
In the realm of system design and architecture, how backend services communicate with each other hold the key to delivering seamless functionality to users. Internally within a company, this can be done through shared databases, remote procedure calls, or queue messaging systems.

However, providing functionality often involves tapping into external services provided by other companies. A prime example is the use of payment acceptance endpoints, typically APIs accessible over HTTPS, offered by payment service providers (PSPs). PayPal, Adyen, and Stripe are some examples of PSPs which require API consumers to talk to their services through the APIs they provide.

But how do we ensure this communication remains robust, even in the face of faults, timeouts or downtime? Designing APIs and communication streams effectively becomes crucial, especially in cases where speed is of utmost importance because it affects the experience of the users of the system. For this discussion, let's consider several things:

- endpoint: a specific place where a web service can both receive requests and respond
- resource: data which we want to perform actions upon, for example a hotel booking
- state: there may be a number of state which the resource is currently in. The actions and their statuses govern the possible transitions between states.
- action: the act of doing something on a resource, for example creating a hotel booking
- status: the current result of the action on a resource - typically there are 3 statuses for each action - pending, success and failed.

In the internet experience of today, we expect webpages to load within the order of 500ms, and websites to be stay responsive to human interactions. However, not all actions can complete within seconds, and sometimes it may even take longer than stipulated connection time limits set along the communication chain. We wouldn't want to hit a timeout error and be left hanging wondering whether the action was completed successfully or not. As service providers, setting up endpoints requires thoughtful planning.

## Restaurant Booking Service Example

Let's take the scenario of designing a restaurant booking service as an example with the following requirements:

- resource: the resource here is a single restaurant booking
- action: actions on the resource can include: create, cancel, update

The states of the booking resource can be something like this:

<center>
<img src="/assets/fsm.drawio.svg" style="width: 50%; min-width: 500px" />
</center>

Time to time, service consumers may send in requests via the `PUT /booking/{bookingCode}` endpoint to create a booking:

<center>
<img src="/assets/seq01.drawio.svg" style="width: 50%; min-width: 500px" />
</center>

Let's say the service level agreement was for the create booking endpoint to return a status within 5 seconds. If a final status was reached within 5 seconds, the response of the create action from the booking service would tell whether the creation was a success or whether it failed.

If the booking creation takes longer than 5 seconds, then the booking service can respond with a "pending" status of the "create" action. Subsequently, the upstream services can retry sending the request again. The `bookingCode` would then serve as an idempotency key, and would ensure that no duplicates of the booking resources would be created from multiple instances of the same request identified by the `bookingCode`.

## Getting Details about the Resource

Another way that upstream services can learn about the current state of the resource is to ask booking service at the `GET /booking/{bookingCode}` endpoint about the current state about the resource.

<center>
<img src="/assets/seq02.drawio.svg" style="width: 50%; min-width: 500px" />
</center>

When and how frequent to poll the endpoint is a key design consideration when building the services.

## Providing Latest Knowledge About Resource to Upstream

To also ensure timely updates on status changes, a callback or webhook design can be factored into the design. This setup facilitates communication on state changes on the resource back to the service consumer, essentially providing a downstream to upstream synchronization of data. This allows information exchange between your service and the service consumer, even if either one party faces a temporary downtime.

In the FSM above, you probably noticed that there's a transition defined from the `ready` state to `archived` when the date of the booking is one day before the current date/time. Perhaps there is a background task that would clean up all bookings in the previous day and move them from `ready` state to `archived`.

In such scenario, the upstream services wouldn't know about the state change, unless the booking service provides an update in the form of a callback.

<center>
<img src="/assets/seq03.drawio.svg" style="width: 50%; min-width: 500px" />
</center>

This robust setup ensures that regardless of any temporary hiccups in communication, crucial information about actions like creating a booking can still be transferred between service provider and consumer when services are back online. The design isn't just about making things work — it's about making sure they work reliably.



Backend services are the procesing blocks that empower the logic of an application served well. They do all sorts of things: talk to databases, keep things in caches, process data from message queues, serve webhooks for external systems, process images, or even send push notifications to users' mobile devices - indispensible part to a well designed software architecture, whether for a web or mobile app.

To make services more reusable and being able to change behaviour without changing code, configuration parameters are typically provided to tell the services what kind of environment it is operating in. For example:

- `NODE_ENV=production`: to inform that a service is running in production environment
- `DB_HOST, DB_PORT`: which database server to connect to
- `DB_USERNAME, DB_PASSWORD`: what user account to use when talking to database
- `JWT_SECRET`: what JWT shared secret to use
- `REQ_TIMEOUT_LIMIT=5000`: what timeout in milliseconds to use when sending requests to other services

Sometimes backend services deployment fail way past deployment because of misconfiguration. It is possible to enter the wrong `DB_HOST` hostname, missed out on setting the value for `JWT_SECRET` or even have a typo in the key or value of any configuration parameters. As services grow larger, we tend not to encounter these errors early on until the particular section of code gets executed under some specific conditions.

For example, if your database connection uses lazy loading and/or connection pooling, chances are that your backend service doesn't establish a connection to the database at start up (or shortly after deployment) until it is needed while serving some request much later.

> We end up celebrating too early at a successful deployment only to find the services failing while handling workload.

## Enter Power-on Self-test

In the world of hardware and embedded systems, it is typical for engineers to embed a set of self-testing routines called [Power-on Self-test](https://en.wikipedia.org/wiki/Power-on_self-test) (POST). Simply put, when the system starts up

Power-on Self-test is much more common than we know it. They are commonly found in

- airplanes
- medical equipment
- computers
- phones
- gaming consoles
- TVs

> Can you imagine yourself in a plane up in the air, only to find the plane stalling mid-air because of fuel amount was in reserves during take-off?

Could the plane's software system checked on the amount of fuel remaining before allowing the pilot to start the engine?

Likewise, to build resilient backend services, it's important to include some form of POST routine at service pre-initialization. By allowing backend services to fail early on shortly after deployment, services wouldn't fail while serving requests or processing important data.

In an Express.js HTTP API service, I'd write a POST routine like this:

```typescript
// post.ts

/// .... imports goes here

const { JWT_SECRET } = process.env;

export async function powerOnSelfTest() {
  // check if JWT secret was configured and can be used to sign and verify payload
  assert(JWT_SECRET);
  const token = jwt.sign("testData", JWT_SECRET);
  const payload = jwt.verify(token, JWT_SECRET);
  assert(payload === "testData");

  // check if database connection and credentials are OK
  const dbConn = createDbConnection();
  try {
    await dbConn.authenticate();
  } catch (e) {
    console.error("Error connecting to DB", e);
    throw e;
  }
}
```

Before starting up the HTTP listener, I'd run the POST routine:

```typescript
// index.ts

/// .... other imports
import { powerOnSelfTest } from "./post";

async function startApp() {
  const app = express();

  const db = await initDbModels(createDbConnection());
  app.locals.db = db;

  /// ... setup routes

  const port = process.env.PORT || 3333;
  const server = app.listen(port, () => {
    console.log(`Listening at port ${port}`);
  });
}

// start up the application
powerOnSelfTest()
  .then(startApp)
  // catch POST and startup errors here
  .catch((err) => console.error(err));
```

Even if all the environment variables provided to the application were correct, the POST routine would at least check for connectivity between the service and dependent servers like the database. Problems like version incompatibility, network connectivity or incorrect infrastructure setup, can also be caught in the POST - _at very little cost_.

Having POST routines in place, I can be _slightly more_ confident that my backend services would work beyond post-deployment.

POST routines work as well as they are written. As [Edsger W. Dijkstra](https://en.wikipedia.org/wiki/Edsger_W._Dijkstra) puts it, POST routines - like any other tests - can only prove the presence, **not** the absence of bugs. You can instead embrace POST routines as another hammer in your toolbox in catching catch potential failures early on.


Sam's Blog

Posts tagged backend

Architecting Asynchronous State Synchronization between Backend Services

Backend Services: Power-on Self-test