Architecture
Description of the basic architecture
At this point the core functionality and the corresponding architecture of openVALIDATION is described. This description does not include the CLI or REST component. A natural language rule and the corresponding schema are expected as input parameters. Afterwards the processing starts, which can be separated into 5 subroutines (preprocessor, schema converter, parser, validation, generator). At the end of the compilation process program code is generated.
Simplified view of the entire compilation process
The preprocessor prepares the natural language rule. At this point, for example, a translation or normalization of the keywords (aliases) takes place. The parser generates the Abstract Syntax Tree (AST). The AST is the logical structure of the grammar, which in our case represents the domain of the validation rules. The AST is then processed further with the help of the generator, so that valid program code is generated at the end of the entire processing procedure.
openVALIDATION makes it possible to describe validation rules in a natural language. Strictly speaking, there are many different languages such as German or English. The languages are extensible, so you can add more languages like Spanish or French with a manageable effort.
In order for openVALIDATION to extract a valid validation rule from a raw text the special keywords must be included in this text. There is a set of fixed keywords. Each of these keywords has one or more aliases in the respective language. For example, the keyword should not has additional aliases such as: must not or have not . Each keyword has a normalized value for which several aliases exist in each natural language. The configuration of these aliases can be found in the resources folder of the openvalidation-core project.
...
ʬoperatorʬequals = IS, EQUAL, EQUALS, IS EQUAL TO
ʬoperatorʬnot_equals = ISN'T, IS NOT, NOT EQUAL, NOT EQUALS, NOT
...
These aliases will be resolved or normalized in the preprocessor.
Any text can be placed around the keywords. We call it semantic sugar. Thanks to this semantic sugar, the creator can make his rules very flexible and describe the set of rules in much more detail. Other editors can better understand the rules while the machine simply ignores the semantic sugar. This flexibility is the core of the grammar of openVALIDATION. It makes it a lot more natural compared to other DSL's.
For example, it would be perfectly sufficient to describe a rule in the following way:
age should not less 18
However, apart from its rather syntactical appearance, it is not so nice for humans to read and may present some questions. Whose age does the rule address? 18 what? For a human to effortlessly understand such a rule, a little more information about its context is required:
applicant's age should not be less than 18 years
This rule is semantically identical to the previous one but sounds a lot more natural. It clarifies the context and is much easier to read and understand. The larger and more complex the set of rules becomes, the more useful the ability to formulate rules in such a way proves to be. Another advantage arises when you dictate the rule instead of writing it. Thus, voice assistants can be equipped with the ability to generate code or to store rules.
A validation rule always requires a data model or schema so that it knows which data is to be validated. Usually a data model already exists before you start to create validation rules. Therefore openVALIDATION compiler expects the corresponding schema in addition to the actual set of rules.
It is not the task of openVALIDATION to create a data model or extract a schema from the given rule set. The schema must be passed to openVALIDATION as an input parameter.
It may be possible that this functionality will be a part of openVALIDATION in the future.
The schema must be currently defined in JSON Schema format. To simplify the use of openVALIDATION, the schema can also be defined in JSON Object Format. From this a possible JSON schema is then automatically derived which is then further processed.
Here is an example of a possible schema:
{
"$schema": "http://json-schema.org/draft-07/schema#",
"type": "object",
"properties": {
"age": {
"type": "number",
}
}
}
The schema defines a single attribute called
age whose type is number. Although the schema is fairly simple the corresponding JSON schema file has a certain complexity to it. Especially if the schema becomes more complicated the amount of complexity skyrockets. However, for such straightforward schemas a simple JSON Object format can be specified from which the above schema is then derived. This is useful if you quickly want to try something out without having such unnecessary complexity stand in your way. The derivation of the following JSON Object creates the above schema:{
age : 0
}
This is a simple way to pass the same information to the compiler. The only attribute
age contains a numerical value of 0. You can also pass other numerical values, since it will be ignored and is only used to determine the data type of the attribute.If you want to realize a deeper integration of openVALIDATION into other systems it is always recommended to use JSON Schema instead of JSON Object. This allows a schema to be specified much more precisely.
openVALIDATION needs the schema to parse the natural language rule. All attributes of the schema are therefore loaded into the DataSchema component. To do this, the schema must be read and converted into the appropriate format. Since the schema itself can be specified in various formats, such as JSON Schema or JSON Object and later also in YAML or XSD, there is a SchemaConverterFactory that provides the corresponding implementation convert the schema depending on its type.
The specified schema is loaded into the DataSchema component
Each converter must implement the ISchemaConverter interface and the DataSchema convert() method. The task of the converter is to run through the hierarchy of the schema and convert each attribute with all relevant meta informations, such as name, type, path/full name, and so on, into a flat structure that is the DataSchema.
Here is an example. The following schema is given:
{
age:0,
address : {
city : ""
}
}
After the conversion the DataSchema contains the following information (pseudo YAML format):
_properties:
DataProperty:
name : "age"
path : ""
fullName : "age"
type : "Decimal"
DataProperty:
name : "address"
path : ""
fullName : "address"
type : "Object"
DataProperty:
name : "city"
path : "address"
fullName : "address.city"
type : "String"
Thanks to this metadata, all relevant information can be extracted from a natural language rule. Among other things, this construct allows recursive name resolution of schema attributes. For example, you could directly use the attribute
city without specifying the full name address.city like in a formal programming language. If city occurs more than once in the schema, it will be validated later and the user will get a compiler message that he has to use the full name because of ambiguity.DataSchema
The DataSchema component stores not only the schema information but also the variables and semantic operators. However, this information is added at a later time after the first parsing routine.
By keeping the type information available during parsing, especially the operands of a condition can be determined and extracted. For example, if you know that the left operand is an attribute from the schema of type decimal and the right operand simply contains text, you can try to extract a numeric value from it. If no numeric value is found, you can then throw a relatively unique compiler message.
Before a natural language rule can be parsed, the corresponding text must first be prepared by the preprocessor. For example, the keywords or the paragraphs must be normalized. Includes must be resolved or it must be checked if there are collisions with different keywords and so on.
The following image illustrates some of the many small routines the preprocessor consists of. This processing logic has also been modularized so there are many small preprocessors that perform the individual tasks one after the other.
The internal design of the preprocessor
Every single step of the whole preprocessor routine is implemented in a separate module. Each of these modules must be derived from the abstract base class PreProcessorStepBase. Afterwards the method String process(String rule) must be overwritten and provided with the respective logic. Currently, the following modules can be found in the package io.openvalidation.core.preprocessing.steps:
Further modules can be added without problems. However, it is very important to keep the order in which the individual modules are executed.
The resolution of keywords is one of the most important tasks of the preprocessor. It involves converting the large number of aliases into a normalized form. The following parsing process only knows the normalized keywords and not their aliases.
Assume the following rule enters the preprocessor:
applicant's age should not be less than 18 years
After the preprocessing has completed the rule will look like this:
the applicant's age
ʬconstraintʬmustnotʬshould_20_not be ʬoperatorʬless_20_thanʬless
than 18 years
The keyword should not became ʬconstraintʬmustnotʬshould_20_not where the last part ...should_20_not is dynamic. It is used to preserve the original text so that it is still available after the parsing process, for example, for compiler messages or for other purposes. The front part ʬconstraintʬmustnotʬ is normalized. This is the part the parser knows.
The resolution of the aliases depends on the culture code used. When using e.g. de, only aliases from the German resource file aliases_de.properties are used.
Parsing is probably the most important and complex processing step in openVALIDATION. One of the reasons for the particular complexity is the flexibility of the grammar that is associated with natural languages. In formal programming languages but also in most DSL's and especially in the strongly typed object-oriented programming languages every word, every character has a very precise meaning. The input must be very exact otherwise there is an immediate compiler error.
While designing the grammar of openVALIDATION it was very important to us to avoid such restrictions as they make the understanding and usage for normal users (non-developers) difficult. Therefore openVALIDATION allows many variants of how to express a rule in a natural way without bothering the user with a compiler message at every little deviation.
if user's age is less than 18 years then underage persons are not admitted
or
the applicant's age must not be less than 18 years
Both rules are semantically identical and are translated by the compiler into the following code. Here in pseudo code:
if (age < 18)
throw Error("underage persons..." or in the 2nd case "the applicant's..")
The compiler first tries to accept the input as it suits the user best, then extracts the relevant part without forcing the user to follow a certain notation. This makes the learning curve for newbies extremely steep. You only need to get a rough idea of the structure of a validation rule, without having to pay attention to every character or every special keyword.
openVALIDATION requires relatively little abstraction ability on the part of the user. The compiler handles this task by itself and thus relieves the user. This is the core of the philosophy behind openVALIDATION:
"Instead of forcing humans to understand the complex inner workings of machines, we should construct machines in a way, so they better understand us humans"
To make this convenience available to the user, the complexity must be shifted into the processing logic of the compiler. Especially the parser does most of the work at this point. It takes care that a complex hierarchical Abstract Syntax Tree (AST) is created from a single string.
The parser takes a normalized rule and creates an AST object
To make this possible the parser itself consists of 3 components:
- 1.ANTLR Grammar(Lexer, Parser, Generator...)
- 2.Parse Tree Transformer
- 3.Post Processors
Each of these sub-components has a specific task and therefore involves a certain complexity. The entire logic for parsing the validation rules is contained in the package io.openvalidation.antlr.
The ANTLR Parser takes care of the initial step of parsing. At this point, a generic AST is created from a text. This is called parse tree in ANTLR. At this point, the text is transferred to a temporary object tree.
The core of the ANTL parsing routine is the corresponding ANTLR grammar of openVALIDATION. It is located in the file main.g4 in the resources folder of the openvalidation-antlr module. Here is a small excerpt of it:
grammar main;
main : PARAGRAPH? (rule_definition|rule_constrained|variable|semantic_operator|comment|unknown)
(PARAGRAPH (rule_definition|rule_constrained|variable|semantic_operator|comment|unknown))*
PARAGRAPH? unknown? EOF;
comment : STRING? COMMENT unknown?;
variable : (lambda|expression)? AS name?;
semantic_operator : unknown? OPERATOR_COMP? AS_OPERATOR name?;
rule_definition : IF? expression? THEN action?
| IF expression? THEN? action?;
...
In this grammar the so-called lexer tokens and the parser rules are defined. While lexer tokens represent the individual keywords, parser rules represent complete signatures of words or the tokens.
Here is an example for the signature of a comment:
comment : STRING? COMMENT unknown?;
unknown : (STRING | LPAREN | RPAREN | WITH_ERROR | COMBINATOR | OPERATOR_ARITH | WITH | OF | CONSTRAINT | IF | THEN | OPERATOR_COMP | AS | COMMENT | FUNCTION | FROM | ORDERED_BY)+;
COMMENT : 'ʬcommentʬ'[a-zA-Z0-9_]+;
STRING : ~('ʬ')+;
A comment consists of an optional lexer token called STRING followed by the lexer token COMMENT and concludes with an optional parser rule unknown. As soon as the string ʬcommentʬxxx appears somewhere in a paragraph, then this parser rule automatically takes effect.
Using this grammar, Java code is generated at design time using antlr4-maven-plugin's Java code. This code is then placed in the target directory target/generated-sources/java/io/openvalidation/antlr/generated and is automatically integrated into the main project.
This code is then aggregated in the class ANTLRExecutor. There the MainASTBuildListener is also included, which is the entry point for the following parse tree transformation logic.
The package io.openvalidation.antlr.transformation contains the transformation logic of the parser. This logic mainly takes care that the generic parse tree (its class name being ParseTree) generated by the ANTLR parser at runtime is transformed into the domain-specific AST model. Like many openVALIDATION components, the transformation logic is quite modular. Each single module takes care of a certain logical area. For example, it is the task of the PTCommentTransformer to transform the corresponding subtree of the parse tree, namely the mainParser.CommentContext into the ASTComment. The structure of these transformers is hierarchical and follows the logical hierarchy of the parse tree, which in turn results from the grammar (main.g4).
Each individual transformer is derived from the abstract base class TransformerBase and must implement the method transform(). There the actual transformation between the two data models takes place. The base method ASTItem createASTItem(ParseTree) can be used to transform the child elements. By calling this method with the corresponding parse tree, the transformation runs hierarchically from top to bottom. The TransformerFactory is used to load the corresponding transformer module depending on the type of the ParseTree.
After each transformation the corresponding post processor is called.
In the transformation step, the AST elements and the entire tree are created only very roughly. It requires many further processing steps to transform the AST into its final form. The post processors take over this task. These are also modular and follow the AST hierarchy. Depending on the type, the individual modules are called only after a certain transformation step has been completed. The naming convention of the post processors ensures that the execution levels can already be recognized by their names. For example, a PostConditionImplicitBoolOperand is called directly after the transformation of a condition, while a PostModelNumbersResolver is called only after the transformation of the ASTModel, i.e. at the very end of the entire transformation.
What exactly happens in the post processors? Let's take a look at the already implemented PostConditionImplicitBoolOperand:
public class PostConditionImplicitBoolOperand extends PostProcessorSelfBase<ASTCondition> {
@Override
protected Predicate<ASTCondition> getFilter() {
return c ->
(c != null
&& !c.hasRightOperand()
&& c.hasLeftOperand()
&& c.hasEqualityComparer()
&& c.getLeftOperand().isPropertyOrVariable()
&& c.getLeftOperand().isBoolean());
}
@Override
protected void processItem(ASTCondition condition) {
if (!condition.hasRightOperand()) {
ASTOperandStatic staticBool = new ASTOperandStatic("true");
staticBool.setDataType(DataPropertyType.Boolean);
staticBool.setSource("");
condition.setRightOperand(staticBool);
}
}
}
The task of this post processor is to complete an incompletely transformed ASTCondition. Such a condition always consists of a left and a right operand. In between there is the comparison operator. The method getFilter() searches for certain conditions, namely for those that have only one boolean operand and a comparison operator that is either Equals or Not Equals. Only if the condition of the method getFilter() is fulfilled, the method processItem() is called. There the already existing but still incomplete ASTCondition will be completed with an operand. Since it is an implicit boolean comparison, only a static true is added as the right operand.
This makes the following rules possible:
the contract must be signed
And here is the corresponding schema:
{signed: true}
These and many other post processors form a very flexible framework, which makes it possible to manipulate the AST at specific points.
Different states of a rule during the parsing process
The Abstract Syntax Tree is the central component of the openVALIDATION compiler. The AST is nothing else than the domain model of openVALIDATION. This domain model represents a logical structure of a validation rule.
Schematic representation of an AST object and the rule contained inside it
In the image you can see that e.g. a rule contains a condition and an error message. The condition has a left operand, a right operand and a comparison operator. This is a very simple example, which only demonstrates the schematic and logical structure of the AST. Usually, the AST is much more complex. For example, a rule can contain many conditions, which are linked with a logical operator AND or OR. There are also nested conditions or condition groups. There are variables, which in turn can contain other conditions, and so on.
The complete ASTModel consists of many individual classes which together form a logical hierarchy. The ASTModel is the aggregate root of this central domain model.
The package io.openvalidation.common.ast contains the ASTModel and all its components
The AST can easily be extended. For this purpose, each element must be derived at least from the class ASTItem. Depending on the position of the extension within the structure, a corresponding base class must be used.
Parsing is primarily about extracting all relevant information from a continuous text and transferring it into an object structure, i.e. into the AST. The AST can then be processed further. Eventually the AST will be validated and if everything is OK, code will be generated from the it.
After the AST has been successfully generated, it has to be validated before program code is generated in the next step with the help of the generator. This validation step ensures that the AST is consistent and valid.
The processing step Validation checks the AST and generates corresponding compiler messages
Here is an example of a possible inconsistency. The schema looks like this:
{
name : ""
}
and here is the rule:
applicant's age should not be less than 18 years
The
age attribute is used in these rules, but the schema only contains the attribute name. The age attribute cannot or must not be found. In this case the openVALIDATION compiler must not generate any program code. At this point, the compiler must throw a corresponding error message: e.g. "The rule must contain at least one attribute from the schema" or something like that. In order for such an error message to be thrown, the AST must be checked accordingly.Such and many other checks are performed in the processing step "Validation". These check mechanisms are implemented in a modular way and are located in the openvalidation-core module.
The package io.openvalidation.core.validation contains the corresponding validators
The ValidatorFactory creates a new instance of the corresponding validator for each individual element of the AST depending on its type. Thus, each validator takes care of a certain area of the AST.
Each single validator is derived from the base class ValidatorBase and must overwrite the inherited method void validate(). If an inconsistency is found, a new exception of type ASTValidationException is thrown immediately. This ensures that no unnecessary subsequent errors are caught.
It is one of the biggest challenges in developing a compiler to produce clean and meaningful compiler messages. These error messages should help the user to correct his input. Among other things, it is important not to return too much and not too few error information. It is often the case that the actual error occurred 10 processing steps back and only its site effects were found.
The exception handling or the generation of meaningful compiler errors in openVALIDATION is still in a very early stage. It is very likely that the architecture will change at one point.
Suggestions for improvement and further discussions on this topic are explicitly welcome!
This is the final processing step. At this stage program code is generated from the AST. To be more precise, an OpenValidationResult is generated, which contains compiler error messages as well as various metadata in addition to the actual code.
The core of the code generation are the generator templates and the framework Handlebars or their java implementation.
Program code is generated from the AST using Handlebars Templates
openVALIDATION is a multilingual cross compiler. It means that both input and output can be done in different languages. Therefore there are several generator templates for different programming languages. Here is the list of the currently supported languages, but not all of them are fully implemented yet:
- Java
- JavaScript/node
- C#
- Python
- Rust
With more to come.
Everything that is necessary to generate is located in the module openvalidation-generation. The corresponding generator templates are located in the resources folder:
Each supported programming language has its own folder with the corresponding name. Generator templates specific to Javascript are located in the folder javascript, the ones for Java in java, and so on. Cross-language templates are located in the folder common.
The templates themselves are named by the type of AST element and are used accordingly during generation. There is also a fallback mechanism that checks whether a template with the corresponding name exists in the language-specific folder. If not, the template is loaded from the common folder. This fallback mechanism enables the consistent reuse of templates for different output languages. Thus, you need much less language-specific generator templates and can implement new programming languages with much less effort.
Besides the actual rules, the generator creates its own specific framework and another code artifact called ValidatorFactory. The framework and the factory exist for each output language. They later facilitate the integration of the generated code into other systems.
Above all, the HUMLFramework or the corresponding framework.hbs is full of specific logic that requires special quality assurance. In order to be able to test it properly, the framework.hbs templates have been outsourced to separate projects within our own openvalidation-framework-tests repositories. There the respective implementations can be tested in their own technology stacks, each with their own unit tests. Afterwards one has to copy the code from the respective file HUMLFramework.xxx into the corresponding directory within the generator as framework.hbs. This is currently a manual step, which is required at this point. In the future, this should also be automated or simplified.
At the end of the entire compilation process, depending on the chosen output language, program code is generated in Java, JavaScript, C# or in one of the other supported programming languages. The code itself is divided into 3 categories:
- Implementation
- Framework
- ValidatorFactory
The implementation code contains the actual program logic of the respective rule set. It is the part of the code that has been translated from a natural language into the desired programming language. And this is what the code looks like:
var HUMLValidator = function() {
var huml = new HUMLFramework();
//rule: applicant's age should not be less than 18 years
huml.appendRule("",
["age"],
"applicant's age should not be less than 18 years",
function(model) { return huml.LESS_THAN(model.age, 18.0);},
false
);
this.validate = function(model){
return huml.validate(model);
}
}
The example above shows the generated JavaScript code. The code consists of a basic structure which, in the case of JavaScript, is mapped by a function called HUMLValidator(default name) to the actual rule set defined by huml.appendRule and by the function validate.
The implementation uses its own framework called HUMLFramework. This framework contains all the necessary basic functions for comparing different values. There are two important reasons why the HUMLFramework is necessary:
- 1.The generated code does not have any dependencies to the 3'rd party libraries. It includes everything necessary and thus makes the integration much easier.
- 2.The framework serves as a kind of normalization layer for cross-language code generation. The structure of the rules behind huml.appendRule looks similar in different programming languages. This allows us to reuse code generation templates when generating code in different programming languages - see Generator.
It is possible to generate several rule sets simultaneously. In this case there are several validators, each with its own rules. With the help of the ValidatorFactory you can access the specific rule set using the unique ID:
var openValidatorFactory = function(){
var _validators = {
Validator1 : new Validator1(),
Validator2 : new Validator2(),
Validator3 : new Validator3(),
};
this.create = function(validatorID){
if (_validators != null){
return _validators[validatorID];
}
return null;
}
}
And here is how to use the ValidatorFactory:
//usage:
var factory = openValidatorFactory();
validator = factory.create("Validator1");
validator.validate({age:18})
Code files can be generated automatically by setting the appropriate parameters. Furthermore, it is possible to create the entire code in one file each. Depending on the programming language, e.g. Java, inner classes will be generated.
Last modified 3yr ago