Meltdown is a Clojure interface to Reactor, an asynchronous programming, event passing and stream processing toolkit for the JVM.
It follows the path of Romulan, an old ClojureWerkz project on top of LMAX Disruptor that's been abandoned.
- Provide a convenient, reasonably idiomatic Clojure API for Reactor
- Not introduce a lot of overhead
- Be well documented
- Be well tested
In addition to providing a core Reactor API for Clojure, Meltdown provides a DSL for stream processing.
Meltdown is young, although the API hasn't changed in a while. However, as Reactor itself is young, breaking API changes are not out of the question.
Meltdown artifacts are
released to Clojars. If you
are using Maven, add the following repository definition to your pom.xml:
<repository>
<id>clojars.org</id>
<url>http://clojars.org/repo</url>
</repository>With Leiningen:
[clojurewerkz/meltdown "1.1.0"]With Maven:
<dependency>
<groupId>clojurewerkz</groupId>
<artifactId>meltdown</artifactId>
<version>1.1.0</version>
</dependency>Meltdown uses Reactor mailing list. Feel free to join it and ask any questions you may have.
To subscribe for announcements of releases, important changes and so on, please follow @ClojureWerkz on Twitter.
Reactor is an event-driven programming toolkit for the JVM which
offers multiple features. At its core, however, it is a message
passing library, that's used to tie event publishers and consumers
together. Handlers can be attached to and detached from a reactor
dynamically. When handler is attached to reactor, selector is
used. Selectors determine what handlers will be invoked.
To start using Meltdown, first define a reactor
using clojurewerkz.meltdown.reactor/create:
(require '[clojurewerkz.meltdown.reactor :as mr])
(let [reactor (mr/create)]
;; your code
)You can subscribe to events triggered within reactor by using
clojurewerkz.meltdown.reactor/on:
(require '[clojurewerkz.meltdown.reactor :as mr]
[clojurewerkz.meltdown.selectors :refer [$]])
(mr/on reactor ($ "key") (fn [event]
(comment "Do something here")))($ "key") here is a selector. In essence, that means that every time
reactor receives event dispatched with the key "key", it will call your
handler.
In order to push events into reactor, use clojurewerkz.meltdown.reactor/notify:
(require '[clojurewerkz.meltdown.reactor :as mr])
;; you can pass any data structures/objects over
;; reactors
(mr/notify reactor "key" {:my "payload"})Selectors determine which consumers will be invoked for an event.
There are multiple types of selectors supported by Reactor. Meltdown primarily focuses on two types: exact match and regular expressions.
The exact match should be used for cases when you want handler to respond to a single key:
(require '[clojurewerkz.meltdown.reactor :as mr :refer [$]])
(mr/on reactor ($ "key") (fn [event] (do-one-thing event)))
(mr/on reactor ($ "key") (fn [event] (do-other-thing event)))
(mr/on reactor ($ "key") (fn [event] (do-something-else event)))
(mr/notify reactor "key" {:my :payload}) ;; will fire all three handlers
(mr/notify reactor "other" {:other :payload}) ;; will fire noneIn the example above all three handlers will receive the event.
Regular expression selectors are used whenever you want to match one or many event keys based on a pattern, for example:
(require '[clojurewerkz.meltdown.reactor :as mr :refer [R]])
(mr/on reactor (R "USA.*") (fn [event] (usa-handler event)))
(mr/on reactor (R "Europe.*") (fn [event] (europe-handler event)))
(mr/on reactor (R "Europe.Sw*") (fn [event] (sw-handler event)))
(mr/notify reactor "USA.EastCoast" {:teh :payload}) ;; will fire USA.*
(mr/notify reactor "Europe.Germany" {:das :payload}) ;; will fire Europe.*
(mr/notify reactor "Europe.Sweden" {:das :payload}) ;; will Europe.Sw* and Europe.* handlers
(mr/notify reactor "Europe.Switzerland" {:das :payload}) ;; will Europe.Sw* and Europe.* handlers
(mr/notify reactor "Asia.China" {:teh :payload}) ;; will fire noneMatch all selector unconditionally matches all keys:
(require '[clojurewerkz.meltdown.reactor :as mr :refer [match-all]])
;; will match all keys
(mr/on reactor (match-all) (fn) [event] ...)To produce a custom selector, reify reactor.event.selector.Selector
which looks like this:
package reactor.event.selector;
/**
* A {@literal Selector} is a wrapper around an arbitrary object.
*
* @author Jon Brisbin
* @author Stephane Maldini
* @author Andy Wilkinson
*/
public interface Selector {
/**
* Get the object being used for comparisons and equals checks.
*
* @return The internal object.
*/
Object getObject();
/**
* Indicates whether this Selector matches the {@code key}.
*
* @param key The key to match
*
* @return {@code true} if there's a match, otherwise {@code false}.
*/
boolean matches(Object key);
/**
* Return a component that can resolve headers from a key
*
* @return A {@link HeaderResolver} applicable to this {@link Selector} type.
*/
HeaderResolver getHeaderResolver();
}Whenever you have more than a single handler registered for a selector, you can define a routing strategy:
:firstrouting strategy will take a first handler whose selector matches the key:broadcastwill dispatch event to every handler whose selector matches the key:round-robinwill pick handlers on the round robin basis. For example, if there're three handlers, first event will get to first, second to second, third to third, fourth to first again, and so on.
You can chose a routing strategy that makes most sense for your
application. :first is usually used when there should be a guarantee
that a single, first-assigned handler should take care of
event. :broadcast makes sense when all handlers should get an event
simultaneously, and perform different actions. And :round-robin would
make sense for things like load-balancing, whenever you would like to
keep all workers equally busy, therefore giving the one that just
received an event a chance to take care of it before it gets the next one.
In order to chose a routing strategy, pass one of the mentioned values
to create function, for example:
(mr/create :event-routing-strategy :broadcast)There are several types of dispatchers, providing you a toolkit for both threadpool-style long-running execution to high-throughput task dispatching.
- default one, synchronous dispatcher, implementation that dispatches events using the calling thread.
:event-loopdispatcher implementation that dispatches events using the single dedicated thread. Together with default synchronous dispatcher, very useful in development mode.:thread-pooldispatcher implementation that usesThreadPoolExecutionwith an unbounded queue to dispatch events. Works best for long-running tasks.:ring-bufferdispatcher implementation that uses LMAX Disruptor RingBuffer to queue tasks to execute. Known to be most high-throughput implementation.
In order to create a reactor backed by the dispatched of your
preference, pass :dispatcher-type to create function, for example:
(mr/create :dispatcher-type :ring-buffer)It is possible to receive get a callback from the callee. In order to
implement request-response with callback, use send-event
receive-event pair of methods.
receive-event is used instead of notify for performance
reasons. It's an expensive operation to check for :respond-to field
for each event. Result of the handler execution will be passed back to
the caller.
For example, if you'd like to send an event with hello-key key, execute a
callback function whenever response is received, you can do it that way:
;; Result of handler execution will be passed back to callback in send-event
(mr/receive-event reactor ($ "hello-key") (fn [_] "response"))
;; Sends "data" to "hello-key" and waits for handler to call back
(mr/send-event reactor "hello-key" "data" (fn [event]
;; do job with response
))Two main concepts in stream processing are channel and stream. You
can publish information to channel, create arbitrary amount of streams
out of any channel or stream.
stream is a stateless event processor, that allows values that are
going through it to be filtered or changed. It's very easy to build
large processing graphs using this concept, since every mutation returns
another stream you can attach consumers to.
In order to compose streams, you can use clojurewerkz.meltdown.streams/map*, filter*, reduce*
and batch* functions. They have signatures similar to the ones you're
used to have in clojure. Applying map* with inc function on the
channel will create a new stream that contains all the values
incremented by one.
Stream processing in Reactor is (mostly) lazy. To begin feeding a channel
the values, you need to flush it with clojurewerkz.meltdown.streams/flush
first.
For example, let's create a channel where we push integers, and two streams with attached consumers that would calculate incremented and decremented values for incoming ints:
(require '[clojurewerkz.meltdown.streams :as ms :refer [create consume accept map*]])
(let [channel (create)
incremented-values (map* inc channel)
decremented-values (map* dec channel)]
(consume incremented-values (fn [i] (println "Incremented value: " i)))
(consume decremented-values (fn [i] (println "Decremented value: " i)))
(accept channel 1)
(accept channel 2)
(accept channel 3)
(ms/flush channel))
;; => Incremented value: 2
;; => Decremented value: 0
;; => Incremented value: 3
;; => Decremented value: 1
;; => Incremented value: 4
;; => Decremented value: 2You can also apply map* to streams that were already "mapped", reduced
filtered or batched:
(require '[clojurewerkz.meltdown.streams :as ms :refer [create consume accept map*]])
(let [channel (create)
incremented-values (map* inc channel)
squared-values (map* (fn [i] (* i i)) incremented-values)]
(consume squared-values (fn [i] (println "Incremented and squared value: " i)))
(accept channel 1)
(accept channel 2)
(ms/flush channel))
;; => Incremented and squared value: 4
;; => Incremented and squared value: 9filter* would filter values that go through it and only pass ones for
which predicate matches further:
(require '[clojurewerkz.meltdown.streams :as ms :refer [create consume accept filter*]])
(let [channel (create)
even-values (filter* even? channel)
odd-values (filter* odd? channel)]
(consume even-values (fn [i] (println "Got an even value: " i)))
(consume odd-values (fn [i] (println "Got an odd value: " i)))
(accept channel 1)
(accept channel 2)
(accept channel 3)
(accept channel 4)
(ms/flush channel))
;; => Got an odd value: 1
;; => Got an even value: 2
;; => Got an odd value: 3
;; => Got an even value: 4reduce* works pretty much same way as reduce in clojure works,
except for it gets values from the stream and holds last accumulator
value:
(require '[clojurewerkz.meltdown.streams :as ms :refer [create consume accept reduce*]])
(let [channel (create)
res (atom nil)
sum (reduce* #(+ %1 %2) 0 channel)]
(consume sum #(reset! res %))
(accept channel 1)
(accept channel 2)
(accept channel 3)
(ms/flush channel)
@res)
;; => 6For buffered operations, for example, when you'd like to have several
values batched together, and only bundled values are of interest for
you, you can use batch* that only emits values when buffer capacity is
reached and buffer is flushed.
If these four given operations are not enough for you and you'd like to
create a custom stream, it's quite easy as well. For that, there's a
custom-stream operation available. For example, you'd like to create
a stream that will only dispatch every 5th value further. For state,
you can use let-over-lambda:
(defn every-fifth-stream
"Defines a stream that will receive all events from upstream and dispatch
every fifth event further down"
[upstream]
(let [counter (atom 0)]
(custom-stream
(fn [event downstream]
(swap! counter inc)
(when (= 5 @counter)
(reset! counter 0)
(accept downstream event)))
upstream)))You can use custom streams same way as you usually use internal ones:
(let [channel (create)
res (atom nil)
incrementer (map* inc channel)
inst-every-fifth-stream (every-fifth-stream incrementer)]
(consume inst-every-fifth-stream #(reset! res %))
(accept channel 1)
(println @res)
(accept channel 2)
(println @res)
(accept channel 3)
(println @res)
(accept channel 4)
(println @res)
(accept channel 5)
@res)
;; => nil
;; => nil
;; => nil
;; => nil
;; => 6You can also build processing graphs in a declarative manner. For example, let's create a graph that will increment all incoming values, then aggregate a sum of them and put that sum to atom.
For these examples, you should use clojurewerkz.meltdown.stream-graph namespace
instead of usual streams one.
(ns my-stream-graphs-ns
(:use clojurewerkz.meltdown.stream-graph))
(let [res (atom nil)
channel (graph (create)
(map* inc
(reduce* #(+ %1 %2) 0
(consume #(reset! res %)))))]
(accept channel 1)
(accept channel 2)
(accept channel 3)
@res)
;; => 9If you see that your graph is too deeply nested, and you'd like to split it
in parts, you can use attach and detach functions. For example, here's
a graph that will increment each incoming value and calculate sums of
even and odd values separately:
(let [even-sum (atom nil)
odd-sum (atom nil)
even-summarizer (detach
(filter* even?
(reduce* #(+ %1 %2) 0
(consume #(reset! even-sum %)))))
odd-summarizer (detach
(filter* odd?
(reduce* #(+ %1 %2) 0
(consume #(reset! odd-sum %)))))
summarizer #(+ %1 %2)
channel (graph (create)
(map* inc
(attach even-summarizer)
(attach odd-summarizer)))]
(accept channel 1)
(accept channel 2)
(accept channel 3)
(accept channel 4)
@even-sum
;; => 6
@odd-sum
;; => 8
)Whenever an Exception occurs inside of your processing pipeline,
downstreams won't receive any events. In order to debug your
pipeline and see what exactly happened, you should subscribe
to the Exception events by using on-error.
Most generic case would be:
(mr/on-error r Exception (fn [event]
(println event)))So whenever any Exception (of any type) occurs inside your
processing pipeline, your handler will be executed. You can also use
sub-classes:
(mr/on-error r RuntimeException (fn [event]
(println event)))In this case your handler will be only triggered when type of an occurred
exception is RuntimeException or one of the derived classes.
You can have more than one handler per Exception type.
You can use reactor for all kinds of events in your systems. Event can
represent an error, alarm or some other piece of information another
part of system should be aware of. Reactor calls do not block, therefore
should be used in cases when you'd like to fire-and-forget the event.
Using reactive approach has it's advantages, for example, if you're responding to TCP socket connection, you can dispatch an event to reactor, continue listening and be certain that a worker will pick it up and take care of it.
You should be aware of the fact that if your handlers never finish (for example, there's an endless loop inside one of handlers), you'll eventually run out of available handlers, and won't be able to proceed.
In Meltdown, we support three basic primitives of functional composition, such as Monoid, Functor and Foldable. Monoids are used to build either sequences (lists, groups), searching for maximum or minimum, finding a sum of the stream and so on.
Monoid consists of 2 operation: mempty (sometimes called mzero) and
mappend. One of our monoid instances is a list:
(extend-protocol Monoid
clojure.lang.PersistentVector
(mempty [_] [])
(mappend [old new] (conj old new)))This monoid has a constructor mempty that returns an empty list, and
mappend which appends new elements of the stream to the list.
Another example of Monoid is MinValue:
(deftype MinValue [a]
clojure.lang.IDeref
(deref [_] a))
(extend-protocol Monoid
MinValue
(mempty [_] (MinValue. (Double/POSITIVE_INFINITY)))
(mappend [old new] (MinValue. (min @old new))))
;; (mconcat (MinValue. 1) [1 2 3 4])
;; => #<MinValue@45061e92: 1>MinValue is just a container for a value that it holds. mempty would
in that case be either an "empty value" (MinEmpty, if you wish), but for simplicity
I decided to construct it with Positive Infinity, since everything is less than positive
infinity. mappend operation finds a minimum of old and new value and puts it
back into MinValue context.
We use functors to distinquish between the ways we "go into" the context. Functor
has only 1 function in it's interface, and it's fmap. For example, want to run inc
(increment by 1) within a list, we would use map, that would apply our function to
the every element of the list. Same thing for our group (which is essentially a
hashmap of {group-name [group-value]}), we would have to apply this function to
each element of [group-value] for each group-name.
If we wanted to fun inc on MinValue, we would just extract the value from MinValue
context, run inc on it and pack it back into MinValue:
(extend-protocol Functor
MinValue
(fmap [v f]
(MinValue. (f @v))))
;; (fmap (MinValue. 5) inc)
;; => #<MinValue@1bf9ff6e: 6>So, Functor would be a convenient way apply functions in different contexts.
Whenever we have some context, we want to some way to fold it. Folding mostly makes sense for collection.
Inside meltdown, we support 2 types of folds by default:
- fold of list (which is essentially
reduce)
fold :: [a] -> ([a] -> b) -> b
- and fold on groups, which would reduce values of an entire group to a single value:
;; Pardon my pseudotypesignature
fold :: {a [b]} -> ([b] -> c) -> {a c}
You may wonder, what it gives us in terms of stream processing and why that even matters. Let's check out a simple example:
(graph (create :env env)
;; (1)
(map* (fn [i] [(if (even? i)
:even
:odd)
i])
;; (2)
(mappend* {}
;; (3)
#(= 5 (last (:odd %)))
;; (4)
(fmap* inc
;; (5)
(fold* +
(consume #(reset! res %)))))))Here we have several steps:
- (1) at first we're transforming our incoming numbers into a vector of
[<parity>, number]. For example, for 1 it would return[:odd 1]and[:even 2]for 2. - (2) these tuples are then passes into
mappend*, which is initialized with{}(group/hash) monoid. Since there on all our items will be collected into the hashmap, holding hash of{<parity> [<items>]}:
{:even [2 4], :odd [1 3 5]}- (3) here, we just specify a "flush" condition. So whenever we see "5", we flush. Although that may be as well any other logical predicate.
- (4) now, we want to apply
incto every element of each group. So, we apply inc to each element in each key-value pair:
{:even (map inc [2 4]), :odd (map inc [1 3 5])}
;; => {:even [3 5], :odd [2 4 6]}- (5) we can now use
foldto get sums of all elements for each group, obtaining our final result:
{:odd 12, :even 8}This was a simplifed example, but even here you may notice that having an ability to iterate over different colleciton types in different ways may help to simplify stream processing code and avoid boilerplate. By retaining semantics of operations (operation that executed on each item within the context, or item that folds all the values within the relevant context), you can build your own abstractions and make your code easier to write, understand, debug and extend.
Basic throughput tests are included. They're set up in the exact same way Reactor's throughput tests are done, therefore you can compare outputs directly. Overhead is insignificant.
Meltdown requires Clojure 1.6+.
Meltdown is part of the group of Clojure libraries known as ClojureWerkz, together with
Meltdown uses Leiningen. Make sure you have it installed and then run tests against supported Clojure versions using
lein all test
Then create a branch and make your changes on it. Once you are done with your changes and all tests pass, submit a pull request on GitHub.
Copyright (C) 2013-2016 Michael S. Klishin, Alex Petrov, and the ClojureWerkz team.
Double licensed under the Eclipse Public License (the same as Clojure) or the Apache Public License 2.0.