Safe-portals solves 3 problems with (JSON) serialization in JS/Typescript:
- Many types are not handled. For example,
Date
objects are deserialized from JSON as strings. We want the original data to be unchanged by going through serialization and deserialization. - The schema of our data is not validated.
- In TypeScript, data emerges from JSON.parse as
any
type. We would prefer to recover type annotations once the data has been validated as matching our expected schema.
Safe-portals can be used wherever data traverses an un-typed boundary (eg from DB, tasks sent to resque, HTTP calls, routing information in URLs), in order to maintain static analysis across the un-typed boundary.
You usually want to import the whole library:
import * as S from 'safe-portals';
A safe-portal describes both the serialization and unserialization operation for a particular type, bridging the gap between arbitrary typed values (like JS Dates, or custom interfaces) and JSONable JS objects (ie only string, number, boolean, null, arrays of these types, and objects indexed by strings and containing only these types). For example:
S.dateIso
Is a safe-portal for a date object. The unserialization operation is called 'read':
S.dateIso.read("2020-03-04T12:27:41.360Z")
> Date Wed Mar 04 2020 12:27:41 GMT+0000 (Greenwich Mean Time)
The serialization operation is the other side of the serialization/unserialization symmetry:
S.dateIso.write(new Date())
> "2020-03-04T12:27:41.360Z"
Note that the safe-portal 'write' operation doesn't generate JSON -- it generates JSONifiable JS objects -- and the 'read' operation doesn't expect to be passed JSON, but objects parsed from JSON (or otherwise JSONifiable objects). This means that a typical usage across an un-typed interface (like HTTP) will look like:
send(JSON.stringify(portal.write(value)));
and
portal.read(JSON.parse(receive()))
Portal for string
type.
Portal for boolean
type.
Portal for number
type, truncated to an integer.
Portal for number
type, allowing decimals.
Portal for Date
type, with seconds since the Unix epoch as the
serialized representation.
Portal for Date
type, with milliseconds since the Unix epoch as the
serialized representation.
Portal for Date
type, with ISO 8601 as the serialized representation.
Portal for string
type, validating that the string is a uuid.
Portal for any
type. Allows passthrough to plain JSON
stringify/parse behavior.
JSON-encoded data.
const s = S.json(S.obj({x: S.int, y: S.str}));
s.read("{\"x\": 123, \"y\": \"hi\"}")
Portal for the type given by the schema
portal, with an ordered list
of data migration that will be run (first to last) before schema.read(),
to lazily migrate the data. For example:
const v1 = S.versioned({
schema: S.obj({ x: S.dateIso }),
migrations: []
});
This serializer handles data of the form {x: new Date() }
. Supposing
we wished to change the serialized representation from dateIso to
dateUnixSecs, we could revise our serializer in this way:
const v2 = S.versioned({
schema: S.obj({ x: S.dateUnixSecs }),
migrations: [
o => ({ x: S.dateUnixSecs.write(S.dateIso.read(o.x)) })
]
});
Adds optionality to the portal argument t
. Eg:
S.optional(S.str)
is a portal for the type string | undefined
Adds nullability to the portal argument t
. Eg:
S.nullable(S.str)
is a portal for the type string | null
Portal for an array of the type that portal t
handles. Eg:
S.array(S.float)
is a portal for the type number[]
Portal for arbitrary object types. Eg:
S.obj({
name: S.str,
date_of_birth: S.dateIso,
date_of_death: S.optional(S.dateIso),
})
Is a portal for the type:
{
name: string;
date_of_birth: Date;
date_of_death: Date | undefined;
}
A fully optional version of obj
. Eg:
S.partial_obj({
name: S.str,
date_of_birth: S.dateIso,
date_of_death: S.optional(S.dateIso),
})
Is a portal for the type:
{
name?: string;
date_of_birth?: Date;
date_of_death?: Date | undefined;
}
A portal for a combination of obj and/or partial_obj serializers. Eg:
S.combine(
S.obj({
name: S.str,
date_of_birth: S.dateIso,
}),
S.partial_obj({
date_of_death: S.dateIso,
})
)
Is a portal for the type:
{
name: string;
date_of_birth: Date,
date_of_death?: Date
}
A portal for an array-as-tuple type. Eg:
S.tuple(S.str, S.float, S.dateIso)
Is a portal for the type:
[string, number, Date]
A portal for a string enum type. Eg:
S.oneOf('apple', 'orange', 'pear')
Is a portal for the type:
'apple' | 'orange' | 'pear'
A portal for a tagged-union type. Eg:
S.variant(
'circle', S.obj({ radius: S.float }),
'rectangle', S.obj({ width: S.float, height: S.float }),
)
Is a portal for the type:
{ type: 'circle', radius: float } |
{ type: 'rectangle', width: float, height: float }
A (experimental) portal for a 'success or failure' type, allowing a means of passing exceptional conditions across serialization boundaries in a type-safe manner.
Consult result.test.ts for example usage.
You can build type-portals for types like tuples, arrays and objects:
const personPortal = S.obj({
name: S.str,
date_of_birth: S.dateIso,
date_of_death: S.optional(S.dateIso),
favourite_foods: S.array(S.str)
});
personPortal.write({
name: 'Bob',
date_of_birth: new Date(1985, 10, 15),
date_of_death: undefined,
favourite_foods: ['pizza', 'broccoli']
});
For javascript users, you get validation. For Typescript users you also get type hints in argument value of write, and return value of read:
const person = personPortal.read(data);
person.date_of_death.getYear()
^-- type error: value could be undefined
We can use the following function to define HTTP endpoints in terms of url prefix, arguments and response types:
export interface Endpoint<Args,Resp> {
url: string,
argumentType: S.Type<Args>,
responseType: S.Type<Resp>,
call: (args: Args) => Promise<Resp>
}
export function makeEndpoint<Args, Resp>(
url: string,
argumentType: S.Type<Args>,
responseType: S.Type<Resp>
): Endpoint<Args, Resp>
{
return {
url,
argumentType,
responseType,
call: async (args: Args): Promise<Resp> => {
const resp = await axios.post(url, { args: argumentType.write(args) });
return responseType.read(resp.data.result);
},
}
}
Then, in code shared by both the HTTP back-end and by the front-end, we can define our endpoints like so:
export const savePerson = makeEndpoint(
'/admin/api/journey/save',
// arguments
S.obj({
guid: S.str,
name: S.str,
date_of_birth: S.dateIso,
date_of_death: S.optional(S.dateIso),
favourite_foods: S.array(S.str)
}),
// response (either success (nothing) or an error string)
S.obj({
error: S.optional(S.str)
})
)
Front-end code can import the above endpoint definition, and make HTTP calls in a way that looks and behaves like any asynchronous typescript function call:
const response = await savePerson.call(person);
if (response.error) {
The arguments and return type of call will of course be statically validated, removing cognitive load on the developer.
The back-end implementation could be implemented using the following utility function (its separation from makeEndpoint ensures that ExpressJS doesn't become a front-end dependency!):
export function endpointHandler<Args, Resp>(
app: express.Application,
endpoint: Endpoint<Args, Resp>,
handler: (args: Args) => Promise<Resp>,
)
{
app.post(endpoint.url, async (req, resp): Promise<void> => {
try {
resp.send({
result: endpoint.responseType.write(
await handler(endpoint.argumentType.read(req.body.args))
)
});
} catch (e) {
resp.sendStatus(500);
throw e;
}
});
}
And the actual endpoint implementation in your controller code will simply be:
endpointHandler(
app, // expressJS app instance
savePerson, // the handler definition created previously
async person => {
// person will be correctly typed
// note that we didn't need to annotate the type here
const error = Person.save(person)
return { error }
}
);
Once again, the arguments and return type from the async person => {
function will be validated and checked, and provide IDE feedback.
Now changes to endpoint URIs, changes to argument shape and validation, and changes to response type become safe operations - you can refactor freely, and the compiler or IDE will keep track of the details.