I found myself wanting to rewrite JSON files with `sed`. The problem is, then I want to run a formatter over them afterwards, and this will change the whole file, not just the area I touched. I would like to propose the nuclear option in remedying this: format everything now. This is a very large change that should make it easier to keep files to a consistent format in the future. I have chosen to use Prettier for this because (a) it has a useful `--write` command and (b) it also does GraphQL, Markdown, YAML, etc. I've elected to exclude two sets of files: 1. `crates/custom-connector/data/*.json`, because they are actually multiple JSON objects, one per line, which Prettier cannot parse. 2. `crates/lang-graphql/tests/**/*.graphql`, because it contains invalid GraphQL, and the parser is intended to work with strangely-formatted GraphQL. The main changes are standardizing whitespace, adding a newline at the end of files, and putting JSON arrays on one line when they fit. V3_GIT_ORIGIN_REV_ID: 92d4a535c34a3cc00721e8ddc6f17c5717e8ff76
5.3 KiB
Roles / namespaces / annotations in v3-engine
In V2 we created a schema per role.
In V3 we create one big schema at build time, where some nodes are annotated
with either generic
info, or namespaced
info that only applies to specific
roles.
The types for these annotations come from the lang-graphql
crate, where they
are associated types on the SchemaContext
trait.
This example used in tests for lang-graphql
shows a schema with no
annotations. It uses a Namespace
type to differentiate different scopes,
however since it has no annotations this won't do anything meaningful:
impl SchemaContext for SDL {
type Namespace = Namespace;
type GenericNodeInfo = ();
type NamespacedNodeInfo = ();
The engine
itself uses the GDS
type, and so we can use the following
annotations:
impl gql_schema::SchemaContext for GDS {
type Namespace = Role;
type GenericNodeInfo = types::Annotation;
type NamespacedNodeInfo = Option<types::NamespaceAnnotation>;
Note that our Namespace
type is Role
- this means that the engine attached
useful information per role, so we can add select permissions for a field that
only apply to the user-1
role, for instance.
(incidentally, there is a comment on the SchemaContext
trait suggesting
Namespace
is renamed to Scope
or Role
- this seems like a useful move to
make all this clearer, personally Namespace
makes me think about different
subgraphs or something)
What is an annotation then, concretely?
The change I have recently been working on is to add preset arguments to
commands. This means that given a delete_user
command with a user_id: Int
argument, we can preset the value for certain roles.
For instance, we might want a user-1
role to only be able to delete
themselves, so we'd preset user_id
with
{ "sessionVariable": "x-hasura-user-id" }
.
Previous to this change, when building the schema for the Command, we'd use
builder.allow_all_namespaced
to create it, like
builder.allow_all_namespace(command_schema_stuff, None)
Instead, for our example, we'd use conditional_namespaced
like this (excuse my
pseudo-Rust):
let role_annotations = HashMap::new();
role_annotations.insert("user-1", Some(ArgumentPresets { "user_id":
"x-hasura-role-id" }));
role_annotations.insert("admin", None);
builder.conditional_namespaced(command_schema_stuff, role_annotations)
Here, we've added an annotation for user-1
with some useful information we can
use later when receiving and running queries. We've also added a None
annotation for admin
- this means there'll be no useful information later, but
we're still signalling that we want this Command
to work for admin
users. If
we had a user-2
role in the schema, they wouldn't be able to use the command.
Removing items from the GraphQL schema, per role
By adding or omitting keys from the role_annotations
above, we've made our
command appear or disappear from the GraphQL schema. This "snipping" as such,
happens in the normalize_request
function in lang-graphql
.
The important thing to know here is that lang-graphql
code does not know about
GDS
or our NamespaceAnnotation
type. All it has to act on is whether a
Role
(or Namespace
, to it's eyes) has a key in any namespaced annotations or
not. Because we're using associated types the contents are "protected" from
lang-graphql
and so it can't peek inside.
We're going to want the user_id
argument to disappear from user-1
's schema
- we do this by using
conditional_namespaced
when contructing schema for theuser_id
command argument itself:
let role_annotations = HashMap::new();
// insert an empty annotation for `admin` to make sure the argument remains in
the schema
role_annotations.insert("admin", None);
// don't add one for `user-1`, so it disappears from the schema (as it has been
replaced with the preset value)
builder.conditional_namespaced(command_argument_schema_stuff, role_annotations)
Reading the annotations at request time
Everything before here happens at "compile time" for the engine, so we do it all once at startup and it remains static for the lifetime of the application.
At some point we are going to need to serve a request though. The steps are as follows:
- Receive request
- Parse into a GraphQL query
- Normalize the query
- Generate IR
- Construct a query plan
- Execute / explain the query plan
The normalize
step was mentioned earlier - it takes place in lang_graphql
and combines the query with the role information to remove all irrelevant
namespace information from the query, and snip any parts of the schema that the
current role can't see.
Generally the helpful place for our annotations is in generating IR
(intermediate representation, compiler fans). For our commands change, we can
look up the ArgumentPresets
we stored earlier like
this.
match field_call.info.namespaced {
None => {}
Some(NamespaceAnnotation::ArgumentPresets(argument_presets)) => {
// use `argument_presets` for current role to generate IR
...
}
}
The IR will then be created from a mixture of the schema for the user's role, and any arguments etc from their request.