Best Practices and Recipes

This documentation is for a previous version of Cerbos. Choose 0.40.0 from the version picker at the top right or navigate to https://docs.cerbos.dev for the latest version.

A collection of tips and code snippets designed to help you write cleaner, more optimised Cerbos policies.

Modelling policies

With Cerbos, access rules are always resource-oriented and the policies you write map to these resources within your system. A resource can be anything, and the way you model your policies is up you — you can achieve the same logical outcome in numerous ways; action-led, role-led, attribute-led, or with combinations thereof.

That said, some patterns will lend themselves more naturally to certain scenarios — let’s take a look at some different approaches. Consider this business model:

Actions

Roles

IT_ADMIN

JR_MANAGER

SR_MANAGER

USER

CFO

run

x

x

x

view

x

x

x

x

x

edit

x

x

save

x

x

share

x

x

x

Representing this as a resource policy could be achieved in a variety of ways. Let’s take a look at each:

Action-led

Here, we focus on an action, and list all the roles that can perform that action:

# Principals in the following three roles can perform the `run` action
  - actions:
      - "run"
    effect: EFFECT_ALLOW
    roles:
      - JR_MANAGER
      - SR_MANAGER
      - CFO

# All principals can perform the `view` action
  - actions:
      - "view"
    effect: EFFECT_ALLOW
    roles:
      - ["*"]

This approach might be suitable if any of the following apply to your system:

  • Your roles are "similar" in what they can do like JR_MANAGER and SR_MANAGER; it’s likely that JR_MANAGER will have a subset of the permissions of SR_MANAGER. There will of course be duplication in either direction, but it’s often easier to reason about this from an action perspective.

  • You have "high-risk" actions — you want to be able to tell at a glance which roles have access to a particular action. The act of explicitly listing roles per action makes it much more difficult to accidentally give unwanted permissions to the wrong user.

  • You have a relatively high number of roles to a low number of actions.

Role-led

Alternatively, we can focus on a role, and list all the actions the role can perform:

# These three actions can be performed by principals in the `JR_MANAGER` role
  - actions:
      - "run"
      - "view"
      - "share"
    effect: EFFECT_ALLOW
    roles:
      - JR_MANAGER

You might opt for a role-led approach if:

  • You have distinct roles where it’s rare for your roles to share common actions.

  • You have a relatively low number of roles to a high number of actions.

Hybrid

Perhaps we want to use a combination of the two:

# Principals in the `SR_MANAGER` or `CFO` roles can perform all actions
  - actions:
      - "*"
    effect: EFFECT_ALLOW
    roles:
      - SR_MANAGER
      - CFO

This might apply if your scenario doesn’t strictly fall into one of the previous two sections; individually, or at all.

Blanket allow, granular deny

We can opt to explicitly state which actions a user cannot do:

# Principals in the `JR_MANAGER` role can perform all actions, other than `edit` and `save`
  - actions:
      - "*"
    effect: EFFECT_ALLOW
    roles:
      - "JR_MANAGER"

  - actions:
      - "edit"
      - "save"
    effect: EFFECT_DENY
    roles:
      - "JR_MANAGER"

This would suit scenarios where a principal can perform nearly every action, and you want to explicitly list disallowed actions.

Attribute-led

Consider the following hypothetical scenario:

An organization models its resources as specific data sets. Each data set is unique, as are the principals trying to access them. The organization uses JWTs extensively to manage and transmit identity/contextual information. The resource policies map 1:1 to each data set, and access is governed by arbitrary information (in this case, passed within the JWT).

Given the dynamic nature of audiences, it’s not practical to enumerate all roles that have access. What we could do instead is to globally allow all roles and actions and then determine access based on attributes passed in the JWT. Take a look at the following example policy:

apiVersion: api.cerbos.dev/v1
resourcePolicy:
  resource: "data_set"
  version: default
  rules:
    - actions: ["*"]
      roles: ["*"]
      effect: EFFECT_ALLOW
      condition:
        match:
          all:
            of:
              - expr: has(request.aux_data.jwt.aud)
              - expr: >
                  "my.custom.audience" in request.aux_data.jwt.aud

In the above, we blanket-allow all actions and roles, but specifically rely on the aud key parsed from the JWT to determine access.

Map of relations

Avoiding complex query planner outputs with sub-expressions:

Before
apiVersion: api.cerbos.dev/v1
resourcePolicy:
  resource: example_resource
  version: default
  rules:
    - actions:
        - "*"
      effect: EFFECT_ALLOW
      roles:
        - USER
      condition:
        match:
          expr: P.attr.relations[R.id] == "owner"

Notice how, in the above match condition, we reference both the Principal and the Resource within a single struct indexing expression: P.attr.relations[R.id]. This is correct both syntactically and logically. However, when using the PlanResources API to produce a query plan, the generated plan can sometimes be quite complicated due to how these expressions are represented internally.

To improve this, we can apply a little bit of mechanical-sympathy and rewrite the expression in a way where the known values (principal attributes) are separated from unknown values (resource attributes) as follows:

After
apiVersion: api.cerbos.dev/v1
resourcePolicy:
  resource: example_resource
  version: default
  rules:
    - actions:
        - "*"
      effect: EFFECT_ALLOW
      roles:
        - USER
      condition:
        match:
          expr: R.id in P.attr.relations.filter(x, P.attr.relations[x] == "owner")

Adding self-service custom roles

Imagine this scenario: you’re an admin in a multi-tenant system, and you want a method by which you can copy an existing role, and then select which permissions/actions to enable or disable for each.

There are two ways of approaching this:

Static Policies / Dynamic Context

This is the idiomatic way of solving this use-case in Cerbos. In the vast majority of cases, it is possible to have the policies statically defined and to pass in dynamic context as attributes of a principal. This dynamic context can be any arbitrary data such as the principal’s location, age, or specific roles it has within the context of an organizational unit (a department, a tenant or a project, for example). This contextual data would be retrieved at request time from another service or a data store. Let’s look at an example.

Here is a resource policy for a resource of type "workspace":

workspace.yaml
apiVersion: "api.cerbos.dev/v1"
resourcePolicy:
  version: "default"
  resource: "workspace"
  rules:
    - actions:
        - workspace:view
        - pii:view
      effect: EFFECT_ALLOW
      roles:
        - USER
      condition:
        match:
          expr: P.attr.workspaces[R.id].role == "OWNER"

Notice how the condition relies on context passed in within the P.attr.workspaces map, with the key being the resource ID, and the value being a predefined value "OWNER". We can grant access to a principal with the USER role, by constructing the following request payload:

  • cURL

  • .NET

  • Go

  • Java

  • JS

  • PHP

  • Python

  • Ruby

  • Rust

cat <<EOF | curl --silent "http://localhost:3592/api/check/resources?pretty" -d @-
{
  "requestId": "quickstart",
  "principal": {
    "id": "123",
    "roles": [
      "USER"
    ],
    "attr": {
      "workspaces": {
        "workspaceA": {
          "role": "OWNER"
        },
        "workspaceB": {
          "role": "MEMBER"
        }
      }
    }
  },
  "resources": [
    {
      "actions": [
        "workspace:view",
        "pii:view"
      ],
      "resource": {
        "id": "workspaceA",
        "kind": "workspace"
      }
    },
    {
      "actions": [
        "workspace:view",
        "pii:view"
      ],
      "resource": {
        "id": "workspaceB",
        "kind": "workspace"
      }
    }
  ]
}
EOF
using Cerbos.Sdk.Builders;
using Cerbos.Sdk;

internal class Program
{
    private static void Main(string[] args)
    {
        var client = new CerbosClientBuilder("http://localhost:3593").WithPlaintext().BuildBlockingClient();
        string[] actions = { "workspace:view", "pii:view" };

        CheckResourcesResult result = client
            .CheckResources(
                Principal.NewInstance("123", "USER")
                    .WithAttribute("workspaces", AttributeValue.MapValue(new Dictionary<string, AttributeValue>()
                    {
                        {
                            "workspaceA", AttributeValue.MapValue(new Dictionary<string, AttributeValue>()
                            {
                                {"role", AttributeValue.StringValue("OWNER")}
                            })
                        },
                        {
                            "workspaceB", AttributeValue.MapValue(new Dictionary<string, AttributeValue>()
                            {
                                {"role", AttributeValue.StringValue("MEMBER")}
                            })
                        }
                    })),

                ResourceAction.NewInstance("workspace", "workspaceA")
                    .WithActions(actions),

                ResourceAction.NewInstance("workspace", "workspaceB")
                    .WithActions(actions)
            );

        foreach (string n in new string[] { "workspaceA", "workspaceB" })
        {
            var r = result.Find(n);
            Console.Write(String.Format("\nResource: {0}\n", n));
            foreach (var i in r.GetAll())
            {
                String action = i.Key;
                Boolean isAllowed = i.Value;
                Console.Write(String.Format("\t{0} -> {1}\n", action, isAllowed ? "EFFECT_ALLOW" : "EFFECT_DENY"));
            }
        }
    }
}
package main

import (
	"context"
	"log"

	"github.com/cerbos/cerbos-sdk-go/cerbos"
)

func main() {
	c, err := cerbos.New("localhost:3593", cerbos.WithPlaintext())
	if err != nil {
		log.Fatalf("Failed to create client: %v", err)
	}

	principal := cerbos.NewPrincipal("123", "USER")
	// We use map[string]interface{} as strictly typed nested maps aren't supported
	principal.WithAttr("workspaces", map[string]map[string]interface{}{
		"workspaceA": map[string]interface{}{
			"role": "OWNER",
		},
		"workspaceB": map[string]interface{}{
			"role": "MEMBER",
		},
	})

	kind := "workspace"
	actions := []string{"workspace:view", "pii:view"}

	batch := cerbos.NewResourceBatch()
	batch.Add(cerbos.NewResource(kind, "workspaceA"), actions...)
	batch.Add(cerbos.NewResource(kind, "workspaceB"), actions...)

	resp, err := c.CheckResources(context.Background(), principal, batch)
	if err != nil {
		log.Fatalf("Failed to check resources: %v", err)
	}
	log.Printf("%v", resp)
}
package demo;

import static dev.cerbos.sdk.builders.AttributeValue.mapValue;
import static dev.cerbos.sdk.builders.AttributeValue.stringValue;

import java.util.Map;

import dev.cerbos.sdk.CerbosBlockingClient;
import dev.cerbos.sdk.CerbosClientBuilder;
import dev.cerbos.sdk.CheckResult;
import dev.cerbos.sdk.builders.Principal;
import dev.cerbos.sdk.builders.ResourceAction;


public class App {
    public static void main(String[] args) throws CerbosClientBuilder.InvalidClientConfigurationException {
        CerbosBlockingClient client=new CerbosClientBuilder("localhost:3593").withPlaintext().buildBlockingClient();

        for (String n : new String[]{"workspaceA", "workspaceB"}) {
            CheckResult cr = client.batch(
                Principal.newInstance("123", "USER")
                    .withAttribute("workspaces", mapValue(Map.of(
                        "workspaceA", mapValue(Map.of(
                                "role", stringValue("OWNER")
                        )),
                        "workspaceB", mapValue(Map.of(
                                "role", stringValue("MEMBER")
                        ))
                    )))
                )
                .addResources(
                    ResourceAction.newInstance("workspace","workspaceA")
                        .withActions("workspace:view", "pii:view"),
                    ResourceAction.newInstance("workspace","workspaceB")
                        .withActions("workspace:view", "pii:view")
                )
                .check().find(n).orElse(null);

            if (cr != null) {
                System.out.printf("\nResource: %s\n", n);
                cr.getAll().forEach((action, allowed) -> { System.out.printf("\t%s -> %s\n", action, allowed ? "EFFECT_ALLOW" : "EFFECT_DENY"); });
            }
        }
    }
}
const { GRPC: Cerbos } = require("@cerbos/grpc");

const cerbos = new Cerbos("localhost:3593", { tls: false });

(async() => {
  const kind = "workspace";
  const actions = ["workspace:view", "pii:view"];

  const cerbosPayload = {
    principal: {
      id: "123",
      roles: ["USER"],
      attributes: {
        workspaces: {
          workspaceA: {
            role: "OWNER",
          },
          workspaceB: {
            role: "MEMBER",
          }
        },
      },
    },
    resources: [
      {
        resource: {
          kind: kind,
          id: "workspaceA",
        },
        actions: actions,
      },
      {
        resource: {
          kind: kind,
          id: "workspaceB",
        },
        actions: actions,
      },
    ],
  };

  const decision = await cerbos.checkResources(cerbosPayload);
  console.log(decision.results)
})();
<?php

require __DIR__ . '/vendor/autoload.php';

use Cerbos\Sdk\Builder\CerbosClientBuilder;
use Cerbos\Sdk\Builder\Principal;
use Cerbos\Sdk\Builder\ResourceAction;
use Symfony\Component\HttpClient\HttplugClient;

$clientBuilder = new CerbosClientBuilder("http://localhost:3592", new HttplugClient(), null, null, null);
$client = $clientBuilder->build();

$principal = Principal::newInstance("123")
              ->withRole("USER")
              ->withAttribute("workspaces", [
                  "workspaceA" => [
                      "role" => "OWNER"
                  ],
                  "workspaceB" => [
                      "role" => "MEMBER"
                  ]
              ]);

$type = "workspace";

$resourceAction1 = ResourceAction::newInstance($type, "workspaceA")
                    ->withAction("workspace:view")
                    ->withAction("pii:view");

$resourceAction2 = ResourceAction::newInstance($type, "workspaceB")
                    ->withAction("workspace:view")
                    ->withAction("pii:view");

$checkResourcesResult = $client->checkResources($principal, array($resourceAction1, $resourceAction2), null, null);

echo json_encode($checkResourcesResult, JSON_PRETTY_PRINT);

?>
import json

from cerbos.sdk.client import CerbosClient
from cerbos.sdk.model import Principal, Resource, ResourceAction, ResourceList
from fastapi import HTTPException, status

principal = Principal(
    "123",
    roles=["USER"],
    attr={
        "workspaces": {
            "workspaceA": {
                "role": "OWNER",
            },
            "workspaceB": {
                "role": "MEMBER",
            }
        }
    }
)

actions = ["workspace:view", "pii:view"]
resource_list = ResourceList(
    resources=[
        ResourceAction(
            Resource(
                "workspaceA",
                "workspace",
            ),
            actions=actions,
        ),
        ResourceAction(
            Resource(
                "workspaceB",
                "workspace",
            ),
            actions=actions,
        ),
    ],
)

with CerbosClient(host="http://localhost:3592") as c:
    try:
        resp = c.check_resources(principal=principal, resources=resource_list)
        resp.raise_if_failed()
    except Exception:
        raise HTTPException(
            status_code=status.HTTP_403_FORBIDDEN, detail="Unauthorized"
        )

print(json.dumps(resp.to_dict(), sort_keys=False, indent=4))
# frozen_string_literal: true

require "cerbos"
require "json"

client = Cerbos::Client.new("localhost:3593", tls: false)

kind = "workspace"
actions = ["workspace:view", "pii:view"]

r1 = {
  kind: kind,
  id: "workspaceA"
}

r2 = {
  kind: kind,
  id: "workspaceB"
}

decision = client.check_resources(
  principal: {
    id: "123",
    roles: ["USER"],
    attributes: {
      workspaces: {
        workspaceA: {
          role: "OWNER"
        },
        workspaceB: {
          role: "MEMBER"
        }
      }
    }
  },
  resources: [
    {
      resource: r1,
      actions: actions
    },
    {
      resource: r2,
      actions: actions
    }
  ]
)

puts JSON.pretty_generate({
  results: [
    {
      resource: r1,
      actions: {
        "workspace:view": decision.allow?(resource: r1, action: "workspace:view"),
        "pii:view": decision.allow?(resource: r1, action: "pii:view")
      }
    },
    {
      resource: r2,
      actions: {
        "workspace:view": decision.allow?(resource: r2, action: "workspace:view"),
        "pii:view": decision.allow?(resource: r2, action: "pii:view")
      }
    }
  ]
})
use cerbos::sdk::attr::{attr, StructVal};
use cerbos::sdk::model::{Principal, Resource, ResourceAction, ResourceList};
use cerbos::sdk::{CerbosAsyncClient, CerbosClientOptions, CerbosEndpoint, Result};

#[tokio::main]
async fn main() -> Result<()> {
    let opt =
        CerbosClientOptions::new(CerbosEndpoint::HostPort("localhost", 3593)).with_plaintext();
    let mut client = CerbosAsyncClient::new(opt).await?;

    let principal = Principal::new("123", ["USER"]).with_attributes([attr(
        "workspaces",
        StructVal([
            ("workspaceA", StructVal([("role", "OWNER")])),
            ("workspaceB", StructVal([("role", "MEMBER")])),
        ]),
    )]);

    let actions: [&str; 2] = ["workspace:view", "pii:view"];

    let kind = "workspace";
    let resp = client
        .check_resources(
            principal,
            ResourceList::new_from([
                ResourceAction(Resource::new("workspaceA", kind), actions),
                ResourceAction(Resource::new("workspaceB", kind), actions),
            ]),
            None,
        )
        .await?;

    println!("{:?}", resp.response);

    Ok(())
}

You can find a full (and extended) example of the above in our SaaS Workspace Policy playground example.

Dynamic Policies

There might be circumstances where you want to create or update resources and actions on the fly; an example of this might be a multi-tenant platform that provides tenants the ability to manage their own policies.

If this is the case, then you can use the Admin API configured alongside a mutable database storage engine to provide this functionality. This would be handled within your application layer, with the desired policy contents provided to the PDP via the API.

For a full example implementation, check out this demo.

Policy repository layout

Cerbos expects the policy repository to have a particular directory layout.

  • The directory must only contain Cerbos policy files, policy test files and schemas. Any other YAML or JSON files will cause Cerbos to consider the policy repository as invalid.

  • If you use schemas, the _schemas directory must be a top-level directory at the root of the policy repo.

  • All policy tests must have a file name ending in _test and a .yaml, .yml or .json extension.

  • Directories named testdata can be used to store test data for policy tests. Cerbos will not attempt to locate any policy files inside those directories.

  • Hidden files and directories (names starting with .) are ignored.

A typical policy repository might resemble the following:

.
├── _schemas
│   ├── principal.json
│   └── resources
│       ├── leave_request.json
│       ├── purchase_order.json
│       └── salary_record.json
├── derived_roles
│   ├── backoffice_roles.yaml
│   └── common_roles.yaml
├── principal_policies
│   └── auditor_audrey.yaml
└── resource_policies
    ├── finance
    │   ├── purchase_order.yaml
    │   └── purchase_order_test.yaml
    └── hr
        ├── leave_request.yaml
        ├── leave_request_test.yaml
        ├── salary_record.yaml
        ├── salary_record_test.yaml
        └── testdata
            ├── auxdata.yaml
            ├── principals.yaml
            └── resources.yaml