConfidence”	String	The value is one	The confidence of the provided
		of: “low”, “high”	estimations. This confidence is for
			the “expectedCardinality” field.

TABLE 23
Operand Format                Meaning
|<InputId>|                   The cardinality of InputId.
                              InputId is an input reference that uses
                              either of Position-Based Referencing
                              or Alias-Based Referencing
|<InputId>.PartitionBy|       The number of partitions created by
                              the PartitionBy clause of InputId
|<ParameterName>|             The number of values listed in the
                              specified parameter
valueFromParameter(<parName>) The value defined inside the specified
                              parameter
val                           Constant value

TABLE 24
		Empty
		Content	Empty Content
Element Name	Presence	Allowed	Representation
“property”	Mandatory	No	N/A
“outputId”	Mandatory	No	N/A
“minCardinality”	Mandatory	No	N/A
“maxCardinality”	Mandatory	No	N/A
“expectedCardinality”	Mandatory	No	N/A
“expConfidence”	Mandatory	No	N/A

Interesting Ordering Property

The interesting ordering property captures whether the output of the DSF inherits or obeys specific ordering or partitioning schemes. This property is used for enabling better query planning by possibly avoiding unnecessary re-ordering or re-distribution.

A function descriptor may contain an interesting-ordering specification document for each candidate output under the “$.interestingOrdering” element, where “$” is the root document (The document of type DESC-D1). If for a given output there is no corresponding document, then this means that no interesting ordering properties are available for that output. The document has the following structure:

{ “property” : “interestingOrdering ” ,
“outputId”: “<Output-unique-Id>”,
“preserveInPartitioning”: “No” | “<inputPosition>” | “<alias>”,
“preserveInOrdering”: “No” | “<inputPosition>” | “<alias>”,
“introducePartitioning”: “No” | <expr>,
“introduceOrdering”: “No” | <expr>
}

The elements in document type ORD-D1 are described in Tables 25 and 26:

TABLE 25
	Value
Element Name	Type	Value Domain	Description
“property”	String	Fixed value =	This property indicates
		“interestingOrdering”	that the document is
			describing the
			interesting-ordering
			specifications.
“outputId”	String	The value is one of	This property specifies
		the unique names	the output whose
		assigned to the	interesting orderings
		candidate outputs.	are described by this
		If there is only one	document.
		candidate output,
		then the value is set
		to “standard”.
“preserveInPartitioning”	String	The string can be a	This property specifies
		constant “No”, or an	whether or not the
		integer number	output of the DSF is
		indicating the input's	preserving the same
		position among the	partitioning on the
		other inputs (if any),	input.
		or an arbitrary alias	“No”: Indicates that no
		string	partitioning is preserved.
			<inputPosition>: An
			integer number
			indicating the input's
			position among the
			other inputs (if any).
			This is the input that
			the output follows its
			partitioning scheme. It
			is used for DSFs using
			Position-Based Referencing
			<alias>: Arbitrary
			string referring to one
			of the inputs' alias.
			This is the input that
			the output follows its
			partitioning scheme. It
			is used for DSFs using
			Alias-Based Referencing
“preserveInOrdering”	String	The string can be a	This property specifies
		constant “No”, or an	whether or not the
		integer number	output of the DSF is
		indicating the input's	preserving the same
		position among the	ordering of the input.
		other inputs (if any),	“No”: Indicates that no
		or an arbitrary alias	ordering is preserved.
		string	<inputPosition>: An
			integer number
			indicating the input's
			position among the
			other inputs (if any).
			This is the input that
			the output follows its
			ordering scheme. It is
			used for DSFs using
			Position-Based
			Referencing (Section
			3.1.2).
			<alias>: Arbitrary
			string referring to one
			of the inputs' alias.
			This is the input that
			the output follows its
			ordering scheme. It is
			used for DSFs using
			Alias-Based Referencing.
“introducePartitioning”	String	The string can be a	This property specifies
		constant “No”, or a	whether or not the DSF
		comma-separated-	is producing the output
		value string	with a specific new
			partitioning.
			“No”: Indicates that no
			specific partitioning is
			introduced.
			<csv-expr>: The
			comma-separated-value
			expression captures the
			output columns on
			which the output is
			partitioned. It takes one
			of the following two
			forms:
			csv-columnNames
			(“Col1, Col2, . . .”): A
			list of column names.
			Constraints: Column
			names work only if the
			names are predefined,
			i.e., do not depend on
			the DSF's inputs or
			parameters.
			csv-columnPositions
			(“2, 4, . . .”): A list of
			column positions.
			Constraints: Column
			positions work only if
			the names are dynamic
			but the positions are
			fixed.
			If neither of the column
			names nor the column
			positions are
			applicable, then the
			value msut be set to
			“No”.
“introduceOrdering”	String	The string can be a	This property specifies
		constant “No”, or a	whether or not the DSF
		comma-separated-	is producing the output
		value string	in a specific new
			ordering.
			“No”: Indicates that no
			specific ordering is introduced.
			<csv-expr>: The
			comma-separated-value
			expression captures the
			output columns on
			which the output is
			ordered. It takes one of
			the following two
			forms:
			csv-columnNames
			(“Col1, Col2, . . .”): A
			list of column names.
			Constraints: Column-
			names format works
			only if the names are
			predefined, i.e., do not
			depend on the DSF's
			inputs or parameters.
			csv-columnPositions
			(“2, 4, . . .”): A list of
			column positions.
			Constraints: Column-
			positions format works
			only if the names are
			dynamic but the
			positions are fixed.
			If neither of the column
			names nor the column
			positions are
			applicable, then the
			value must be set to “No”.

TABLE 26
		Empty
		Content	Empty Content
Element Name	Presence	Allowed	Representation
“property”	Mandatory	No	N/A
“outputId”	Mandatory	No	N/A
“preserveInPartitioning”	Mandatory	No	N/A
“preserveInOrdering”	Mandatory	No	N/A
“introducePartitioning”	Mandatory	No	N/A
“introduceOrdering”	Mandatory	No	N/A

For DSF execution through QUERYGRID, the Interesting Ordering properties may not be useful because the current implementation of the QUERYGRID does not preserve the partitioning or ordering of the transferred data. However, the Interesting Ordering properties are still useful for the TERADATA local functions.

Miscellaneous Design and Setup Considerations

Input Referencing Mechanism (Position-Based vs. Alias-Based)

Each function descriptor is consistent in the referencing mechanism used for the inputs (the ON clauses). Throughout a given descriptor, the referencing is either position-based or alias-based.

Mixing the two referencing mechanisms in a single descriptor may result in missing optimization opportunities due to a possible mismatch between the descriptor instructions and a given query.

If there are DSFs that mandate alias when multiple ON clauses are used and do not mandate the alias when only one ON clause is used, the descriptor uses the alias-based referencing mechanism (the general case). And queries use alias even if only one ON clause is used.

Storage Limitations and Workarounds

Function descriptors are stored in a dictionary table, in TERADATA the UDFInfo table in a TERADATA database system. Since size is very critical in dictionary tables, the descriptor size is limited to roughly 3K (i.e., 3000) bytes. This size accommodates most descriptors without exceeding the limit.

In the cases where the limit is exceeded, some JSON segments from the descriptor may need to be removed. The list below gives some guidelines on which segments can be removed and the related impact:

Output Schema Specifications: This segment does not enable optimizations, but rather it is needed for contract-bypassed compilation. The instructions of the output schema can be removed from the descriptor, and the system should gracefully fall back to the contract function mechanism.

Input Schema Specifications: There are several scenarios in which providing the instructions for input schema specifications is neither effective nor critical, as discussed above. In these cases, the corresponding JSON segments can be removed.

Interesting Ordering Properties: The QUERYGRID does not guarantee the ordering sent to or generated from the coprocessor functions, as discussed above. Therefore, such JSON segment can be removed.

Uploading Function Descriptors to Database Dictionaries

The mechanism of uploading the function descriptor to the database dictionary table varies depending on the source of the function. If the function is local to the database system, then the descriptor may be uploaded as part of the CREATE FUNCTION DDL command. In contrast, if the function is remote (i.e., a DSF), then the descriptor may be uploaded as part of the installation procedure that installs the DSF into the database system, as discussed in connection with FIG. 4.

Using DSFs When the Output Schema Cannot Be Expressed Using the Markup Language

As mentioned above, the markup language is designed to cover the needs for the majority of the DSFs. There are some DSFs that have been observed, which are not covered by the markup language as described above, i.e., their properties cannot be expressed using the instructions described above.

In the cases where the output schema is the property that cannot be expressed using the markup language, then the contract function mechanism described above in connection with FIGS. 5A and 5B is mandatory, and it is automatically activated (as the fallback mechanism). In the cases where other properties, e.g., input schema, predicate push, etc., are the ones that cannot be expressed using the markup language, then the no-optimization mode takes precedence and the corresponding DSFs can still execute but without optimizations, as discussed above in connection with FIGS. 5A and 5B.

FIG. 8 shows the specifications of a Pivot DSF invocation. FIG. 9 shows the output schema specifications for the Pivot DSF. An example of a DSF in which the output schema cannot be expressed using the markup language is illustrated in FIGS. 8 and 9.

FIG. 10 shows an example Pivot DSF invocation and output schema. The output columns are: the list of columns in the “PartitionColumns” parameter followed by a number of columns, each of which is named “colName_i”, where “colName” is the value defined in the “TargetColumns” parameter, and “i” ranges from 0 to the value defined in the “NumberOfRows” parameter. If “TargetColumns” lists multiple columns, e.g., ‘X’ and ‘Y’, then the output will be [ . . . , X_0, X_1, . . . , Y_0, Y_1, . . . ].

Capturing such output schema uses looping capabilities and nested looping.

FIG. 11 shows the specifications of a PCAPlot DSF. FIG. 12 shows the output schema for the PCAPlot DSF. FIG. 13 shows an example of the invocation of PCAPlot DSF and the resulting output schema. The PCAPlot DSF uses the looping capability.

FIG. 14 shows the specifications of a NaiveBayesPredict DSF. FIG. 15 shows the output schema for the NaiveBayesPredict DSF. FIG. 16 shows an example of the invocation of NaiveBayesPredict DSF and the resulting output schema. The number of columns in the output schema and their names depend on the NaiveB ayes model passed to the DSF in the “Model” parameter, which cannot be captured by the markup language. In the NaiveBayesPredict DSF, the model values can be explicitly specified in parameter “Response”, and in this case, the “Model” parameter is not needed. Therefore, the output schema specifications for NaiveBayesPredict may still provide instructions on how to infer the property under the condition that the “Response” parameter exists (using CASE instruction). Otherwise, the function descriptor returns an empty output schema document.

LOOP Instruction

The LOOP instruction is used when looping or iterating over a set of instructions is needed, such as in the Pivot DSF, discussed in connection with FIGS. 8-10 and the PCAPlot DSF, discussed in connection with FIGS. 11-13. The structure of the LOOP instruction document, type LOOP-D1, is illustrated in FIG. 17.

The elements of document type LOOP-D1 are described in Tables 27 and 28.

TABLE 27
	Value
Element Name	Type	Value Domain	Description
“instruction”	String	Fixed value = “LOOP”	This property indicates
			that the document is a
			LOOP instruction.
“numIterations”	String	The values takes one of	This property specifies
		three possible forms:	the number of iterations
		“constant”,	to execute.
		“valueFromParameter(. . .)”,	“constant”: A constant
		“EnumFromParameter(. . .)”	positive integer value is
			specified to determine the
			number of iterations.
			“valueFromParameter(<parName>)”:
			The specified parameter must
			return a positive integer
			value, which determines
			the number of iterations.
			“EnumFromParameter(<parName>)”:
			The specified parameter must
			return a comma-
			separated list of values.
			The number of these
			values determines the
			number of iterations. The
			values within the list are
			treated as strings.
“startIndex”	String	The value is either “0” or	This property specifies
		“1”	the start value of the
			iteration index
			(iteration.index variable),
			which can be referenced
			in the arguments field.
“arguments”	Array	JSON documents of type	The entire array of
	of	ADD-D1 or CASE-D1	documents executes for
	documents		each iteration.
			The documents can
			reference two reserved
			keywords, namely
			“iteration.index” and
			“iteration.value”, e.g., to
			be concatenated with
			column names.
			“iteration.index”:
			Returns an integer
			representing the iteration
			number. This variable
			starts with the value
			specified in “startIndex”
			field, and gets
			incremented by 1 after
			each iteration.
			“iteration.value”: This
			variable is applicable
			only for
			“EnumFromParameter( )”,
			and NULL otherwise.
			With each iteration, it
			carries one of the values
			from the CSV list
			specified in
			“numIterations” field (in
			ordered manner).

	TABLE 28
			Empty
			Content	Empty Content
	Element Name	Presence	Allowed	Representation
	“instruction”	Mandatory	No	N/A
	“numIterations”	Mandatory	No	N/A
	“startIndex”	Mandatory	No	N/A
	“arguments”	Mandatory	No	N/A

Demonstration of the Enabled Optimizations

Projection Push Optimization

Projection push can significantly reduce the data transfer overhead if the base table has many columns while the DSF only needs few of these columns for its internal processing.

The following example uses a GLM (Generalized Linear Model) DSF to illustrate this optimization:

SELECT * FROM GLM@coprocessor (
         ON (SELECT * FROM admissions_train) as InputTable
         USING
         OutputTable (‘glm_admissions_model’)
         InputColumns (‘admitted’,‘masters', ‘id’, ‘gpa’,
         ‘programming’)
         CategoricalColumns (‘masters', ‘programming’)
         Family (‘LOGISTIC’)
) as glmtb;

The input table to the DSF, i.e., admissions_train in the example, has the following table schema:

admissions_train(id int,
masters char(100),
gpa float,
programming char(100),
admitted int,
fatcol char(10000));

The input schema property, which is relevant for projection push optimization, of the function descriptor document for the GLM DSF is shown below.

{“functionName”: “GLM”,
“inputSchema”: [
{ “property” : “inputSchema” ,
“inputId”: “1”,
“surplus”: “ignored”,
“nullElimination”: “No”,
“instructions”: [ { “instruction”: “ADD”,
“placement”: “end”,
“arguments”: [
{“source”: “parameter” ,
“name”: “InputColumns.*” ,
“dataType”: “unchanged”,
“manipulations”: [ ]
},
{“source”: “parameter” ,
“name”: “Weight.*” ,
“dataType”: “unchanged”,
“manipulations”: [ ]
}
]
}
]
}
],
...
}

As defined in the function descriptor, only the columns specified in the InputColumns and Weight parameters are required for the function execution. Therefore, the function only needs five columns, i.e., {id, masters, gpa, programming, admitted}. The last column in the table, i.e., “fatcol”, is not needed by the function. As a result, the query can be equivalently re-written as shown below, which eliminates “fatcol” from the InputColumns list, thereby reducing the amount of data required to be sent to the coprocessor.

SELECT * FROM GLM@coprocessor (
         ON (SELECT admitted, masters, id, gpa,
         programming FROM
         admissions_train) as InputTable
         USING
         OutputTable (‘glm_admissions_model’)
         InputColumns (‘admitted’,‘masters', ‘id’, ‘gpa’,
         ‘programming’)
         CategoricalColumns (‘masters', ‘programming’)
         Family (‘LOGISTIC’)
) as glmtb;

Predicate Push Optimization

The predicate push optimization can significantly reduce the data transfer overhead due to the elimination of unneeded records. Moreover, it can enable the generation of more efficient query plans by possibly leveraging available access paths. The following example query illustrates this optimization using the Sessionize DSF illustrated in FIG. 1 and repeated below.

SELECT *
FROM ( SELECT * FROM SESSIONIZE@coprocessor (
ON (SELECT * FROM web_click)
PARTITION BY userid
ORDER BY clicktime
USING
TimeColumn (‘clicktime’)
TimeOut (‘60’)
)AS output) AS q
WHERE userid = 0;

A possible function descriptor document for the Sessionize DSF is shown below. Only the predicate push property, which is relevant for predicate push optimization, is shown.

{“functionName”: “SESSIONIZE”,
“predicatePush”: [{“property” : “predicatePush” ,
“outputId”: “standard”,
“targetInputId”: “1”,
“rowIndependence”: “No”,
“partitionIndependence”: “Yes”
}],
...
}

Since the partitionIndependence flag is set to “Yes”, rows across different partitions are processed independently by the Sessionize function. As the user is interested only in rows in partition userid=0, the post-function predicate (in our example “userid=0”) can be applied on the function's input and only the required rows are sent to the remote coprocessor. As a result, the query can be equivalently re-written to the version shown below.

SELECT *
FROM ( SELECT * FROM SESSIONIZE@coprocessor (
ON (SELECT * FROM web_click WHERE userid = 0)
PARTITION BY userid
ORDER BY clicktime
USING
TimeColumn (‘clicktime’)
TimeOut (‘60’)
)as output) as q
;

Further examples consistent with the present teaching are set out in the following numbered clauses.

Clause 1. A method comprising:

a database system receiving a request from a user, wherein the request invokes a data set function (DSF) and uses a property to be provided by the DSF;

the database system determining that a function descriptor is available for the DSF, wherein the function descriptor is expressed as markup language instructions, and wherein the function descriptor defines the property of the DSF; and

the database system using the function descriptor to define a property for the DSF.

Clause 2. The method of clause 1 further comprising:

a developer creating the DSF to execute on a remote system;

the developer writing the descriptor for the DSF; and

the database system receiving and storing the descriptor for the DSF.

Clause 3. The method of clause 1 further comprising:

a developer creating the DSF to execute on the database system;

the developer writing the descriptor for the DSF; and

the database system receiving and storing the descriptor for the DSF.

Clause 4. The method of any of clauses 1-3, wherein the markup language is an instruction-based language.

Clause 5. The method of any of clauses 1-4, further comprising using the function descriptor to define an output schema for the DSF.

Clause 6. The method of any of clauses 1-5, further comprising using the function descriptor to define an input schema for the DSF.

Clause 7. The method of any of clauses 1-6, further comprising using the function descriptor to determine to push a predicate in the request from the DSF's output to the input of the DSF.

Clause 8. The method of any of clauses 1-7, further comprising using the function descriptor to determine to push a projection in the request from the input of the DSF to the output of the DSF.

Clause 9. The method of any of clauses 1-8, further comprising using the function descriptor to estimate a cardinality of the property.

Clause 10. The method of any of clauses 1-3, further comprising using the function descriptor to determine if the DSF inherits or obeys specific ordering or partitioning schemes.

Clause 11. A non-transitory computer-readable tangible medium, on which is recorded a computer program, the computer program comprising executable instructions, that, when executed, perform a method comprising:

a database system receiving a request from a user, wherein the request invokes a data set function (DSF) and uses a property to be provided by the DSF;

the database system determining that a function descriptor is available for the DSF, wherein the function descriptor is expressed as markup language instructions, and wherein the function descriptor defines the property of the DSF; and

the database system using the function descriptor to define a property for the DSF.

Clause 12. The computer program of clause 11 wherein the method further comprises:

a developer creating the DSF to execute on a remote system;

the developer writing the descriptor for the DSF; and

the database system receiving and storing the descriptor for the DSF.

Clause 13. The computer program of clause 11 wherein the method further comprises:

a developer creating the DSF to execute on the database system;

the developer writing the descriptor for the DSF; and

the database system receiving and storing the descriptor for the DSF.

Clause 14. The computer program of any of clauses 11-13, wherein the method further comprises using the function descriptor to define an output schema for the DSF.

Clause 15. The computer program of clauses 11-14, wherein the method further comprises using the function descriptor to define an input schema for the DSF.

Clause 16. The computer program of clauses 11-15, wherein the method further comprises using the function descriptor to determine to push a predicate in the request from the DSF's output to the input of the DSF.

Clause 17. The computer program of clauses 11-16, wherein the method further comprises using the function descriptor to determine to push a projection in the request from the input of the DSF to the output of the DSF.

Clause 18. The computer program of clauses 11-17, wherein the method further comprises using the function descriptor to estimate a cardinality of the property.

Clause 19. The computer program of clauses 11-18, wherein the method further comprises to using the function descriptor to determine if the DSF inherits or obeys specific ordering or partitioning schemes.

Clause 20. A method comprising:

a database system receiving a request from a user, wherein the request invokes a data set function (DSF) and uses a property to be provided by the DSF;

the database system determining that a function descriptor is not available for the DSF;

the database system determining that a contract function is available for the DSF; and

the database system using the contract function to optimize the request.

Note that, while the above description is directed to a few specific uses for the techniques described herein (such as predicate push/pull), it will be understood that there are numerous other applications for the techniques described herein, such as in join planning using not just cardinality but also historical estimates or in join order.

The operations of the flow diagrams are described with references to the systems/apparatus shown in the block diagrams. However, it should be understood that the operations of the flow diagrams could be performed by embodiments of systems and apparatus other than those discussed with reference to the block diagrams, and embodiments discussed with reference to the systems/apparatus could perform operations different than those discussed with reference to the flow diagrams.

The word “coupled” herein means a direct connection or an indirect connection.

The text above describes one or more specific embodiments of a broader invention. The invention also is carried out in a variety of alternate embodiments and thus is not limited to those described here. The foregoing description of an embodiment of the invention has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise form disclosed. Many modifications and variations are possible in light of the above teaching. It is intended that the scope of the invention be limited not by this detailed description, but rather by the claims appended hereto.

Claims

1. A method comprising: a database system receiving a request from a user, wherein the request invokes a data set function (DSF) and uses a property to be provided by the DSF;the database system determining that a function descriptor is available for the DSF, wherein the function descriptor is expressed as markup language instructions, and wherein the function descriptor defines the property of the DSF; andthe database system using the function descriptor to define a property for the DSF.

2. The method of claim 1 further comprising: a developer creating the DSF to execute on a remote system;the developer writing the descriptor for the DSF; andthe database system receiving and storing the descriptor for the DSF.

3. The method of claim 1 further comprising: a developer creating the DSF to execute on the database system;the developer writing the descriptor for the DSF; andthe database system receiving and storing the descriptor for the DSF.

4. The method of claim 1, wherein the markup language is an instruction-based language.

5. The method of claim 1, further comprising using the function descriptor to define an output schema for the DSF.

6. The method of claim 1, further comprising using the function descriptor to define an input schema for the DSF.

7. The method of claim 1, further comprising using the function descriptor to determine to push a predicate in the request from the DSF's output to the input of the DSF.

8. The method of claim 1, further comprising using the function descriptor to determine to push a projection in the request from the input of the DSF to the output of the DSF.

9. The method of claim 1, further comprising using the function descriptor to estimate a cardinality of the property.

10. The method of claim 1, further comprising using the function descriptor to determine if the DSF inherits or obeys specific ordering or partitioning schemes.

11. A non-transitory computer-readable tangible medium, on which is recorded a computer program, the computer program comprising executable instructions, that, when executed, perform a method comprising: a database system receiving a request from a user, wherein the request invokes a data set function (DSF) and uses a property to be provided by the DSF;the database system determining that a function descriptor is available for the DSF, wherein the function descriptor is expressed as markup language instructions, and wherein the function descriptor defines the property of the DSF; andthe database system using the function descriptor to define a property for the DSF.

12. The computer program of claim 11 wherein the method further comprises: a developer creating the DSF to execute on a remote system;the developer writing the descriptor for the DSF; andthe database system receiving and storing the descriptor for the DSF.

13. The computer program of claim 11 wherein the method further comprises: a developer creating the DSF to execute on the database system;the developer writing the descriptor for the DSF; andthe database system receiving and storing the descriptor for the DSF.

14. The computer program of claim 11, wherein the method further comprises using the function descriptor to define an output schema for the DSF.

15. The computer program of claim 11, wherein the method further comprises using the function descriptor to define an input schema for the DSF.

16. The computer program of claim 11, wherein the method further comprises using the function descriptor to determine to push a predicate in the request from the DSF's output to the input of the DSF.

17. The computer program of claim 11, wherein the method further comprises using the function descriptor to determine to push a projection in the request from the input of the DSF to the output of the DSF.

18. The computer program of claim 11, wherein the method further comprises using the function descriptor to estimate a cardinality of the property.

19. The computer program of claim 11, wherein the method further comprises using the function descriptor to determine if the DSF inherits or obeys specific ordering or partitioning schemes.

20. A method comprising: a database system receiving a request from a user, wherein the request invokes a data set function (DSF) and uses a property to be provided by the DSF;the database system determining that a function descriptor is not available for the DSF;the database system determining that a contract function is available for the DSF; andthe database system using the contract function to optimize the request.