Skip to content
This repository has been archived by the owner on Sep 3, 2024. It is now read-only.

Latest commit

 

History

History
137 lines (123 loc) · 9.71 KB

README.md

File metadata and controls

137 lines (123 loc) · 9.71 KB
Raw

SDK generator PoC

PoC of generating full object implementation based on object definition.

Description

There is an example file ready for generation database_role_def.go which creates files:

How it works

Creating object generation definition

To create definition for object generation:

  1. Create file object_name_def.go (like example database_role_def.go file).
  2. Put go generate directive at the top: //go:generate go run ../main.go. Remember that you may have to change the path to main.go file.
  3. Create object interface definition.
  4. Add key-value entry to definitionMapping in main.go:
  5. You are all set to run generation.
Invoking generation

To invoke example generation (with first cleaning all the generated files) run:

make clean-generator-poc run-generator-poc

To invoke generation inside SDK package (with cleaning), e.g. for session_policies run (mind the _(underscore)):

make clean-generator-session_policies run-generator-session_policies

Next steps

Essentials
  • generate each branch of alter in tests (instead of basic and all options)
  • clean up predefined operations in generator (now casting to string)
  • (?) add mapping from db to plain struct (for now only "// TODO: Mapping" comment is generated)
  • add arguments to the generator, so we'll be able to specify which files should be generated / re-generated, because after we fill things that need our input we don't want to re-generate those files and override the changes, also adding small changes is very challenging, e.g. for new validation rule you have to re-generate unit-tests to get one new function, revert to old tests (the one with filled tests), copy new test case (of course we could add that one by hand but if we add one case, or modify more cases this becomes more challenging)
  • add support for Enums
  • generate ShowID function with 3 implementation variations (the last one is the rarest one and can be postponed)
    • use Show function with Like
    • use Show without any options and filter with Go for + if
    • in some cases we could need more filters -> see alerts.go (but we can implement it later)
  • handle arrays
  • handle more validation types
  • write new valueSet function (see validations.go) that will have better defaults or more parameters that will determine checking behaviour which should get rid of edge cases that may cause bugs in the future
    • right now, we have valueSet function that doesn't take into consideration edge cases, e.g. with slice where sometimes we would like to do something like alter x set y = () (set empty array to unset y). Those edge cases have cause on our validation, and it determines sometimes if we'll return an error or not, which can lead to bugs!
  • refactor generation of Describe, so it will tak context and request as arguments
    • all the interface functions should have context and request as arguments for the sake of API consistency and generation simplicity
  • split templates into multiple templates (e.g. ImplementationTemplate) to improve readability
  • check if SelfIdentifier implementation is correct (mostly type, because it's derived from interface obj) by checking if there's a resource with different types of identifiers across queries (e.g. Create , Alter )
  • we should specify prefix / postfix standard for top-level items in _def.go files to avoid any conflicts in the package
  • remove name argument from QueryStruct in the Operation, because Opt structs in the Operation will have name from op name + interface field and not query struct itself
  • Derive field name from QueryStruct, e.g. see network_policies_def where we can remove "Set" field, but we have to make a convention of creating nested struct with name pattern like e.g. NetworkPoliciesSet or NetworkPolicySet, then we could automatically remove prefix and we'll name field with postfix, so "Set" in this case
  • Add more operations (every operation ?) in the database_role_def.go example
  • Divide into packages or add common prefix for similar files (e.g. struct_plain.go, struct_db.go or builders_keyword.go, builders_parameter.go)
  • Make a clear division between DSL files and model files (etc. QueryStruct(DSL) and Field(Model)) and divide them into separate packages (?)
  • Simplify ImplementationTemplate (templates.go) and separate into multiple templates variables / definitions
  • Add parameter to DtoTemplate (templates.go) to generate the right path to the dto generator's main.go file
  • Right now to avoid generated structs duplication, arrays containing struct names have been introduced (template_executors.go), find a better solution to solve the issue (add more logic to the templates ?)
Improvements
  • automatic names of nested structs (e.g. DatabaseRoleRename)
  • check if generating with package name + invoking format removes unnecessary qualifier
  • consider merging templates StructTemplate and OptionsTemplate (requires moving Doc to Field)
  • expand unit tests generation
  • experiment with Snowflake table (any table) representation in Go in order to implement DbStruct -> PlainStruct convert function
    • see if *string can have similar effect as sql.NullString (check go-snowflake connector ?)
      • if yes, then we should be using pointers instead of abstractions like sql.NullString and we can modify ShowMapping and DescribeMapping to generate convert function with automatic conversion (as we have in DTOs). warehouses.go is a good place to start with when planning mapping strategy, because there's a lot of different mapping cases.
  • when calling .SelfIdentifier we can implicitly also add validateObjectIdentifier validation rule
  • enforce user to use KindOf... functions with interface
  • queryStruct should be spilled into Operation interface file, because the idea was to have model which is unaware of DSL used to create it.
  • generate full tests for common types (e.g. setting/unsetting tags)
  • generate common resources for integration tests
  • cleanup the design of builders in DSL (e.g. why transformer has to be always added?)
  • generate getters for requests, at least for identifier/name
  • generate integration tests in child package (because now we keep them in testint package)
  • struct_to_builder is not supporting templated-like values. See stages_def.go where in SQL there could be value, where 'n' can be replaced with any number
    • SKIP_FILE_n - this looks more like keyword without a space between SQL prefix and int
    • SKIP_FILE_n% (e.g. SKIP_FILE_123%) - this is more template-like behaviour, notice that 'n' is inside the value (we cannot reproduce that right now with struct_to_builder capabilities)
  • fix builder generation
    • we can add flatten option in cases where some sql structs had to be nested to create correct sql representation
      • for example encryption options in stages_def.go (instead of calling .WithEncryption(NewEncryptionRequest(encryption)) we could call .WithEncryption(encryption))
    • operation names (or their sql struct names) should dictate more how constructors are made
  • better handling of list of strings/identifiers
    • there should be no need to define custom types every time
    • more clear definition of lists that can be empty vs cannot be empty
  • add empty ids in generated tests (TODO in random_test.go)
Known issues
  • generating two converts when Show and Desc use the same data structure
  • wrong generated validations for validIdentifierIfSet for cases like
A := QueryStruct("A").
	Identifier("Name").
	Validation(ValidIdentifierIfSet, "Name")
B := QueryStruct("B")
    .ListQueryStructField(A) // A []A - validations will be wrong because this is array
  • cannot re-generate when client.go is using generated interface
  • spaces in templates (especially nested validations)
  • request mapping fails (.toOpts()) when nested object is not optional (pointer) e.g.
type NestedReq struct {
}

type SomeReq struct {
    NestedReq NestedReq // Not a pointer and in toOpts right now we're always do a check if req.NestedReq != nil which is not correct for non pointer type
}
Known limitations
  • automatic array conversion is not recursive, so we're only supporting one level mapping
    • []Request1{ foo Request2, bar int } won't be converted, but []Request1{ foo string, bar int } will