ReactiveJournal supports Reactive libraries by adding functionality to record and replay reactive streams.
ReactiveJournal is a Maven project so you can git clone
the project and build in the usual way.
The intention is for this project to make its way into Maven Central (work in progress).
Go to the releases section of the project. With each release there will be an uber jar that you can download with the ReactiveJournal classes and all dependencies.
Once downloaded you can test that it works by running:
java -cp ./reactivejournal-x.x.x.jar org.reactivejournal.examples.helloworld.HelloWorld
Picture this scenario...
You've just released your shiny new Reactive project. Streams of data start to flow into the system, your 'Reactives' are busy processing the data, enriching it, making decisions based on it, creating new streams of data for Reactives further down the line to consume. It's the well oiled engine of an F1 car, a thing of beauty, poetry in motion. Exactly what a Reactive system was designed to do.
But then...
Something starts going wrong. Perhaps data received isn't quite as clean as it should be, the system gets dirty, components clog up and fail, data backs up. Processes slow down, before long we have total melt down. Our F1 engine is a heap a smouldering metal.
But what went wrong... this is what we need to be able to our system:
-
Recreate the exact situation so that can step through it and understand where the root of the problem is.
-
Capture that scenario fix it and include it in a test case so that it never happens again.
-
Monitor the slowdown of the system by observing the queues build up. Recover the system without losing any data in the meantime.
-
Deal with the surge in data that caused the component to start runing out of memory.
-
Run multiple processes to process our data rather than relying on the one which went over.
-
All the above without slowing the system down or having to do any development.
Journaling addresses all these points. ReactiveJournal was designed to makes journaling for Reactive programs as simple as it could possible be.
Testing is a primary motivation for ReactiveJournal. ReactiveJournal allows developers to black box test their code by recording all inputs and outputs in and out of their programs.
An obvious use case are unit tests where ReactiveJournal recordings can be used to create
comprehensive tests (see HelloWorldTest for an example). This example makes use of
ReactiveValidator
which allows unit tests to compare their results against previously
recorded results in the journal.
Another powerful use case is to enable users to replay production data into test systems. By simply copying over the journal file from a production system and replaying all or part of the file into a test system the exact conditions of the primary system will be reproduced.
ReactiveJournal can be recorded on one JVM and can be replayed (in real-time if required) on one or more JVMs provided they have access to the journal file location.
The remote connection can either read from the beginning of the recording or just start with live updates from the recorder. The remote connection (the 'listener') can optionally write back to the journal effecting a two way conversation or RPC. There can be multiple readers and writers to the journal.
ReactiveJournal uses Chronicle-Queue (a memory mapped file solution) serialisation meaning that the process of moving data from one JVM to another is exceedingly efficient and can be achieved in single digit micro seconds.
If you need to pass data between JVMs on the same machine this is not only the most efficient way to do so but you will also provide you with a full recording of the data that is transferred between the JVMs.
Setting up remote listeners allows a system to have live (hot) backup processes. Another powerful strategy is to set up processes running different code bases so that you can live test new code or play with alternative algorithms and compare results with the exisiting sytem.
If you have a fast producer that you can't slow down but your consumer can't keep up there are a few options available to your system.
Most often you end up implementing strategies that hold buffers of data in memory allowing the consumer to catch up eventually. The problem with those sort of strategies are one, if your process crashes you lose all the data in your buffer. Therefore if you need to consume the fast data in a transactional manner this will not be an option. Two, you may run out of memory if the buffers get really big. At the very least you will probably need to run your JVM with a large memory setting that many be inefficient. For latency sensitive applications it will put pressure on the GC which will not be acceptable.
In such scenarios you can publish your data to ReactiveJournal which shouldn't have any problems keeping up with most feeds (typically writing in the order of 1,000,000/s). You can then read your data from ReactiveJournal knowing that the data is safely stored on disk without having to keep any data in memory buffers.
See more about this topic below in the Examples section
- Recording to the journal is transactional i.e. no data will be lost if the program crashes
- Recording and playback is so fast that it won't slow down the host program.
- Recording and playback can be achieved without any gc overhead
- RxRecorder can be easily added (or even retro-fitted) into any
Reactive
project
An ReactiveJournal
is created as follows:
ReactiveJournal reactiveJournal = new ReactiveJournal(String dir);
The directory is the location where the serialised file will be created
ReactiveRecorder
allows any Reactive Publisher
to be journaled to disk using
the record
function:
ReactiveRecorder reactiveRecorder = reactiveJournal.createReactiveRecorder();
reactiveRexcorder.record(Publisher)
For notes on threading see FAQ below.
ReactivePlayer
is used to playback the journal recording:
ReactivePlayer reactivePlayer = reactiveJournal.createReactivePlayer();
Publisher = reactivePlayer.play(new PlayOptions());
There are a number of options that can be configured using PlayOptions
. These
include filtering the stream by time and stream. Playback speed can also be
controlled using this configuration.
ReactiveJournal
is created and stored to disk using the low latency Chronicle-Queue library.
The data can be examined in plain ASCII using the writeToDisk function:
reactiveJournal.writeToDisk(String fileName, boolean printToSdout)
Full code example code HelloWorldApp.
package org.reactivejournal.examples.helloworld;
import org.reactivejournal.impl.PlayOptions;
import org.reactivejournal.impl.ReactiveJournal;
import org.reactivejournal.impl.ReactiveRecorder;
import org.reactivestreams.Publisher;
import org.reactivestreams.Subscriber;
import org.reactivestreams.Subscription;
import java.io.File;
import java.io.IOException;
/**
* Simple Demo Program
*/
public class HelloWorld {
private static final String tmpDir = System.getProperty("java.io.tmpdir");
public static void main(String[] args) throws IOException {
//1. Create the reactiveRecorder and delete any previous content by clearing the cache
ReactiveJournal reactiveJournal = new ReactiveJournal(tmpDir + File.separator + "HW");
reactiveJournal.clearCache();
//2. Create a HelloWorld Publisher
Publisher<String> helloWorldFlowable = subscriber -> {
subscriber.onNext("Hello World!");
subscriber.onComplete();
};
//3. Pass the Publisher into the reactiveRecorder which will subscribe to it and record all events.
ReactiveRecorder reactiveRecorder = reactiveJournal.createReactiveRecorder();
reactiveRecorder.record(helloWorldFlowable, "");
//4. Subscribe to ReactiveJournal and print out results
Publisher recordedObservable = reactiveJournal.createReactivePlayer().play(new PlayOptions());
recordedObservable.subscribe(new Subscriber() {
@Override
public void onSubscribe(Subscription subscription) {
subscription.request(Long.MAX_VALUE);
}
@Override
public void onNext(Object o) {
System.out.println(o);
}
@Override
public void onError(Throwable throwable) {
System.err.println(throwable);
}
@Override
public void onComplete() {
System.out.println("Hello World Complete");
}
});
//5. Sometimes useful to see the recording written to a file
reactiveJournal.writeToFile(tmpDir + File.separator + "/hw.txt",true);
}
}
The results of running this program can be seen below:
[main] INFO org.reactivejournal.impl.ReactiveJournal - Deleting existing recording [/tmp/HW/.reactiveJournal]
Hello World!
[main] INFO org.reactivejournal.impl.ReactiveJournal - Writing recording to dir [/tmp//hw.txt]
Hello World Complete
[main] INFO org.reactivejournal.impl.ReactiveJournal - VALID 1 1499202194782 Hello World!
[main] INFO org.reactivejournal.impl.ReactiveJournal - COMPLETE 2 1499202194784 EndOfStream{}
[main] INFO org.reactivejournal.impl.ReactiveJournal - Writing to dir complete
When playing back from ReactiveJournal there are a number of options that can be set summarised in the table below:
They are explained more fully in the FAQ and in the javadoc for PlayOptions
Option Name | Values | Default | Description |
---|---|---|---|
filter | any string | n/a | All messages are written with a filter tag |
pauseStrategy | SPIN / YIELD | YIELD | Determines the pause strategy if no events in the journal |
playFromNow | true / false | false | Only play message from now - ignore older messages |
playFromSeqNo | any long | Long.MIN_VALUE | Only play messages from seqNo |
playUntilSeqNo | any long | Long.MAX_VALUE | Only play messages until seqNo and then complete |
playFromTime | any long | Long.MIN_VALUE | Only play messages from time in millis |
playUntilTime | any long | Long.MAX_VALUE | Only play messages until time in millis and then complete |
replayRate | FAST / ACTUAL_TIME | FAST | Play back messages fast or with the intervals they were recorded |
sameThread | true / false | false | Play back messages on the subscriber thread or a new one |
using | any Object | null | Use this Object to populate the message - save on GC |
Items that can be serialised to ReactiveJournal are those that can be serialised to Chronicle-Queue.
These are:
- AutoBoxed primitives, Strings and byte[]
- Classes implementing
Serialisable
- Classes implementing
Externalizable
- Classes implementing
Marshallable
See Chronicle Queue Docs for full documentation
ReactiveJava
has been designed be used with any Reactive implentation
such as RxJava and Reactor.
For example if you had an RxJava Flowable
, because Flowable
implements Publisher
,
you could use the Flowable
as input to
ReactiveRecorder.record()
which takes a Publisher
. ()If you have an RxJava
Observable
you would have to convert it to a Flowable
with Observable.toFlowable()
first.)
ReactivePlayer
returns a Publisher
that can be converted to Flowable
with
Flowable.fromPublisher(reactivePlayer.play(options))
. There is utility class
RxPlayer
that does this for the RxJava user.
There are 2 ways you might want to set up your ReactiveJournal
.
-
Record your
Publisher
input intoReactiveJournal
and then subscribe toReactiveJournal
for its stream of events. This effectively insertsReactiveJournal
into the critical path of your program. This will certainly be the setup if you are using ReactiveJournal to handle back pressure. This is demonstrated in the example program HelloWorldApp_JournalPlayThrough -
Have
ReactiveJournal
as a second subscriber to yourPublisher
. This has the benefit of keeping all functions on the same thread. This might be the setup if you are usingReactiveJournal
to record data for testing purposes. If you are using RxJava you might want to use theConnectableObservable
paradigm as you won't want ReactiveRecorder kicking off the connection until all the other connections have been setup. This is demonstrated in the example program HelloWorldApp_JounalAsObserver
The RxPlayer
can play
in two modes:
ACTUAL_TIME
This plays back the stream preserving the time gaps between the events. This is important for back testing and reproducing exact conditions in unit tests.FAST
This plays the events as soon as they are received. This mode should be set when you are using ReactiveJournal for remote connections or when using RxJounal to deal with back pressure.
The intention is for ReactiveJournal
to support low latency programs. The two main features to allow
for this are:
- Dedicating a CPU core to RxPlayer by using the
PlayOptions.setPauseStrategy(PauseStrategy.SPIN)
setting so that we don't have any context switching on the thread reading from the journal. - Setting the
PlayOptions.using()
so that there is no allocation for new events. This should enable programs to be written that have minimal GC impact, critical for reliable low latency.
There are few core example applications in the code that work through the typical use cases and are worth considering in more detail.
This program demonstrates how to set up a simple 'play through' example.
We have an input Flowable
with a stream of Byte
s. These are recorded in the journal
by ReactiveRecorder
.
We then subscribe to ReactiveJournal
with ReactivePlayer
giving us an Flowable
of Byte
s
which are processed by the BytesToWordsProcessor
. The output of the processor is
also recorded into ReactiveJournal
so we have a full record of all our input and outputs to
the program.
Note that we use recordAsync
rather than record
because otherwise we would
block the main thread until all the event stream had completed recording and only
then would we proceed to process the items. Although in this trivial example
it's hard to see the effect this has I encourage you to play with the INTERVAL_MS
setting to see what happens as you increase the delay to something noticeable.
Then try and change recordAsync
to async
and you will see the effect of
the threading.
We then display the results of the program to stdout as well as writing to a file.
This recording will be valuable when it comes to writing a unit test for
BytesToWordsProcessor
which we'll see in another example.
This is very similar to the last example except that we processes everything
on the same thread. We can do this because rather than the BytesToWordsProcessor
subscribing to ReactiveJournal
it subscribes directly to the Observable<Byte>
input.
This is a less intrusive way to insert RxRecorder into your project but of course will not handle the back pressure problem.
This example demonstrates how to use RxRecorder in a unit test. The journal file
we created in the previous examples is used as input to test the BytesToWordsProcessor
.
The results of BytesToWordsProcessor
are fed into ReactiveValidator
which compares
the output to the output which was recorded in the journal reporting any
differences.
We have effectively black boxed the inputs and outputs to BytesToWordsProcessor
and can
be confident that any changes we make to the processor will not break the existing
behaviour.
This example is designed to show how ReactiveJournal can be used to tranfer data between JVMs.
Start HelloWorldApp_JournalPlayThrough
but increase the INTERVAL_MS
to 1000. Then
run HelloWorldRemote
.
HelloWorldRemote
has been configured with this option:
new PlayOptions().filter(HelloWorldApp_JounalAsObserver.INPUT_FILTER).playFromNow(true);
The playFromNow means that it will only consume current events and depending on how long a gap you have between starting the 2 programs you will see output which looks something like this:
In these example programs we deal with the situation where we find ourselves with a fast producer and slow consumer.
In all these example we setup a scenario in FastProducerSlowConsumer
where the
producer emits Long
values every millisecond. We also create a Consumer
which
processes the Long values with a variable delay which is significantly slower
than the rate that they are being produced.
In other words we have the classic Fast Producer Slow Consumer scenario which needs to be handled by applying back pressure.
The following example programs have all been written to 'solve' the back pressure problem we have created.
Firstly let's consider how RxJava handles back pressure out of the box.
A quick reminder, in RxJava2 the code was split into 2 sections:
Observable
- no back pressure. Use when back pressure is not an issue because the code is more efficient not having to deal with this complication.Flowable
- handles back pressure. Use when you have to address the back pressure issue.
Clearly we will only be looking at the Flowable
part of RxJava2 in this example.
This example program demonstrates how the 5 BackpressureStrategy
modes handle back
pressure.
-
BUFFER
this will, as its name implies, hold the items in an in-memory buffer waiting for availablility on the consumer to process them. This is good choice for handling spikes in event traffic where the consumer will eventually be able to catch up with the producer. The problems using this strategy are:- If the program crashes the events in the buffer will be lost. Even if the program terminates normally careful attention has o be paid to draining the buffer.
- If the queue builds up too much the JVM will run out of memory and crash.
- It forces the program to run with a large memory setting to hold the buffer which can be a problem for programs where latency is an issue especially coupled with the next point.
- The program will not be able to be designed in an allocation-free manner. Every item will have to be created in a 'new' object which will then put pressure on the GC.
-
LATEST
andDROP
deal with back pressure by making the slow consumer keep up with the fast producer. This is done by dropping events from the stream. This is a good choice where events on the stream are replaceable and you don't need to process every item. The problems with this straegy are:- If you want to back test your program against all the values in the stream to see if you might get better results by processing more events.
- As with buffer you can't write GC friendly code.
-
ERROR
andMISSING
deal with back pressure by putting the program into an error state as soon as back pressure is encountered. This is useful when you don't expect any back pressure and you want the program to error on encountering back pressure.
In this program we set up ReactiveJournal
to handle back pressure in the buffer mode
but solving all the problems that we saw with the standard RxJava BUFFER
mode.
The FastProducer can be created with
the BackpressureStrategy.MISSING
because we don't expect that the producer will ever be
slowed down by the consumer, which in this case is RxRecorder
.
The Consumer, rather than subscribing directly to the FastProducer, subscribes to
RxPlayer
. Note that RxPlayer.play
returns an Obserable
as there is no need for it to
handle back pressure because bakc pressure has already been applied using ReactiveJournal
as the
buffer.
Lets look at the problems BackpressureStrategy.MISSING
and see how they are solved.
- Even if the program crashes everything written to ReactiveJournal is safe. The events will be stored
to disk and you can just restart the program and carry on consuming the queue at the point
you crashed. If there is a OS/Machine level issue it is possible that a few messages might
get lost that are waiting to be written to disk.
If that is a problem you should make sure that replication is setup on your system. - There is no in-memory buffer so there is no need to run with extra heap memory and the program
certainly won't run out memory because of
ReactiveJournal
. - When you call
RxPlayer.play
one of the the options isusing
. This allows you to pass in the object that will be used for every event. This means that no new objects will be allocated even if you have millions of items in your stream. (Of course if you want to hole a reference to the event you will need to clone).
In addition to those benefits you will have the ususal benefits of using 'RxJornal' in that you will have a full record of the stream to use in testing and you will be able to use remote JVMs.
As its name implies this demo program shows you how to handle back pressure using ReactiveJournal
but rather than buffer you just want the latest item on the queue.
All you have to do is set up the program exactly as we did in the previous example
ReactiveJournalBackPressureBuffer
but rather than the slow subscriber subscribing to the Observable
that comes from RxPlayer.play
we insert a Flowable
inbetween. The Flowable
is created
with BackpressureStrategy.LATEST
.
See code snippet from the example below:
//1. Get the stream of events from the RxPlayer
ConnectableObservable journalInput = reactiveJournal.createRxPlayer().play(options).publish();
//2. Create a Flowable with LATEST back pressure strategy from the ReactiveJournal stream
Flowable flowable = journalInput.toFlowable(BackpressureStrategy.LATEST);
//3. Record the output of the Flowable into the journal (note the different filter name)
recorder.record(flowable, "consumed");
long startTime = System.currentTimeMillis();
//4. The slow consumer subscribes to the Flowable
flowable.observeOn(Schedulers.io()).subscribe(onNextSlowConsumer::accept,
e -> System.out.println("RxRecorder " + " " + e),
() -> System.out.println("RxRecorder complete [" + (System.currentTimeMillis()-startTime) + "]")
);
You might have noticed that as well as the Slow Consumer subscribing to the Flowable to make sure it uses the LATEST strategy we also record the values we actaully consumer into ReactiveJournal.
As with the plain RxJava implementation of LATEST (without ReactiveJournal) the Slow Consumer only sees the latest updates from the Fast Producer. However if you use RxRecorder (as in this example) you have:
- A full record of all the events that were emitted by the Fast Producer.
- A full record of all the events that were actaully consumed by the Slow Producer.
Both these streams can be played back with RxPlayer
by specifying the appropriate filter
in the PlayOptions
when calling play
.
This leads to being ablse to try the following...
In this example we experiment by replaying the event stream recorded in ReactiveJournal
and
observing the effects of lowering the latency of the SlowConsumer.
We have a recording of the FastProducer created whilst running ReactiveJournalBackPressureBuffer
.
The SlowConsumer subscribes to this using a Flowable
with BackpressureStrategy.LATEST
as in the provious example.
When we run with the SlowConsumer at a latency of 5ms we get this result:
Received [100] items. Published item[100]
Received [200] items. Published item[391]
Received [300] items. Published item[791]
Received [400] items. Published item[1175]
Received [500] items. Published item[1560]
Received [600] items. Published item[1946]
Received [700] items. Published item[2340]
RxRecorder complete [3909ms]
The Slow Consumer has managed to consume about 700 events.
If we reduce the latency of SlowConsumer to 3ms we get this result:
Received [100] items. Published item[100]
Received [200] items. Published item[265]
Received [300] items. Published item[491]
Received [400] items. Published item[719]
Received [500] items. Published item[958]
Received [600] items. Published item[1192]
Received [700] items. Published item[1428]
Received [800] items. Published item[1664]
Received [900] items. Published item[2046]
Received [1000] items. Published item[2288]
RxRecorder complete [3666ms]
The Slow Consumer has now managed to consume about 1000 events.
Whilst this is a trivial example I'll let your imagination extend the scenarios to real world situations where this sort of ability to replay data against real load will be invaluable.
Special thanks to my friend and ex-collegue Peter Lawrey for inspiring me with his Chronicle libraries which underpin ReactiveJournal.
To those behind RxJava in particular to Tomasz Nurkiewicz for his talks and book which opened my eyes to RxJava.