FeatureHub’s original and primary architecture reflects its focus on it being a streaming update platform, one where updates to feature values are streamed out to listening clients in near-realtime.
Since its release, other patterns for use have emerged and we are, with the 1.5.0 release, making a few additions and alterations to match these expectations.
With the 1.5.9 release we have moved to using Cloud Events and supporting an increased number of async layers - NATS (preferred), Google PubSub (Release), AWS Kinesis (beta).
FeatureHub is available as a bundle (for streaming this is the Party Server) or as individual pieces for better scalability and isolation of your Administration side from your applications that require their features.
The extended streaming deployment is designed to scale to millions, even billions of requests, while mostly isolating your backend system from being overwhelmed by requests. With release 1.5.9 we have moved to the preferred "Dacha2" Lazy Cache system, and deployed it conceptually looks like this:
|Communication between Edge and Dacha(Cache) is shown via REST on this image, which can be configured optionally. By default, it is via NATS.|
The non-streaming platform (Party-Server-Ish) is designed to scale to less - tens of thousands, possibly more if you have a limited number of environments, or a larger number of read replicas. It is also designed to be much simpler and cheaper to deploy on environments like Google Cloud Run or Azure Container Instances. Deployed, conceptually it looks like this:
The way that FeatureHub is architected is designed for various different implementation sizes and scales, but fundamentally there is a separation of concerns of all the main components, so they can be scaled independently as and when needed.
We discuss the main deployment options of FeatureHub in the installation section and what each part is for.
This is the main admin server and is the source of truth for the application. All users login here (via local, OAuth2 or SAML), all portfolios, applications, environments, groups, features, etc are all controlled via this. This is always bundled with a UI and backend server and is configured to talk to some external database.
If MR server goes down, it won’t affect the operation of end-user clients, all their data is in the cache (or in the database if you use party-server-ish or edge-rest).
The "Admin" API is defined in an OpenAPI schema and can be generated for a wide variety of platforms. We currently include generated clients for Dart, Typescript, C# and Java, but it is not limited to these.
NATS is the Cloud Native Open Source messaging platform that has been around for a very long time, is very fast and is very adept at scaling to huge volume in a hugely distributed fashion. We use it for FeatureHub to transfer environments, features and service accounts around the network to feed Dacha and Edge.
Dacha is where the data that is required by every SDK is cached, and you need at least one of these for an operational FeatureHub system. It can be run in-process (using the Party Server design), or separately. Edge always talks to Dacha which holds permissions, environments, features, and pre-calculated etags for appropriate requests.
There are two choice for Dacha: Dacha1 and Dacha2 (Dacha2 is available from v1.5.9).
It must use NATS as it relies on features only NATS has
When it starts it completely fills its internal cache, either from another NATS or via the MR. This makes it completely isolate your servers from MR, no deliberate "miss" traffic can impact your Management Repository
Edge is able to talk to Dacha ove NATS
Filling its internal cache can take some time with hundreds or thousands of environments, and MR must be available for it to do so, so it can lead to a complicated start for a new k8s cluster or rollout. This can delay it from being healthy depending on how fast it can fill its cache, which can lead to operational complexity.
Dacha2 is introduced in 1.5.9 and exists to support multiple async layers. It is a lazy cache:
it supports multiple async layers (NATS, Google Pub/Sub, AWS Kinesis (beta), we are looking at others)
it is Cloud Events first
it caches misses as well as hits to ensure consistent misses do not make it to MR
it automatically updates itself as new environments, features, and service account changes are broadcast from MR, so a newly created environment will be a "cache hit" by default.
Edge is intended to be where the communication with the SDKs live. It is intended to be high volume endpoint but retain little data - only who is connected to it and which environment they are listening to for feature updates. Access to Edge is given by a combination of Service Account and Environment IDs (the API key). That combination is given a permission structure back in MR, and is usually simply READ. For test accounts, a service account can also have the ability to change features as it may need to while doing end-to-end tests.
It does not attempt to retain the active feature list for each Service Account + Environment. It is highly multithreaded and concentrates requests to Dacha.
It is expected that you will normally run at least two of these in any kind of environment.
Edge-REST provides only GET and PUT (for updating features for tests) API options. It allows the SDK to poll for updates but not get realtime updates, and will talk directly to the database. It can be deployed on its own or as part of party-server-ish.
If you would like to serve features faster globally or would like to cache feature flags on CDN, FeatureHub has affiliated with Fastly - real-time content delivery network. Integration with Fastly can save costs on deployment infrastructure and make the FeatureHub application stack considerably faster around the world.
We can provide you with the environment variables, and the configuration steps necessary to integrate Fastly with FeatureHub. Pulumi configuration, which can be translated easily into Terraform is also available on demand. Please contact us on email@example.com for further information.