Skip to content

Latest commit

 

History

History
 
 

automate-gateway

README.md

Public-facing API Gateway for Chef Automate

Chef Automate's public-facing API is available in two forms: GRPC and RESTful HTTP. Both flavors of the API are provided by the automate-gateway component which sits in the application layer of Automate's architecture. In this document, we'll use "gateway" to refer to the automate-gateway component.

Here is a high-level sequence diagram showing communication paths from a user's browser through to domain services. The gateway is accessible via the external load balancer (automate-load-balancer). Its job is to respond to API calls by leveraging one or more domain services. If you're familiar with Model, View, Controller (MVC) architectures, it may help you to think of the gateway as a controller and the domain services as models.

sequence-diagram

API Documentation

Today, we generate docs for the HTTP API only. They can be viewed from within an Automate deployment by browsing to https://$YOUR_A2/api/v0/openapi/ui/.

(TODO: add notes on how to document your endpoints)

Compatibility

Our goal is to maintain backward compatibility such that existing clients and tooling that make use of our HTTP or GRPC API continue to work across upgrades of Chef Automate.

Clients of our GRPC API can expect stability of:

  • Method name
  • Request message type
  • Response message type

GRPC request and response message types can evolve by addition of new fields. Fields cannot be removed from either. Note that for response messages, values must continue to be set in responses.

Clients of our HTTP API can expect stability of:

  • URL path and parameters
  • Request body
  • Expected fields in response body. The API may respond with additional fields.

Changes that alter the semantics of an API, even if signatures and messages do not change, are considered breaking changes.

A change is compatible if:

  1. Adds new functionality that is not required.

  2. Does not change existing semantics.

    a. No change in default value or behavior.

    b. Does not change interpretation of fields or values.

  3. Validation does not get more strict.

  4. Consider a request body serialized before your change, and then sent after the change. Does the system still work and respond as expected?

  5. A single property cannot be represented using multiple fields.

Getting Started

  1. Launch the Automate Habitat studio dev environment by following the DEV_ENVIRONMENT.md instructions.

  2. Test the gateway with a request using the gateway_get helper: gateway_get /version

  3. Run the gateway unit tests: go_component_unit automate-gateway

  4. If you add or update any of the .proto files that define the API, you can regenerate code like this: compile_go_protobuf_component automate-gateway. If you see an error about invalid ELF header try unsetting the CC environment variable (unset CC) and retry.

Structure

$ tree -L 1
.
├── api          - proto files + auto-generated go bindings
├── cmd          - cli to start the service
├── habitat      - Habitat configuration
├── config       - configuration and self-signed certs
├── gateway      - core grpc server
├── scripts      - build scripts
├── handler      - implementation of the backend services
├── protobuf     - protobuf helpers
└── third_party  - swagger ui

Extend the API

All public APIs are described via a GRPC service definition in a .proto file.

Add a new endpoint to an existing service

  1. Add an authz smoke test for the API.
  2. Add the method and any new message types for the request and response to the .proto file under a2/components/automate-gateway/api (Example).
  3. Annotate your API with an http option, defining its HTTP method and endpoint (e.g. option (google.api.http).get = "/cfgmgmt/nodes";).
  4. Annotate your API with the authorization resource and action mappings (e.g. option (chef.automate.api.policy).resource = "cfgmgmt:nodes"; option (chef.automate.api.policy).action = "read";). In the example file just above, the chef.automate.api.policy options are for IAM v1 while the chef.automate.api.iam.policy options are for IAM v2. During the current transition time you must supply both sets of values.
  5. Compile the proto file from within your habitat studio, using compile_go_protobuf_component automate-gateway.
  6. Add an implementation for the new method in the appropriate file under handler/....
  7. If there is no default policy governing the API's resource and action, define one in storage.go and add a migration to a2/components/authz-service. Note This whole step is necessary only if the API should be accessible to non-admins.

Add a new service

  1. Place your service in api/... following the existing pattern.
  2. Follow above steps for adding a new endpoint to existing service. Be sure to generate code by running the following from the project root in the dev studio: compile_go_protobuf_component automate-gateway
  3. Wire your service into RegisterRestGWServices and RegisterGRPCServices in gateway/services.go

Authorization for API Endpoints

APIs are permissioned using Automate's authorization system. Each public API needs to be mapped to an authorization resource and action. That resource and action must correspond to a default policy defined in the system, so you need to either map to a resource and action governed by an existing default policy, or define a new default policy governing the mapped resource and action.

Our authorization policies state which subjects have the right to perform which actions on which resources. This means the choices we make in resources and action names define the vocabulary our customers will have to use to manage permissioning. That makes these definitions a product concern.

An Example Let's say we have two APIs: list all nodes (1) and get node statistics (2).

Should those APIs map to the same resource and action (e.g. resource: nodes:*, action: read) or should they have different mappings?

To answer that question, it is important to understand whether our customers would expect that those interactions be permissioned separately from each other, or instead that if someone was allowed to list nodes, it should follow that they are also allowed to see statistics about those nodes.

To put it another way, is the customer going to want "Alex is allowed to read nodes" to both mean Alex is able list all the nodes and get statistics on those nodes?

Another Example If a user is allowed to read information about a node, should they also be able to read information about the converge runs for that node?

If so, maybe the resource and action for both the read node and read node runs APIs should be nodes:* and read. If not, then maybe we should have separate node:* resource and a node:{id}:run resources (the actions might still both be read).

Testing the API

The sorts of testing we want in place for exposed gateway APIs are meant to ensure:

  1. Does each API have a resource/action mapping, and is it mapping to the correct resource and action? This testing is done automatically by inplace infrastructure. See How Automate Ensures Integrity of Protobuf Files

  2. Does each API's resource/action mapping and the default policy governing it result in the expected access (e.g. if only admins should be able to access it, is that the case)? Add tests of this sort as smoke tests.

Documenting the API

Comments on service methods will be integrated into the swagger docs for HTTP.

Data generation

From within the Habitat studio there are a few options for generating Chef converge runs. Running chef_load_nodes 100 will generate 100 nodes that converged over the last hour through the data-collector endpoint. To have ongoing ingestion of nodes, run chef_load_start. This will start the ingestion of 10 nodes and 10 Chef actions every five minutes. If you only want to send one static Chef converge run, use the send_chef_run_example command. This sends data from the following file with only the IDs updated: https://github.com/chef/automate/blob/master/components/ingest-service/examples/converge-success-report.json. This node will go missing quickly because of the old date for the end_time.