What is an Interceptor?

It doesn’t take very long to bump into the word "Interceptor" when you’re working with Pedestal. They’re the most important construct in the library, the essential unit-of-work that virtually everything else in Pedestal is built upon.

What You Will Learn

After reading this guide, you will be able to:

  • Explain the role of interceptors in Pedestal

  • Compare interceptors to middleware

  • Create interceptors that affect execution of remaining interceptors

  • Understand how to handle errors

Interceptor Goals

Pedestal exists to support the following goals:

  • Enable code reuse, especially for cross-cutting concerns.

  • Keep code as functional — side-effect free — as possible.

  • Ensure that code can be easily tested.

  • Support asynchronous request flows.

Comparison to Ring

Many developers come to Pedestal from an earlier Clojure web framework, Ring; Ring has a very simple model:

  • A servlet container receives an HTTP request and creates a request map

  • A handler is passed the request map, and returns a response map

  • The response map is used to generate an HTTP response

Cross-cutting concerns, such as logging and authentication, are handled in Ring via middleware. Middleware is a special class of function with a particular purpose. A middleware function accepts a handler function, wraps it with new behavior, and returns a new handler function.

For example, a middleware function could parse the :body key of the request (which starts as a Java InputStream) and pass it though a JSON parser, replacing :body with the parsed content. Any handlers downstream of that middleware can expect the :body key to be parsed JSON.

Because a Ring handler is a stack of functions that invoke functions[1] there’s no room to support asynchronous flows. The essence of an asynchronous flow is that a function can return early, and later, resume from where it left off - a significant challenge for a language that doesn’t support this concept, such as Java.

As we’ll see shortly, interceptors provide a different model where the return early and resume later is more explicit.

Core Definition

An interceptor provides one step in the processing of an overall request. Interceptors are formed into a pipeline; each interceptor is invoked to do its one step, then control returns to Pedestal, which then invokes the next interceptor in the pipeline.

An interceptor might be specific to a particular route, or it might be more generic and wide-ranging, such as an interceptor that logs requests, or enforces authentication.

Even saying "one step" is not exactly correct; Pedestal breaks request processing into two phases: enter and leave.

The enter phase is related to processing of the incoming request; this covers cross-cutting concerns such as logging and authentication, but also routing, and route-specific interceptors and handlers.

The leave phase begins once a response has been created and represents an "unwinding", in much the same way that a stack of function calls eventually unwind towards the initial caller.

There’s also an error phase, which occurs when an exception is thrown by an interceptor; this allows other interceptors to report or record the exception, and generate a useful response; the error phase is really a subset of the leave phase.

In practical terms, an interceptor is an instance of the Interceptor record — a specialized Clojure map. This map has keys :name, :enter, :leave, and :error.

You rarely create an Interceptor directly, instead you pass a map, or some other value[2] to the interceptor function and an Interceptor record is returned [3].

An Interceptor should have a :name, and must have at least one of :enter, :leave, and :error.

:name

A keyword used (especially in logging) to identify the interceptor; generally, this is a namespace qualified keyword.

:enter

A callback function, passed the context map, returns a possibly modified context map.

:leave

A callback function, passed the context map, returns a possibly modified context map.

:error

A callback function, passed the context map and an exception, returns a possibly modified context map.

Why not a Protocol?
This may look like a situation where a Clojure Protocol would make sense. However, a Protocol would have to define all four parts (a method corresponding to each of :name, :enter, :leave, and :error). This would not only be cumbersome, but would prevent some optimizations that occur inside Pedestal.

So the context map flows through each interceptor; first through the enter phase, then later through the leave phase:

interceptors

Queue and Stack

The previous diagram shows a single interceptor, but the whole point here is that the interceptors work in a pipeline:

interceptor stack

This pipeline is composed of two parts: the interceptor queue and the interceptor stack. These are both stored inside the context map.

The queue is a list of interceptors that are waiting to execute in the enter phase. The stack is a list of interceptors that have already executed - it is used by the leave and error phases.

As an interceptor is popped of the queue, it is pushed onto the stack. When execution switches from the enter phase to the leave (or error) phase, the logic changes to popping each successive interceptor off the stack to invoke the :leave or :error callback.

Transition from Enter to Leave

Not all interceptors in the queue will execute: any interceptor can attach a response map (as key :response) to the context. That switches gears to the leave phase.

Likewise, simply running out of interceptors in the enter phase will switch to the leave phase.

Dynamic Queue

Earlier we said that interceptors are the "unit of work" for Pedestal. Even routing is one unit of work; a router interceptor works by matching an incoming request to a route, and the route provides a list of interceptors which are simply added to the queue.

It is perfectly acceptable for other interceptors to extend the interceptor queue in just the same way.

Sharing Information between phases

Remember that an interceptor’s :enter callback is executed to completion before subsequent interceptor’s are invoked; this means that any local symbol assigned via a let is out of scope for the later interceptors.

If an interceptor needs to record information during the enter phase and access it again during the leave phase, there is no recourse but to add that data to the context map.

A simple example is a timing interceptor that tracks how long it take to process a request:

(def timing-interceptor
  (interceptor
    {:name ::timing (1)
     :enter (fn [context]
              (assoc context ::start-ms (System/currentTimeMillis))) (2)
     :leave (fn [context]
              (let [{::keys [start-ms]} context (3)
                    elapsed-ms (- (System/currentTimeMillis) start-ms)]
                (log/debug :elapsed-ms elapsed-ms) (4)
                (dissoc context ::start-ms)))})) (5)
1 Give the interceptor a unique, namespace qualified name.
2 Update the context with a new key, namespace qualified to avoid collisions.
3 Destructuring trick, put :keys into a namespace to destructure using that namespace.
4 A side effect, but also the point of this interceptor.
5 Good hygiene is to dissoc anything that was previously added by the interceptor.

Asynchronous Results

Any callback also has the option to work asynchronously; this is quite simple: return a core.async channel that will eventually convey the updated context map rather than simply returning the updated context map.

Most often, going asynchronous is simply a matter of using the clojure.core.async/go macro:

(def user-data-interceptor
  (interceptor
    {:name ::user-data
     :enter (fn [context]
                (go (1)
                  (let [db (:db context) (2)
                        user-id (get-in context [:request :query-parameters :user-id])
                        user-ch (db/get-user db user-id)] (3)
                    (assoc context :user (<! user-ch))))) (4)
     :leave #(dissoc % :user)}))
1 A go block returns a channel that conveys the result.
2 Assumption is that some other interceptor put the :db key into the context.
3 db/get-user is asynchronous and returns a channel that conveys the user.
4 <! waits for the result from user-ch, which is then applied to the context. A simple key is used since other interceptors need access to this user data.

When an interceptor returns a channel, Pedestal will return the request-processing thread to the servlet container, so that it can be used to process other incoming requests. It will then wait for the channel to convey the new context, and continue from there. All remaining interceptors for the request will execute inside a thread from core.async's dispatch thread pool.

This only a thumbnail sketch; it doesn’t address likely scenarios such as what if the user doesn’t exist? What if there’s a database failure? What if reading the user from the database takes a really long time?

Error Handling

When an interceptor throws an exception, the exception is caught by Pedestal, and the pipeline shifts to the error phase.

In the error phase, Pedestal works up the stack of previously executed interceptors.

For interceptors with an :error callback, that callback is passed the context map and the exception.

The callback can:

  1. Return the context map; the exception will be ignored and Pedestal will switch to the leave phase.

  2. Attach the exception to the context map to allow a different intercept to handle the exception.

  3. Throw a new exception if unable to handle the original exception.

Further details are in the error handling reference.

Other Uses for Interceptor Pipelines

One of Pedestal’s core values is to create flexible utilities and use them in specific ways. One example of this approach is that the interceptor pipeline is not inherently tied to an HTTP request/response cycle; that is a specific application of the pipeline, setup via the Servlet Interceptor.

The same approach could be used for any number of other purposes, including:

  • Handling messages sent to a JMS queue or Kafka topic

  • Transforming a document

  • Sending an outgoing HTTP request and processing the response

The main differences are:

  • What data is stored in the initial context map

  • What are the termination conditions of the enter phase (via the terminate-when function)

Where To Next?

More details are available in the reference section:


1. Ring has, more recently, itself started to support asynchronous flows and WebSockets.
2. For example, a handler function in a route will be wrapped as an interceptor.
3. In most cases, even this step is performed for you, such as interceptors that are referenced as part of a route