Crates.io | typify-impl |
lib.rs | typify-impl |
version | 0.2.0 |
source | src |
created_at | 2021-11-08 05:39:32.246031 |
updated_at | 2024-09-26 20:14:05.747145 |
description | typify backend implementation |
homepage | |
repository | https://github.com/oxidecomputer/typify |
max_upload_size | |
id | 478332 |
size | 7,838,460 |
Typify compiles JSON Schema documents into Rust types. It can be used in one of several ways:
using the cargo typify
command
via the macro import_types!("types.json")
to generate Rust types directly in
your program
via a builder interface to generate Rust types in build.rs
or xtask
via the builder functions to generate persistent files e.g. when building API bindings
If generation fails, doesn't compile or is generally lousy: Please file an
issue and include the JSON Schema and Rust output (if there is any). Use cargo typify
command to generate code from the command-line. It's even more helpful
if you can articulate the output you'd ideally like to see.
Typify translates JSON Schema types in a few different ways depending on some basic properties of the schema:
Integers, floating-point numbers, strings, etc. Those all have straightforward representations in Rust. The only significant nuance is how to select the appropriate built-in type based on type attributes. For example, a JSON Schema might specify a maximum and/or minimum that indicates the appropriate integral type to use.
String schemas that include a known format
are represented with the
appropriate Rust type. For example { "type": "string", "format": "uuid" }
is
represented as a uuid::Uuid
(which requires the uuid
crate be included as a
dependency).
JSON Schema arrays can turn into one of three Rust types Vec<T>
, HashSet<T>
,
and tuples depending on the schema properties. An array may have a fixed length
that matches a fixed list of item types; this is well represented by a Rust
tuple. The distinction between Vec<T>
and HashSet<T>
is only if the
schema's uniqueItems
field is false
or true
respectively.
In general, objects turn into Rust structs. If, however, the schema defines no
properties, Typify emits a HashMap<String, T>
if the additionalProperties
schema specifies T
or a HashMap<String, serde_json::Value>
otherwise.
Properties of generated struct
that are not in the required
set are
typically represented as an Option<T>
with the #[serde(default)]
attribute
applied. Non-required properties with types that already have a default value
(such as a Vec<T>
) simply get the #[serde(default)]
attribute (so you won't
see e.g. Option<Vec<T>>
).
The oneOf
construct maps to a Rust enum. Typify maps this to the various
serde enum types.
The 'allOf' construct is handled by merging schemas. While most of the time, typify tries to preserve and share type names, it can't always do this when merging schemas. You may end up with fields replicated across type; optimizing this generation is an area of active work.
The anyOf
construct is much trickier. If can be close to an enum
(oneOf
),
but where no particular variant might be canonical or unique for particular
data. While today we (imprecisely) model these as structs with optional,
flattened members, this is one of the weaker areas of code generation.
Issues describing example schemas and desired output are welcome and helpful.
Schemas derived from Rust types may include an extension that provides information about the original type:
{
"type": "object",
"properties": { .. },
"x-rust-type": {
"crate": "crate-o-types",
"version": "1.0.0",
"path": "crate_o_types::some_mod::SomeType"
}
}
The extension includes the name of the crate, a Cargo-style version requirements spec, and the full path (that must start with ident-converted name of the crate).
Each of the modes of using typify allow for a list of crates and versions to be
specified. In this case, if the user specifies "crate-o-types@1.0.1" for
example, then typify would use its SomeType
type rather than generating one
according to the schema.
Each mode of using typify has a method for controlling the use of types with
x-rust-type
annotations. The default is to ignore them. The recommended
method is to specify each crate and version you intend to use. You can
additionally supply the *
version for crates (which may result in
incompatibilities) or you can define a policy to allow the use of all "unknown"
crates (which may require that addition of dependencies for those crates).
For the CLI:
$ cargo typify --unknown-crates allow --crate oxnet@1.0.0 ...
For the builder:
let mut settings = typify::TypeSpaceSettings::default();
settings.with_unknown_crates(typify::UnknownPolicy::Allow)
.with_crate("oxnet", typify::CrateVers::Version("1.0.0".parse().unwrap()));
For the macro:
typify::import_types!(
schema = "schema.json",
unknown_types = Allow,
crates {
"oxnet" = "1.0.0"
}
)
The version
field within the x-rust-type
extension follows the Cargo
version requirements specification. If the extension specifies 0.1.0
of a
crate and the user states that they're using 0.1.1
, then the type is used;
conversely, if the extension specifies 0.2.2
and the user is only using
0.2.0
the type is not used.
Crate authors may choose to adhere to greater stability than otherwise provided
by semver. If the extension version is >=0.1.0, <1.0.0
then the crate author
is committing to the schema compatibility of the given type on all releases
until 1.0.0
. It is important that crate authors populate the version
field
in a way that upholds type availability. For example, while *
is a valid
value, it is only conceivably valid if the type in question were available in
the first ever version of a crate published and never changed incompatibly in
any subsequent version.
The x-rust-type
extension may also specify type parameters:
{
"$defs": {
"Sprocket": {
"type": "object",
"properties": { .. },
"x-rust-type": {
"crate": "util",
"version": "0.1.0",
"path": "util::Sprocket",
"parameters": [
{
"$ref": "#/$defs/Gizmo"
}
]
}
},
"Gizmo": {
"type": "object",
"properties": { .. },
"x-rust-type": {
"crate": "util",
"version": "0.1.0",
"path": "util::Gizmo"
}
}
}
}
With the util@0.1.0
crate specified during type generation, schemas
referencing #/$defs/Sprocket
would use the (non-generated) type
util::Sprocket<util::Gizmo>
.
The parameters
field is an array of schemas. They may be inline schemas or
referenced schemas.
x-rust-type
in your libraryThe schema for the expected value is as follows:
{
"description": "schema for the x-rust-type extension",
"type": "object",
"properties": {
"crate": {
"type": "string",
"pattern": "^[a-zA-Z0-9_-]+$"
},
"version": {
"description": "semver requirements per a Cargo.toml dependencies entry",
"type": "string"
},
"path": {
"type": "string",
"pattern": "^[a-zA-Z0-9_]+(::[a-zA-Z0-9+]+)*$"
},
"parameters": {
"type": "array",
"items": {
"$ref": "#/definitions/Schema"
}
}
},
"required": [
"crate",
"path",
"version"
]
}
The version
field expresses the stability of your type. For example, if
0.1.0
indicates that 0.1.1
users would be fine whereas 0.2.0
users would
not use the type (instead generating it). You can communicate a future
commitment beyond what semver implies by using the Cargo version requirement
syntax.
For example >=0.1.0, <1.0.0
says that the type will remain structurally
compatible from version 0.1.0
until 1.0.0
.
You can format generated code using crates such as
rustfmt-wrapper and
prettyplease. This can be particularly useful
when checking in code or emitting code from a build.rs
.
The examples below show different ways to convert a TypeSpace
to a string
(typespace
is a typify::TypeSpace
).
rustfmt
Best for generation of code that might be checked in alongside hand-written
code such as in the case of an xtask
or stand-alone code generator (such as
cargo-typify
).
rustfmt_wrapper::rustfmt(typespace.to_stream().to_string())?
prettyplease
Best for build.rs
scripts where transitive dependencies might not have
rustfmt
installed so should be self-contained.
prettyplease::unparse(&syn::parse2::<syn::File>(typespace.to_stream())?)
If no human will ever see the code (and this is almost never the case).
typespace.to_stream().to_string()
Typify is a work in progress. Changes that affect output will be indicated with a breaking change to the crate version number.
In general, if you have a JSON Schema that causes Typify to fail or if the generated type isn't what you expect, please file an issue.
There are some known areas where we'd like to improve:
JSON schema can express a wide variety of types. Some of them are easy to model in Rust; others aren't. There's a lot of work to be done to handle esoteric types. Examples from users are very helpful in this regard.
Bounded numbers aren't very well handled. Consider, for example, the schema:
{
"type": "integer",
"minimum": 1,
"maximum": 6
}
The resulting types won't enforce those value constraints.
A string schema with format
set to uuid
will result in the uuid::Uuid
type; similarly, a format
of date
translates to
chrono::naive::NaiveDate
. For users that don't want dependencies on
uuid
or chrono
it would be useful for Typify to optionally represent those
as String
(or as some other, consumer-specified type).