If you've reached this point, it's likely that you have successfully completed the onboarding guide, connected your application with Permit, and checked for permissions. Now, let's dive into the different deployment options available for incorporating Permit into your app.
As you can see in the diagram below there a few key components that are part of every deployment, the only components you have to have are your App with an enforcement point; but some additional can be added to enjoy more features.
|Your App||MUST||Otherwise what's the point? 😜😅|
|Policy Enforcement Point(PEP)||MUST||Otherwise what's the point? 😜😅 These can be Frontend code, Backend code, reverse proxy, or API Gateway|
|Policy Decision Point (PDP)||OPTIONAL||Permit can host it for you. But it's recommended to use locally in production.|
|Policy Git repository||OPTIONAL||Permit can host it for you. You'd host the repo is you want to use the Gitops / manually add policy as code,|
The PDP provides fast, zero-latency, scalable, secure policy decision making for your software.
The PDP opens an outgoing HTTPS websocket connection to Permit.io, which is used to subscribe to updates. The updates happen asynchronously in the background - they are not part of the authorization query critical path.
We often refer to the cloud-hosted PDP option as full-SaaS or cloud-PDP, and the local PDP option as hybrid SaaS. Hybrid SaaS is the default recommended layout.
By default when you first try out Permit, you start with the "cloud" PDP we host for you at "https://cloudpdp.api.permit.io" . This is a great way to start, and can be scalable for some use cases; but we do recommend running with your own local PDP in production or in any performance critical deployments.
The most common way to deploy the PDP locally is as a sidecar - i.e. you run one PDP container next to each of your own microservices. This is also the easiest way to scale your authorization layer with your application.
In this layout you of course enjoy zero-latency between your application and the PDP - this together with improved stability and security (no dependency on other clouds) is the main reason to use a local PDP compared to a cloud pdp.
Running the sidecar in Kubernetes?
There are a few things you need to make sure you do:
- API Service - You need to open up it's port 7000 (often mapped to 7766 when running locally)
- OPA Service - You can open up port 8181 to expose OPA directly
- Use the
/healthcheckendpoint on the API to check for readiness
If you need more advanced policy level healthchecks, you can read more about them as part of OPAL here.
In this slightly more advanced deployment, you run multiple instances of the same PDP (with the same environnement key) behind a load-balancer and direct your services to it. Just like with any other microservice - this enables you scale out the PDPs elastically to match the demand coming from your application. Unless your application is bound to face extreme workloads, and/or highly dynamic workloads, usually there isn't much added benefit for this deployment style over the sidecar one; but even if this is just your preference it's reason enough to work with this layout. When possible prefer to have the various microservices / apps on the same physical node as the PDPs, or close to it as possible to minimize latency.
Deploy as a Cluster with Sharding / Chaining
For truly massive data sets needed in the PDP, you can apply sharding (using Permit environments and/or OPAL topics) to split the data between multiple PDPs within the same cluster, and reroute the requests to the relevant PDP service-mesh style. In addition, you can apply chaining, using OPA's http.send as part of the policy in one PDP to fetch the missing data from another.
When running your app on a Serverless platform (e.g. AWS Lambda, GCP cloud-run), you can simply run a the PDP as a container on your platform's container service, fro example - ECS-Fargate (which BTW is also considered serverless by AWS' definition). The PDP and the app will enjoy all the same latency reductions as if it was running on another cloud function - as long as it's part of the same VPC. If for some reason it is important to that the PDP itself will also run as part of the cloud-function service; please reach to us on the Slack community for additional support.
Single Star PDP
In cases where you don't expect to have a lot of data and you can secure a strong (CPU and memory) machine to host the PDP you can even use one PDP, in a star topology, for your entire deployment. This is not very common practice, but still legitimate. When possible prefer to have the various services / apps on the same physical node as the PDP, or close to it as possible to minimize latency.