As a developer, one of your tasks is decomposing an application into logically coherent, reusable, loosely-coupled components that can be understood and tested in isolation. Another task is coordinating these components -- composing them in a way that’s coherent, such that the system as a whole remains comprehensible and it’s possible to grow, debug, and maintain the application with minimal confusion.
donut.system is a depdendency injection library that helps you manage this source of complexity. It helps you in the following ways:
- It provides system and component abstractions that give your application a comprehensible structure
- It provides a means of composing components
- It provides a structure for defining component behavior
- It makes loose coupling possible
- It makes it easier to test your system
- It aids in understanding the scope of your system and how the pieces interact
Examples of components you might need in your system include:
- A database threadpool
- A messaging or queueing system
- A job scheduler
Something that these components have in common is that they have stateful behaviors, claiming resources like threads and reading and writing to other resources. They could also have dependencies that determine the order in which they're started and stopped: your job scheduler might use your database as its data store, and therefore can't be started until after your db threadpool is created. donut.system makes sure that these behaviors happen in the correct order.
Table of Contents
- Basic Usage
- Advanced usage
- Purpose
- Objections
- Alternatives
- Why use this and not that?
- Composing systems
- Creating multiple instances of groups of components
- Acknowledgments
- Status: alpha
- Community
- TODO
To use donut.system, you first define a system that contains component groups. Component groups contain component definitions. Component definitions include signal handlers that specify component behaviors.
Here's an example of a system definition:
(ns donut.examples.single-component
(:require
[donut.system :as ds]))
(def system
{::ds/defs ;; <-- components defined under this key
{:app ;; <-- component group name
{:printer ;; <-- component name
;; ::ds/start and ::ds/stop are signal handlers
#::ds{:start (fn [_]
(future
(loop []
(println "hello!")
(Thread/sleep 1000)
(recur))))
:stop (fn [{:keys [::ds/instance]}]
(future-cancel instance))}}}})NOTE: donut.system makes heavy use of namespaced keywords. If the
#::ds{:start ...}syntax above is new to you, please read this doc.
This example defines a system var (the name system is arbitrary). Its value
is a map that has one key, ::ds/defs. This is where your component definitions
live. Systems are implemented as maps that contain the ::ds/defs key.
The value of ::ds/defs is a map, where the keys are names for component
groups. In this case, there's only one component group, :app. :app is an
arbitrary name with no special significance; you can use whatever keywords you
want for component group names.
Under the :app component group we have a map of where each key is the name of
the component and each value is the component's definition. A component
definition specifies the component's behavior. In this example, the :printer
component definition is a map that has two keys, ::ds/start and ::ds/stop.
These keys are names of signal handlers, which you'll learn about momentarily.
::ds/start and ::ds/stop are both associated with a function. These
functions are where you specify a component's behavior.
Let's actually interact with this system and see its behavior:
(let [running-system (ds/signal system ::ds/start)]
(Thread/sleep 5000)
(ds/signal running-system ::ds/stop))If you run the above in a REPL, it will print "hello!" once a second for five
seconds and then stop. The function ds/signal takes a system as its argument
and "sends" the given signal (::ds/start) to the components in the system,
calling the corresponding signal handler function. This signal and send
terminology is metaphorical; there's no network or sockets or anything like that
involved.
The return value of a signal handler becomes the component's instance. A
component instance is typically some object that you can use to stop the
component. (ds/signal system ::ds/start) returns an updated system map that
includes component instances. If you send another signal to the updated system
map, it can use those instances. In the example above, we call (ds/signal running-system ::ds/stop) to send the ::ds/stop signal, and its signal
handler cancels the future returned by the ::ds/start signal handler.
Let's look at a slightly more complicated example. This system has two
components, a :printer component and a :stack component. When the system
receives the :donut.system/start signal, the :printer pops an item off the
:stack and prints it once a second:
(ns donut.examples.printer
(:require [donut.system :as ds]))
(def system
{::ds/defs
{:services
{:stack #::ds{:start (fn [{:keys [::ds/config]}]
(atom (vec (range (:items config)))))
:config {:items 10}}}
:app
{:printer #::ds{:start (fn [opts]
(let [stack (get-in opts [::ds/config :stack])]
(future
(loop []
(prn "peek:" (peek @stack))
(swap! stack pop)
(Thread/sleep 1000)
(recur)))))
:stop (fn [{:keys [::ds/instance]}]
(prn "stopping")
(future-cancel instance))
:config {:stack (ds/ref [:services :stack])}}}}})
;; start the system, let it run for 5 seconds, then stop it
(comment
(let [running-system (ds/signal system ::ds/start)]
(Thread/sleep 5000)
(ds/signal running-system ::ds/stop)))As before, system is a map that contains just one key, ::ds/defs.
::ds/defs is a map of component groups, of which there are two: :services
and :app. The :services group has one component definition, :stack, and
the :app group has one component definition, :printer.
Component definitions can contain ::ds/start and ::ds/stop signal handlers,
as well as a ::ds/config. The :printer component's :ds/config contains a
ref to the :stack component. You'll learn more about refs below; they allow
one component to refer to and use another component.
You start the system by calling (ds/signal system ::ds/start). This produces an
updated system map (bound to running-system) which you then use when stopping
the system with (ds/signal running-system :stop).
The rest of this README covers donut.system's pieces in more detail.
Components have definitions and instances.
A component definition (component def or just def for short) is an entry in
the ::ds/defs map of a system map. A component definition can be a map, as
this system with a single component definition shows:
(def Stack
#::ds{:start (fn [{{:keys [items]} ::ds/config}] (atom (vec (range items))))
:config {:items 10}})
(def system {::ds/defs {:services {:stack Stack}}})Components are organized under component groups. I cover some interesting
things you can do with groups below, but for now you can just consider them an
organizational aid. This system map includes the component group :services.
(Note that there's no special reason to break out the Stack component
definition into a top-level var. I just thought it would make the example more
readable.)
A def map can contain signal handlers, which are used to create component instances and implement component behavior. A def can also contain additional configuration values that will get passed to the signal handlers.
In the example above, we've defined a ::ds/start signal handlers. Signal
handlers are just functions with one argument, a map. What is included in this
map?
This map includes the key ::ds/config, and its value is taken from the
::ds/config key in your component definition. In the example above, that means
that the map will contain {:items 10}. You can see that the ::ds/start
signal handler destructures ::ds/config out of its first argument, and then
looks up :items.
(Other key/value pairs get added to the signal handler's map, and I'll cover those as we need them.)
This approach to defining components lets us easily modify them. If you want to
mock out a component, you just have to use assoc-in to assign a new
::ds/start signal handler.
Signal handlers return a component instance, which is stored in the system map
under ::ds/instances. Try this to see a system's instances:
(::ds/instances (ds/signal system :start))This is how you can access component instances for tests.
Component instances are added to the signal handler's argument under the
::ds/instance key. When you apply the ::ds/start signal to a Stack
component, it creates a new atom, and when you apply the ::ds/stop handler the
atom is passed in under ::ds/instance key. In the example above, the
::ds/stop signal handler destructures this:
(fn [{::ds/keys [instance]}] (reset! instance []))This is how you can allocate and deallocate the resources needed for your
system: the ::ds/start handler will create a new object or connection or
thread pool or whatever, and place that in the system map under
::ds/instances. The ::ds/stop handler can retrieve this instance, and it can
then call whatever functions or methods are needed to to deallocate the
resource.
You don't have to define a handler for every signal. Components that don't have a handler for a signal are essentially skipped when you send a signal to a system.
Component defs can contains refs, references to other components that resolve to that component's instance when signal handlers are called. Let's look at our stack printer again:
(def system
{::ds/defs
{:services
{:stack #::ds{:start (fn [{:keys [::ds/config]}]
(atom (vec (range (:items config)))))
:config {:items 10}}}
:app
{:printer #::ds{:start (fn [opts]
(let [stack (get-in opts [::ds/config :stack])]
(future
(loop []
(prn "peek:" (peek @stack))
(swap! stack pop)
(Thread/sleep 1000)
(recur)))))
:stop (fn [{:keys [::ds/instance]}]
(prn "stopping")
(future-cancel instance))
:config {:stack (ds/ref [:services :stack])}}}}})The last line includes {:stack (ds/ref [:services :stack])}. ds/ref is a
function that returns a vector of the form [:donut.system/ref component-key],
where component-key is a vector of the form [group-name component-name].
These refs are used to determine the order in which signals are applied to
components. Since the :printer refers to the :stack, we know that it depends
on a :stack instance to function correctly. Therefore, when we send a
:start signal, it's handled by :stack before :printer.
Within :printer's :start signal handler, stack refers to the atom created
by the :stack component.
When you call (ds/signal system ::ds/start), the following happens:
- The
::ds/startsignal handler for[:services :stack]gets called. It returns an atom, which becomes the component instance for[:services :stack]. - Internally, that atom is added to the system map under
[::ds/instances :services :stack]. - The
::ds/startsignal handler for[:app :printer]gets called with a single argument, a map. That map includes the key path[::ds/config :stack], and its value is the component instance for[:services :stack]-- the atom created at step 1.
If you have a component [:group-a :component-a] whose instance is a map like
{:level-1 {:level-2 {:level-3 ...}}} then you can refer to values within the
map with a ref like (ds/ref [:group-a :component-a :level-1 :level-2 :level-3]).
Note that a ref must be reachable for it to be resolved, meaning that it must
be possible to use (get-in system [::ds/defs :path :to :ref]) to retrieve the
ref. Something like this wont' work:
{::ds/defs {:app {:printer #::ds{:start (fn [_] (ds/ref [:services :stack]))}}}}It won't work because you ds/ref resides inside a function definition that
isn't reachable by get-in.
If a component is defined using any value other than a map that contains the
:donut.system/start key, that value is considered to be the component's
instance. This can be useful for configuration. Consider this system:
(ns donut.examples.ring
(:require [donut.system :as ds]
[ring.adapter.jetty :as rj]))
(def system
{::ds/defs
{:env {:http-port 8080}
:http {:server #::ds{:start (fn [{{:keys [handler options]} ::ds/config}]
(rj/run-jetty handler options))
:stop (fn [{::ds/keys [instance]}]
(.stop instance))
:config {:handler (ds/local-ref [:handler])
:options {:port (ds/ref [:env :http-port])
:join? false}}}
:handler (fn [_req]
{:status 200
:headers {"ContentType" "text/html"}
:body "It's donut.system, baby!"})}}})The component [:env :http-port] is defined as the value 8080. It's referred
to by the [:http :server] component. When the [:http :server]'s :start
handler is applied, it destructures options from its first argument. options
will be the map {:port 8080, join? false}.
This is just a little bit of sugar to make it easier to work with donut.system. It would be annoying and possibly confusing to have to write something like
(def system
{::ds/defs
{:env {:http-port #::ds{:start (constantly 8080)}}}})We've seen how you can specify signal handlers for components, but what is a
signal? The best way to understand them is behaviorally: when you call the
ds/signal function on a system, then each component's signal handler gets
called in the correct order. I needed to convey the idea of "make all the
components do a thing", and signal handling seemed like a good metaphor.
Using the term "signal" could be misleading, though, in that it implies the use of a communication primitive like a socket or a semaphor. That's not the case. Internally, it's all just plain ol' function calls. If I talk about "sending" a signal, nothing's actually being sent. And anyway, even if something were getting sent, that shouldn't matter to you in using the library; it would be an implementation detail that should be transparent to .
donut.system provides some sugar for built-in signals: instead of calling
(ds/signal system ::ds/start) you can call (ds/start system).
There's a more interesting reason for the use of signal, though: I want signal handling to be extensible. Other component libraries use the term lifecycle, which I think doesn't convey the sense of extensibility that's possible with donut.system.
Out of the box, donut.system recognizes ::ds/start, ::ds/stop,
::ds/suspend, and ::ds/resume signals, but it's possible to handle arbitrary
signals -- say, :your.app/validate or :your.app/status. To do that, you just
need to add a little configuration to your system:
(def system
{::ds/defs {;; components go here
}
::ds/signals {:your.app/status {:order :topsort}
:your.app/validate {:order :reverse-topsort}}})::ds/signals is a map where keys are signal names and values are configuration
maps. The configuration keys are:
:order values can be :topsort or :reverse-topsort. This specifies the
order that components' signal handlers should be called. :topsort means that
if Component A refers to Component B, then Component A's handler will be called
first; reverse is, well, the reverse.
:returns-instance? this determines whether the return value of the signal
handler should be used to update the system's instances, under ::ds/instances.
The map you specify under ::ds/signals will get merged with the default signal
map, which is:
(def default-signals
"which graph sort order to follow to apply signal, and where to put result"
{::start {:order :reverse-topsort
:returns-instance? true}
::stop {:order :topsort
:returns-instance? true}
::suspend {:order :topsort
:returns-instance? true}
::resume {:order :reverse-topsort
:returns-instance? true}
::status {:order :reverse-topsort}})Systems organize components and provide a consistent way to initiate component behavior. You send a signal to a system, and the system ensures its components handle the signal in the correct order.
As you've seen, systems are implemented as maps. I sometimes refer to these maps
as system maps or system states. It can be useful, for example, to think of
ds/signal as taking a system state as an argument and returning a new state.
donut.system follows a pattern that you might be used to if you've used interceptors: it places as much information as possible in the system map and uses that to drive execution. This lets us do cool and useful stuff like define custom signals.
One day I'd like to write more about the advantages of taking the "world in a map" approach. In the mean time, this Lambda Island blog post on Coffee Grinders does a good job of explaining it.
donut.system/named-system is a multimethod you can use to register system
maps. This can be useful for defining dev, test, and prod systems:
(defmethod ds/named-system :test
[_]
{::ds/defs ...})Often you'll want to customize a config; you'll want to replace a component with
a mock, for example. You can pass an additional argument to ds/system to
specify overrides:
(ds/system :test {[:services :queue] mock-queue})You don't have to override an entire component. You can also override just a signal handler:
(ds/system :test {[:services :queue ::ds/start] (fn mock-start-queue [_])})Overrides are a map where keys are def paths, and values are whatever value
you want to be assoc'd in to that path under ::ds/defs. The above code is
equivalent to this:
(update (ds/named-system :test)
::ds/defs
(fn [defs]
(reduce-kv (fn [new-defs path val]
(assoc-in new-defs path val))
defs
{[:services :queue :start] (fn mock-start-queue [_])})))The signal helpers ds/start, ds/stop, ds/suspend, and ds/resume can take
either a system name or a system map, and can take optional overrides:
(ds/start :test) ;; <- system name
(ds/start {::ds/defs ...}) ;; <- system map
;; use named system, with overrides
(ds/start :test {[:services :queue] mock-queue})The start helper also takes an optional third argument to select a subset of components start:
(ds/start :test
{[:services :queue] mock-queue}
#{[:app :http-server]}) ;; <- component selectionComponent selection is explained below.
The donut.system.repl namespace has conveniences for REPL workflows. By
default, it will start and stop a named-system named :donut.system/repl, but
you can also specify a system:
(require '[donut.system :as ds])
(require '[donut.system.repl :as dsr])
;;---
;; By default, the named-system :donut.system.repl is used
(defmethod ds/named-system :donut.system/repl
[_]
{::ds/defs {:group {:component {::ds/start (fn [_] (println "starting :donut.system/repl"))
::ds/stop (fn [_] (println "stopping :donut.system/repl"))}}}})
(dsr/start)
;; => starting :donut.system/repl
(dsr/stop)
;; => stopping :donut.system/repl
;; you can still override components
(dsr/start {[:group :component ::ds/start] (fn [_] (println "override"))})
;; => override
;;---
;; You can also use a different named-system
(defmethod ds/named-system :dev
[_]
{::ds/defs {:group {:component {::ds/start (fn [_] (println "starting :dev"))
::ds/stop (fn [_] (println "stopping :dev"))}}}})
(dsr/start :dev)
;; => starting :dev
(dsr/stop)
;; => stopping :dev
;; you can still override components
(dsr/start :dev {[:group :component ::ds/start] (fn [_] (println "override dev"))})
;; => override devdonut.system.repl/restart will:
- Stop the running system
- Call
(clojure.tools.namespace.repl/refresh :after 'donut.system.repl/start)
This will reload any changed files and then start your system again.
You can use the library beholder to watch your file system for changes and automatically reload changes and restart your system while you're developing it. Here's how I do it:
First, create the file dev/src/user.clj and put this in it:
(ns user)
(defn dev
"Load and switch to the 'dev' namespace."
[]
(require 'dev)
(in-ns 'dev)
:loaded)Then create dev/src/dev.clj and put this in it:
(ns dev
{:clj-kondo/config {:linters {:unused-namespace {:level :off}}}}
(:require
[clojure.tools.namespace.repl :as nsrepl]
[dev.repl :as dev-repl]
[donut.system :as ds]
[donut.system.repl :as dsr]
[donut.system.repl.state :as dsrs]
[fluree.http-api.system :as sys])
(:refer-clojure :exclude [test]))
(nsrepl/set-refresh-dirs "dev/src" "src" "test")
(defn routes
[]
(get-in dsrs/system [::ds/defs :env :http/routes]))
(def start dsr/start)
(def stop dsr/stop)
(def restart dsr/restart)
(defmethod ds/named-system :donut.system/repl
[_]
(ds/system :dev))
(when-not dsrs/system
(dsr/start))Next create dev/src/dev/repl.clj and put this in it:
(ns dev.repl
(:require [clojure.tools.namespace.repl :as repl]
[donut.system.repl :as dsr]
[nextjournal.beholder :as beholder]))
(repl/disable-reload!)
(defonce persistent-state (atom {}))
(defn- source-file? [path]
(re-find #"(\.cljc?|\.edn)$" (str path)))
(defn- restart*
[path]
(when (source-file? path)
(try
(dsr/restart)
(catch Exception e
(println "Exception reloading:")
(println e)))))
(defn- restart [ns]
(fn [{:keys [path]}]
(binding [*ns* ns]
(restart* path))))
(def watcher
(beholder/watch (restart *ns*) "src" "resources" "dev/src" "test"))
(comment
(beholder/stop watcher))merge this configuration into your deps.edn file:
{:aliases
{:dev
{:extra-paths ["dev/src" "test"]
:extra-deps {com.nextjournal/beholder {:mvn/version "1.0.0"}
org.clojure/tools.namespace {:mvn/version "1.1.0"}}}}}By "merge" I mean that if you already have a :dev alias, add the values to it
in a way works for your project.
Once you've done this, you start a REPL with the :dev alias. If you use emacs,
you can add the following to your .emacs.d to have CIDER always include the dev
alias for REPLs:
(setq cider-clojure-cli-aliases ":dev")After the REPL has started, call the (dev) function from the user namespace,
which is the default namespace. Calling (dev) will load the dev namespace
and switch to it, then start your system. It will also get beholder to do its
thing, watching the filesystem and reloading your namespaces and restarting your
system.
As you develop your project, it's likely an exception will get thrown when you're trying to start your system. This can cause some resources to be claimed without an obvious way to recover them. For example, your system might start an HTTP server on port 8080, then throw an exception, leaving you without a clear way to stop the HTTP server.
You can try to stop a failed system with the function
donut.system/stop-failed-system. Here's its source:
(defn stop-failed-system
"Will attempt to stop a system that threw an exception when starting"
[]
(when-let [system (and *e (::system (ex-data *e)))]
(stop system)))If you're trying to start a system using donut.system.repl/start, it will
automatically try to stop a failed system if an exception gets thrown.
Where do you actually put your donut.system-related code? And how do you handle configuration?
I recommend creating a your-project.system namespace to define your base system. It
might look something like this:
(ns you-project.system
(:require
[aero.core :as aero]
[clojure.java.io :as io]
[donut.system :as ds]
[ring.adapter.jetty :as rj]))
;; Use aero for all configuration
(defn env-config [& [profile]]
(aero/read-config (io/resource "config/env.edn")
(when profile {:profile profile})))
;; define all behavior in base-system
(def base-system
{::ds/defs
{:env {}
:http
{:server
#::ds{:start (fn [{{:keys [handler options]} ::ds/config}]
(rj/run-jetty handler options))
:stop (fn [{::ds/keys [instance]}]
(.stop instance))
:config {:handler (ds/ref [:http :handler])
:options {:port (ds/ref [:env :http-port])
:join? false}}}
:handler
#::ds{:start (fn [_]
;; handler goes here
)}}}})
(defmethod ds/named-system :base
[_]
base-system)
(defmethod ds/named-system :dev
[_]
(ds/system :base {[:env] (env-config :dev)}))
(defmethod ds/named-system :donut.system/repl
[_]
(ds/system :dev))
(defmethod ds/named-system :test
[_]
(ds/system :dev
{[:http :server] ::disabled}))Note that this system contains an :env group. Other components can reference
values in the :env group for their configuration. The [:http :server]
component does this for its port.
Additionally, refs can "reach" farther into the referenced component. For example, this would work:
(def base-system
{::ds/defs
{:env
{:http {:port 8080}}
:http
{:server
#::ds{:start (fn [{{:keys [handler options]} ::ds/config}]
(rj/run-jetty handler options))
:config {:handler (ds/ref :handler)
:options {:port (ds/ref [:env :http :port])
:join? false}}}}}})Note the second-to-last-line includes (ds/ref [:env :http :port]) - this will
correctly reference the HTTP port.
As your system grows, you'll probably want to move components into separate namespaces. Your system map might then look something like this:
(def base-system
{::ds/defs
{:env {}
:http
{:server http/server
:handler http/handler}}})How do you test an application that uses donut.system? There are three main concerns:
- Starting and stopping your system
- Accessing component instances
- Mocking components
Let's look at each, using this test system:
(defmethod ds/named-system ::test
[_]
{::ds/defs
{:group-a
{:component-a
{::ds/start (fn [_] (atom []))}}
:group-b
{:component-b
{::ds/start (fn [opts]
;; add an element to the `[:group-a :component-a]` atom on
;; start
(swap! (get-in opts [::ds/config :component-a])
conj
:foo))
::ds/config {:component-a (ds/ref [:group-a :component-a])}}}}})There are three main options you can choose from to start and stop your system:
(deftest your-test
(let [system (ds/start ::test)]
(is (= [:foo]
@(get-in system [::ds/instances :group-a :component-a])))
(ds/stop system)))The donut.system namespace has a dynamic var, *system*, and a macro that
handles some of the machinery of working with it:
(deftest using-with-*system*
(ds/with-*system* ::test
(is (= [:foo]
@(get-in ds/*system* [::ds/instances :group-a :component-a])))))The macro's first argument is either a system map or a system name. The macro
will start the system and bind the started system map to ds/*system*. It will
also stop the system.
The function ds/system-fixture returns a function that can be used as a
clojure.test fixture:
(use-fixtures :each (ds/system-fixture ::test))
(deftest using-fixture
(is (= [:foo]
@(get-in ds/*system* [::ds/instances :group-a :component-a]))))Just be careful not to mix this method with method 2. If you do that you'll end up starting two different systems, and that could cause hard-to-debug problems.
Once you have a started system, you can access component instances under the
system's ::ds/instances key. You can also use the function ds/instance:
(deftest retrieving-instances
(ds/with-*system* ::test
;; one way to retrieve an instance
(is (= [:foo]
@(get-in ds/*system* [::ds/instances :group-a :component-a])))
;; another way to retrieve an instance
(is (= [:foo]
@(ds/instance ds/*system* [:group-a :component-a])))))The advantage of using ds/instance is that it will throw an exception if
you're trying to get an instance for an undefined component, which can help you
catch typos.
When you're writing tests, you'll sometimes want to mock out components. For
example, if you have an Amazon SQS queue, you might want to mock out the client rather
than trying to connect to an actual SQS queue over the network. When you use the
ds/start or ds/system functions, you can provide a map of component
overrides, as covered above in the config helpers section.
Here's what that might look like:
(deftest with-override
;; method 1
(let [test-atom (atom [])]
(ds/start ::test {[:group-a :component-a] test-atom})
(is (= [:foo] @test-atom)))
;; method 2 - the first argument to `ds/with-*system*` can be either a system
;; name or a system map. In this example we're getting a system map.
(let [test-atom (atom [])]
(ds/with-*system* (ds/system ::test {[:group-a :component-a] test-atom})
(is (= [:foo] @test-atom)))))The topics covered so far should let you get started defining components and systems in your own projects. donut.system can also handle more complex use cases.
All component definitions are organized into groups. As someone who compulsively lines up pens and straightens stacks of brochures, I think this extra level of tidiness is inherently good and needs no further explanation.
The inclusion of component groups unlocks some useful capabilities that are less obvious, though, so let's talk about those. Component groups make it easier to:
- Create multiple instances of a component
- Send signals to selections of components
- Designate system stages
I'll describe what I mean by "multiple instances" here, and I'll explain the rest in later sections.
Let's say for some reason you want to run multiple HTTP servers. Here's how you could do that:
(ns donut.examples.multiple-http-servers
(:require
[donut.system :as ds]
[ring.adapter.jetty :as rj]))
(def HTTPServer
#::ds{:start (fn [{{:keys [handler options]} ::ds/config}]
(rj/run-jetty handler options))
:stop (fn [{::ds/keys [instance]}]
(.stop instance))
:config {:handler (ds/local-ref [:handler])
:options {:port (ds/local-ref [:port])
:join? false}}})
(def system
{::ds/defs
{:http-1 {:server HTTPServer
:handler (fn [_req]
{:status 200
:headers {"ContentType" "text/html"}
:body "http server 1"})
:port 8080}
:http-2 {:server HTTPServer
:handler (fn [_req]
{:status 200
:headers {"ContentType" "text/html"}
:body "http server 2"})
:port 9090}}})First, we define the component HTTPServer. Notice that it has two refs,
(ds/local-ref [:handler]) and (ds/local-ref [:port]). These differ from the
refs you've seen so far, which have been created with ds/ref. Refs created
with ds/local-ref are, well, local refs, and will resolve to the component
of the given name within the same group.
This little sprinkling of abstraction creates more possibilities for component modularity and reuse. You could create multiple instances of an HTTP server without groups, sure, but it would be more tedious and typo-prone. The fact is, some components actually are part of a group, so it makes sense to have first-class support for groups.
The system function takes an optional third argument that lets you specify
what components you want to use:
(ds/system :named-system {} #{[:group-1 :component-1]})
(ds/system
{;; first argument can also be a system map
}
{}
#{[:group-1 :component-1]})The purpose of specifying components like this is to limit what components receive signals. This might come in handy in testing, where you might want to work with only a subset of all system components.
When you select components, the entire subgraph of component dependencies get selected too; you don't have to include all those dependencies in your selection. For example with this:
(ds/signal (ds/system :test {} #{[:group-1 :component-1]}) ::ds/start)The ::ds/start signal gets sent to the component [:group-1 :component-1] as
well as all the components it depends on.
You can also select component groups by using just the group's name for your selection, like so:
(ds/system system {} #{:group-1})It might be useful to signal parts of your system in stages. For example, you might want to instantiate a logger and error reporter and use those if an exception is thrown when starting other components:
;; This is mostly pseudocode
(def system
{::ds/defs
{:boot {:logger #::ds{:start ...
:stop ...}
:error-reporter #::ds{:start ...
:stop ...}}
:app {:server #::ds{:start ...}}}})
(let [booted-system (ds/start system {} #{:boot})
logger (get-in booted-system [::ds/instances :boot :logger])
error-reporter (get-in booted-system [::ds/instances :boot :error-reporter])]
(try (ds/signal booted-system :start)
(catch Exception e
(log logger e)
(report-error error-report e))))Note that you would need to make the ::ds/start handlers for :logger and
:error-reporter idempotent, meaning that calling ::ds/start on an
already-started component should not create a new instance but use an existing
one. The code would look something like this:
(fn [{::ds/keys [config instance]}]
(or instance
(create-logger config)))The select-components function takes two arguments, a system and a set of
component-ids. It returns a new system with component selection noted, so that
when you send signals to the new system the signals are only sent to the selected
components and the components they depend on (recursively):
(ds/select-components system #{[:group-a :component-a] [:group-b :component-b]})If you call ds/start on this, then only [:group-a :component-a] and
[:group-b :component-b] will receive the start signal, as well as all the
components they depend on.
If you include a keyword in the selected components set, like
(ds/select-components system #{:boot}), then all components in that group will
be selected.
The ds/start function can optionally take a set of selected components as a
third argument.
If you want to remove the component selection, you can either dissoc the key
::ds/selected-components from your system map or call select-components with
nil: (ds/select-components system nil)
You can define pre- and post- handlers for signals:
(def system
{::ds/defs
{:app {:server #::ds{:pre-start (fn [_] (prn "pre-start"))
:start (fn [_] (prn "start"))
:post-start (fn [_] (prn "post-start"))}}}})You can use these lifecycle handlers to gather information about your system as it handles signals, and to perform validation. Let's look at a couple use cases: printing signal progress and validating configs.
Here's how you might print signal progress:
(defn print-progress
[{::ds/keys [system]}]
(prn (::ds/component-id system)))
(def system
{::ds/defs
{:group {:component-a #::ds{:start "component a"
:post-start print-progress}
:component-b #::ds{:start "component b"
:post-start print-progress}}}})
(ds/signal system ::ds/start)
;; =>
[:group :component-a]
[:group :component-b]The function print-progress is used as the :post-start handler for both
:component-a and :component-b. It destructures ::ds/system, then prints
(::ds/component-id system).
That's right: signal handlers are passed the entire system under the
::ds/system key of their argument. The current component's id gets assoc'd
into the system map under ::ds/component-id prior to calling a signal handler.
The handler argument also has a collection of "channel" functions merged into it
which we can use to gather information about components and perform validation.
Look at how we destructure ->info and ->validation from the third argument
in these :post-start handlers:
(def system
{::ds/defs
{:group {:component-a #::ds{:start "component a"
:post-start (fn [{:keys [->info]}]
(->info "component a is valid"))}
:component-b #::ds{:start "component b"
:post-start (fn [{:keys [->validation]}]
(->validation "component b is invalid"))
;; This `:config` is only here to create the
;; dependency order for demonstration purpose
:config {:ref (ds/ref :component-a)}}
:component-c #::ds{:start "component-c"
:post-start (fn [_]
(prn "this won't print"))
;; This `:config` is only here to create the
;; dependency order for demonstration purpose
:config {:ref (ds/ref :component-b)}}}}})
(::ds/out (ds/signal system ::ds/start))
;; =>
{:info {:group {:component-a "component a is valid"}},
:validation {:group {:component-b "component b is invalid"}}}Notice that :component-c's :post-start handler doesn't get called. As it
predicts, the string "this won't print" doesn't get printed.
It's not obvious what's going on here, so let's step through it.
-
:component-a's:post-startgets called first. It destructures the->infofunction out of the third argument.->infois a channel function and its purpose is to allow signal handlers to place a value somewhere in the system map in a convenient and consistent way.->infoassoc'd into the system map before a signal handler is called, and it closes is over the "output path", which includes the current component id. This is why when you call(->info "component a is valid"), the string"component a is valid"ends up at the path[::ds/out :info :group :component-a]. -
(->info "component a is valid")returns a system map, and that updated system map is conveyed forward to other components' signal handlers, until a final system map is returned byds/signal.But what if you want to use
:post-startto perform a side effect? What then?? Do these functions always have to return a system map?No. The rules for handling return values are:
- If a system map is returned, convey that forward
- Otherwise, check whether the signal handler is flagged as returning an
instance. This is configured under
[::ds/signals :signal-name :returns-instance?]. If that value is true, use the return value to update the instance value. - Otherwise, ignore the return value.
-
(->validation "component b is invalid")is similar to->infoin that it places a value in the system map. However, it differs in that it also has implicit control flow semantics: if at any point a value is placed under[::ds/out :validation], then the library will stop trying to send signals to that component's descendants. (It's actually a little more nuanced than that, and I cover those nuances below.)
One way you could make use of these features is to write something like this:
(ns donut.examples.validate
(:require
[donut.system :as ds]
[malli.core :as m]))
(defn validate-config
[{:keys [->validation ::ds/config]}]
(when-let [schema (:schema config)]
(when-let [errors (m/explain schema (dissoc config :schema))]
(->validation errors))))
(def system
{::ds/defs
{:group {:component-a #::ds{:pre-start validate-config
:start "component a"
:config {:schema [:map [:foo any?] [:baz any?]]}}
:component-b #::ds{:pre-start validate-config
:start "component b"
:config {:schema [:map [:foo any?] [:baz any?]]}}
:component-c #::ds{:start "component-c"}}}})We can create a generic validate-component function that checks whether a
component's definition contains a :schema key, and use that to validate the
rest of the component definition.
You can add ::ds/base key to a system map to define a "base" component
definition that will get merged with the rest of your component defs. The last
example could be rewritten like this:
(ns donut.examples.validate
(:require
[donut.system :as ds]
[malli.core :as m]))
(defn validate-config
[{:keys [->validation ::ds/config]}]
(when-let [schema (:schema config)]
(when-let [errors (m/explain schema config)]
(->validation errors))))
(def system
{::ds/base #::ds{:pre-start validate-config}
::ds/defs
{:group {:component-a {:start "component a"
:schema [:map [:foo any?] [:baz any?]]}
:component-b {:start "component b"
:schema [:map [:foo any?] [:baz any?]]}
:component-c {:start "component-c"}}}})Sometimes you don't want a component to stop and start every time a system restarts. For example, if you have a threadpool component, you don't want to tear it down and recreate it constantly. A couple scenarios where this isn't desirable:
- You've set up a reloaded REPL workflow and don't want to restart your threadpool every time you save a file
- You're starting and stopping a system for every test, and don't want to restart that threadpool between tests
To cache a component, pass its def to the ds/cache-component function. This
test demonstrates:
(deftest caching
(reset! ds/component-instance-cache {})
(let [counter (atom 0)
system {::ds/defs
{:group
{:component (ds/cache-component
{::ds/start (fn [_] (swap! counter inc))
::ds/stop (fn [_] (swap! counter + 10))})}}}]
(ds/start system)
(is (= 1 @counter))
(ds/stop system)
(is (= 1 @counter))
(ds/start system)
(is (= 1 @counter))
;; if you clear the cache then the stop signal will go through
(reset! ds/component-instance-cache {})
(ds/stop system)
(is (= 11 @counter))))One of donut.system's overarching goals is to provide a foundation for a richer ecosystem of composable libraries so that an application developer can easily integrate some vertical slice of functionality with minimal fiddling. The plugin system is meant to provide a clear interface for this kind of extension.
To use a plugin, add it to a vector under ::ds/plugins in your system map:
{::ds/defs {}
::ds/plugins [some-plugin]}I want it to be easy to understand what a plugin has done to your system. Right
now, the function donut.system.plugin can take a system as an argument and
produce descriptions of how each plugin has modified the system.
Plugins modify a system map, adding or modifying values. They're defined as maps with the following keys:
:donut.system.plugin/name
A keyword
:donut.system.plugin/doc
Not currently used, but this is where a docstring goes
:donut.system.plugin/system-defaults
This gets merged with a system via (merge system-defaults system), meaning
that any values in your system map take precedence over those in the plugin.
:donut.system.plugin/system-merge
This gets merge with a system via (merge system system-merge), meaning that
plugin values will take precedence over those already in the system.
:donut.system.plugin/system-update
This is a function that takes a system as an argument and returns a new system. For cases where you need some extra logic in updating a system definition.
Example plugin definition:
(def test-harness-plugin
{:donut.system.plugin/name
::test-harness-plugin
:donut.system.plugin/doc
"Configures system so that donut.endpoint.test.harness can find the
components needed to construct and dispatch requests."
:donut.system.plugin/system-defaults
{::ds/registry {:donut/endpoint-router [:routing :router]
:donut/http-handler [:http :handler]}
::ds/defs {::config {:default-request-content-type :transit-json}}}})Woe be unto you if you ever have to compose a system from subsystems. But if you do, I've tried to make it straightforward. Check it out:
(ns donut.examples.subsystem
(:require [donut.system :as ds]))
(defn mk-print-thread
[prefix stack]
(doto (Thread.
(fn []
(prn prefix (peek @stack))
(swap! stack pop)
(Thread/sleep 1000)
(recur)))
(.start)))
(defn print-worker-system
[print-prefix]
{::ds/defs
{:workers
{:print-worker #::ds{:start (fn [{{:keys [stack]} ::ds/config}]
(mk-print-thread print-prefix stack))
:stop (fn [{::ds/keys [instance]}]
(.stop instance))
:config {:stack (ds/ref [:services :stack])}}}}})
(def system
{::ds/defs
{:services {:stack #::ds{:start (fn [_] (atom (vec (range 20))))
:stop (fn [{::ds/keys [instance]}] (reset! instance []))}}
:printers {:printer-1 (ds/subsystem-component
(print-worker-system ":printer-1")
#{(ds/ref [:services])})
:printer-2 (ds/subsystem-component
(print-worker-system ":printer-2")
#{(ds/ref [:services :stack])})}}})In this example, we're creating two subsystems ([:printers printer-1] and
[:printers :printer-2]) that pop items from a shared stack component defined
in the parent system, [:services :stack].
We generate definitions for the subsystems with the function
print-worker-system, which returns a system definition with one component,
[:workers :print-worker]. The component def has a key, :stack, which
references [:services :stack], but notice that there is no [:services :stack] component in the print-worker-system definition.
Internally, the parent system wraps these subsystems with a call to
ds/subsystem-component. ds/subsystem-component returns a component def, a
map with a ::ds/start signal handler that "forwards" the signal to the
subsystem. The component def also includes the key ::ds/mk-signal-handler, a
privileged key that acts as default signal handler. ::ds/mk-signal-handler is
responsible for forwarding all other signals to the subsystem.
ds/subsystem-component takes an optional second argument, a set of refs that
should be imported into the subsystem. This is how the subsystems can reference
the parent system's component [:services :stack].
Now that we've covered how to use the library, let's talk about why you'd use it.
When building a non-trivial Clojure application you're faced with some questions that don't have obvious answers:
- How do I write code that's understandable and maintainable?
- How do I manage resources like database connections and thread pools?
- How do I manage test environments?
donut.system helps you address these problems by giving you tools for encapsulating behavior in components and composing components into systems.
We can make application code more understandable and maintainable by identifying a system's responsibilities and organizing code around those responsibilities so that they can be considered and developed in isolation - in other words, defining a system architecture an implementing it with healthy doses of loose coupling and encapsulation.
It's not obvious how to do implement and convey your system's architecture in a functional programming language like Clojure, where it's pretty much one giant pool of functions, and boundaries (namespaces, marking functions private) are more like swim lanes you can easily duck under than walls enforcing isolation.
Using a component library like donut.system is one way for you to introduce such boundaries. When you program with components, you clarify your application's functional concerns, you codify (literally!) the relationships between different parts of your system, and you make the interfaces between them explicit. You avoid creating a codebase where any random function can access any random state - part of why you got into Clojure in the first place.
Components facilitate writing loosely-coupled code. The benefits of that are well documented, but I'll briefly mention a couple here:
- Loosely-coupled code is easier to understand because it reduces the scope of the system you have to have in your head to understand what something is doing.
- Loosely-coupled code is easier to maintain because it reduces the scope of impact from changes.
Components also aid discoverability. A system definition serves as a map that outlines the major "territories" of functionality, as well the entry point to each.
donut.system helps allocate and deallocate resources like database connections and thread pools in the correct order. It also provides a systematic approach to accessing resources. When building an application, you have to manage these tasks somehow; a component library like donut.system gives you the tools to manage them in a consistent way.
I have a half-baked thought about component libraries serving a purpose similar to tools like systemd, though in a much more limited scope. I'm not sure exactly where you want to go with it, but: component libraries are useful in building an application for reasons similar to why systemd is useful in managing a machine. In both cases, you want some consistent method for starting and stopping the actors in a computing environment. This work is not central to whatever business problem you're trying to solve, but it still has to get done, so it's nice to be able to use a tool that does that work for you that you can learn once and use across different projects.
donut.system (and other component libraries) provide a kind of light-weight virtual environment for your application. Usually there's one-to-one relationship between a running process and a running application; component systems make it possible to run many instances of an application within a single process.
The biggest benefit this brings is the ability to run dev and test systems at the same time. I can start a dev system with an HTTP server and a dev db connection from the REPL, and from the same REPL run integration tests with a separate HTTP server and db connection. It's a huge workflow improvement.
donut.system's component definitions are just data, which means that it's possible for libraries to provide components that work with donut.system without actually including a donut.system dependency. A library like cronut, for example, could include the following map for easy consumption in a donut.system project:
(def CronutComponent
:donut.system{:start (fn [{:donut.system/keys [config]}] (initialize config))
:stop (fn [{:donut.system/keys [instance]}] (shutdown instance))})What if you want to define a component group without depending on donut.system? You might want to do this if you have a collection of related components that have local refs to each other. Here's how you could do that:
(def CoolLibComponentGroup
{:component-a #:donut.system{:start (fn [_] ...)}
:component-b #:donut.system{:start (fn [{{:keys [component-a]} :donut.system/config}])
:config {:component-a [:donut.system/ref :component-a]}}})The key is that refs are represented with the vector [:donut.system/ref ref-key].
Whether or not this is actually a good idea remains to be seen, but my hope is that it will provide a better foundation for writing higher-level, composable libraries.
Over the years, I've encountered two main objections to this approach:
- It forces premature abstraction
- It's too complex
TODO address these concerns. (They're not necessarily wrong!)
Other Clojure libraries in the same space:
I cover how donut.system compares to the alternatives in docs/rationale.org.
TODO
TODO
donut.system takes inspiration from Component, Integrant, and Clip.
This library has been used in production but is not widely used. The interfaces may change, but change is unlikely.
PRs welcome! Also check out the #donut channel in Clojurians Slack if you wanna chat or if you have questions.
- async signal handling
- more examples
- discuss the value of dependency injection