The joyful act of decoupling your application plugin from its deployment will make your hapi app more portable and simpler to test.
hapi offers a rich API for splitting your application into plugins. If the first piece of advice to someone getting started with hapi is, "implement your application in a plugin" then the second would have to be "keep your plugins and your server separate." Why, you ask? Read on to find out!
While "plugin" can mean all sorts of different things in different frameworks, a hapi plugin represents a robust component or cross-cutting concern within your web application. And importantly, hapi will ensure that plugins snap together nicely without any conflicts—they're hapi's modular building-blocks.
A hapi plugin is a composable package of routes, server methods, cache provisions, request lifecycle extensions, authentication schemes/strategies, etc.; just about any piece of the hapi server
API can be called within the confines of a plugin. Multiple plugins may then be registered onto a server and deployed together.
As a result of this ability to "componentize" of an application, plugins also allow us to abstract away the deployment of the services that comprise our applications. Each plugin should be decoupled as a self-contained unit that can actually be deployed on its own. Plugins with this property are desirable, as they tend to live tidily within a greater application, are more reusable, simpler to test, and may be pulled into their own deployments if they need to scale independently. Summarily, organizing an application into plugins with well-defined boundaries is a relatively low-cost activity that provides us with lots of flexibility in deployment and development, and additionally reflects good organizational practices (i.e. separation of concerns).
It is common to write an application in the form of plugins, but still couple those plugins implicitly to the server on which they will be deployed. This tendency undermines the portability and reusability of the plugin, typically in two ways,
- The configuration of the server/deployment is mixed with the configuration of the plugins.
- The server/deployment is seen as a fixed hub for plugins, encouraging inter-plugin dependencies and reliance on plugin registration order.
As a result separations are crossed precisely where we hoped they wouldn't be—at plugin boundaries. If our plugin isn't portable or reusable, then it might as well not have been a plugin at all. Each plugin should live a modular life of its own!
The solutions here are pretty straightforward. Let's tackle our bad habits one at a time.
The first step is to untangle our plugin's configuration. The only input to our plugin from user-land are the plugin's options
—the ones that are passed to the plugin's register
function,
// Pinger plugin
module.exports = {
name: 'my-pinger',
register(server, options) {
// Below a plugin option is used to specify if the pinger route
// should support just GET requests, or both POST and GET requests
server.route({
method: options.supportPost ? ['get', 'post'] : 'get',
path: '/pinger',
handler: (request, h) => {
return { ping: 'pong' };
}
});
}
};
// And here's what it looks like to register the pinger
// plugin with options to an existing hapi server
await server.register({
register: MyPinger,
options: { supportPost: true }
});
Get used to turning every piece of dynamic configuration for your plugin into a plugin option. If there's a config.js
file somewhere in your project that's used by all your application plugins, or by a plugin and also your server
's deployment, then it's likely that your plugin is becoming tangled via configuration. The solution is simply to redirect that configuration into your plugin through its registration options.
An example of making this adjustment can be seen below. Observe that the plugin email-plugin.js
ends-up entirely self-contained, no longer sharing dependencies or configuration with server.js
.
module.exports = {
port: process.env.PORT || 3000,
email: {
worldLeader: 'president@whitehouse.gov',
from: 'no-reply@hapipal.com',
creds: {
host: 'smtp.gmail.com',
port: 465,
secure: true,
auth: {
user: 'paldo@hapipal.com',
pass: process.env.SMTP_PASSWORD
}
}
}
};
const Nodemailer = require('nodemailer');
-const Config = require('./config');
module.exports = {
name: 'email-plugin',
register(server, options) {
// Configure nodemailer
- const mailer = Nodemailer.createTransport(Config.email.creds);
+ const mailer = Nodemailer.createTransport(options.creds);
// Add a route to send an email to a configurable world leader
server.route({
method: 'post',
path: '/email/world-leader',
handler: async (request) => {
await mailer.sendMail({
- from: Config.email.from,
- to: Config.email.worldLeader,
+ from: options.from,
+ to: options.worldLeader,
subject: 'We have some thoughts for you!',
text: request.payload.text
});
return { message: 'sent' };
}
});
}
};
const Hapi = require('@hapi/hapi');
const EmailPlugin = require('./email-plugin');
const Config = require('./config');
(async () => {
const server = Hapi.server({ port: Config.port });
await server.register({
- register: EmailPlugin
+ register: EmailPlugin,
+ options: Config.email
});
await server.start();
console.log(`Server started on ${server.info.uri}`);
})();
See how it would now be simpler to test the email plugin with various different configurations, or to pull it out into a separate deployment? Just register the same plugin with different options
. It feels good!
A related issue to sharing configuration across plugins is sharing resources, such as a database connection, DB-connected models (e.g. Objection ORM, Mongoose, etc.), or a library that has been configured with credentials (e.g. an SDK such as aws-sdk). It is often desirable to share resources in this way in order to centralize credentials/configurations or maintain resource pools (e.g. a fixed number of database connections).
It's common to require()
those shared resources directly in each of your plugins, which is a form of coupling between your server and plugin particularly if those resources require configuration/credentials. It also may create implicit, hard-to-recognize coupling among your plugins.
There are a few ways to counter this anti-pattern.
-
Pass resources around as plugin options
This is the simplest solution, and it should work for a broad variety of use-cases. Configure your resources at the server/deployment level, and simply pass those resources into your plugins through their options.
-
Place shared resources on
server.app
hapi actually has a special place to put run-time server state that is available to all plugins within the application:
server.app
. You can place your resources here once they're ready to be used. A nice way to impose some order here is to create a plugin whose job it is to receive config/credentials as plugin options, create and configure the resource, then place the resource onserver.app
during server initialization (onPreStart
). Then other plugins have the guarantee that the resource is ready to use precisely after the server has started/initialized. -
Make shared resources available via server and request decorations
Similar to the previous pattern using
server.app
, a plugin can also create server or request decorations viaserver.decorate()
to expose a shared resource. Request decorations are particularly convenient when accessing those resources in route handlers. Advanced techniques using realms, when combined with decorations, can allow resources to be either shared or sandboxed within a single plugin—see schwifty as a good example of this, configuring knex instances and database-connected models at both the server and plugin levels under a robust concept of ownership.
Pro tip! However your plugin gets access to its shared resources, use
server.bind()
to allow those resources to show-up conveniently onh.context
and as thethis
context within server methods, route handlers, and server/request extensions defined by your plugin.
While mitigating inter-plugin dependencies is beyond the scope of this article, it's certainly related to the separation of server and plugin. Luckily we have an entire separate article precisely on that topic! Check it out—"Handling plugin dependencies".
hapi pal offers a boilerplate with a strong separation of server and plugin. It's a great way to author a plugin that is deployable and testable on its own, but prepared to be integrated into a larger application when the time comes. See the server/
directory for all things "deployment" and the lib/
directory for all things "plugin." It also provides flavors that maintain this separation while sharing resources, such as a textbook integration with Objection ORM, incorporation of Swagger, and a fancy templated site using browserify and SASS.
Once you've started writing your applications in terms of hapi plugins, do yourself a favor and take advantage of the wonderful benefits of pluginization: flexibility of deployment, simpler testing, tidy separation of your application's concerns, reusable server functionality, and a clear path to service-oriented architecture and microservices. The first step is to ensure you don't couple your server with its deployment. And it's really not so hard—you primarily want to separate deployment configuration and plugin configuration by utilizing plugin options liberally. We covered some examples of that, and several techniques to share resources among the plugins that comprise your application. Have fun plugin-izing!