A basic CEP package written in erlang
Erlang Shell JavaScript
Latest commit 6f7bd54 May 24, 2013 @danmacklin Found a bug caused by the recent refactor.
Must improve the time based window unit tests, as I only found the bug
by running the stress test!!!!!!


erlang_cep v0.2 Beta

A simple CEP (Complex Event Processing) engine written in erlang OTP inspired by esper and CouchDB.


  • Simple API.
  • CEP rules programmed in javascript or erlang.
  • Supports many data feeds. Each feed having none or many time or size based windows.
  • Three step CEP rules. Configuration sets-up the window. Row Functions filter on entry to a window. Reduce Functions aggregate data when a window breaches it's configuration thresholds.
  • Standard, MatchRecognise and Every CEP rules.
  • Configurable pattern recognition.
  • Search API.
  • Joined Windows.
  • Publish and subscribe mechanism notifies interested processes when a CEP rule matches.

How Does it Work?

The feed_api is used to create Feeds and Windows.

Each feed can have none or many windows.

A Feed is a stream of Json encoded data (other formats are possible, but might need some more testing in this version).

A Window is an in-memory data storage abstraction. Each window has a Row Function, Reduce Function and associated configuration. Functions can be implemented in javascript or erlang.

A Row Function is a filter. Data is added to the window if it passes the filter.

A Reduce Function is a data aggregator. It aggregates data when a CEP rule fires.

Interested processes subscribe to a window, and are notified when a CEP rule fires.

A good example is a Stock Market Feed with a time based window. The window is programmed to fire when three trades with the symbol GOOG and a volume of 100 shares or more make it into the window. When the window fires the Reduce Function runs and averages the sale prices. The average sale price is distributed to interested processes via publish and subscribe.

Query Types

erlang_cep supports three types of query.

  • "Standard" - Performs simple matches. I.E. match every Google trade with a sale price > £10.

  • "matchRecognise" - Perform matches that make a progression. I.E. match every Google trade with a sale price greater than £10, where the next sale price is greater than the previous sell price. Fires when this progression happens 3 times in a row.

  • "every" - Performs standard matches where the results are output at a regular interval. I.E. in a 60 second window give me the average sale price for all Google trades.

Window Types

erlang_cep supports two types of window.

  • size - Sized windows have a fixed size. Each time data is added the position within the window is incremented. Once the position equals the window size the position resets to 0 and the data within the window is overwritten and expired. If you think of it as a ring buffer you won't be too far away.

  • time - Time based windows expire data after a set period of time. A 60 second time based window stores any number of data points over a 60 second period.

Window Configuration

Windows are configured through a list of tuples. Version 0.2 supports the following configuration parameters.

  • {numberOfMatches, Numeric} - The number of matches required to fire the Reduce Function.
  • {windowSize, Numeric} - The size of the window (number of elements for size based windows, time in seconds for time based windows),
  • {windowType , size / time} - The type of the window size or time.
  • {consecutive, consecutive / nonConsecutive} - Specifies whether matches need to be made up from consecutive or nonConsecutive data elements. For example if this configuration is set to consecutive, and numberOfMatches is set to 3, then the Reduce Function will fire the Row Function when three consecutive elements match.
  • {matchType, standard / matchRecognise / every} - The type of CEP rule.
  • {resetStrategy, restart / noRestart} - After the Reduce Function runs should the window be reset or carry on.

The default windowType is size. The default consecutive parameter is consecutive. The default matchType is standard. The default resetStrategy is restart.

If a query fails to parse an error atom is returned.

Some examples :-

Set up a standard four element size window, which will fire it's Reduce Function after three consecutive matches and then restart.

QueryParameterList = [{numberOfMatches, 3}, {windowSize, 4}]

Set up a matchRecognise four element window, which will fire it's Reduce Function after three consecutive matches and then restart.

QueryParameterList = [{numberOfMatches, 3}, {windowSize, 4}, {matchType, matchRecognise}],

Set up a matchRecognised four second time based window, which will fire it's Reduce Function after three consecutive matches and then restart.

QueryParameterList = [{numberOfMatches, 3}, {windowSize, 4}, {matchType, matchRecognise}, {windowType, time}],

Row Functions

Row Functions are defined in javascript as binaries or erlang functions. There are three types of Row Function "standard" / "every" and "matchRecognise".

Standard Row Functions

"Standard" Row Functions are used to filter data on entry to a window. They can be implemented in javascript or erlang functions.

Javascript Row Functions

Each javascript Row Function must accept the following parameters :-

  1. parameters - An array of values initialised when the window is first started. Use these parameters to build functions without hard coded values. i.e. if (price > parameters[0]) rather than if (price > 1.00)
  2. joins - An array or arrays of joined data. [ [FeedName, WindowName, [ Joined Rows ]] ]. Each joined row is an array of one or more elements.
  3. row - A string containing the new data.
  4. otherRow - A string holding data from the previous match. Only used in "matchRecognise" windows, where the new data is compared against the last successful match, rather than a hard coded rule.
  5. sequence - The current number of matches within this window. Use this number and an array of parameters to perform complex matches.
  6. matchRecogniseFirst - A boolean set to true if this is the initial check within a matchRecognise window, false if checking data against previous values.

As an example this javascript function will match if the price is greater than 1.00.

<<"var rowFunction = function(parameters, joins, row, otherRow, sequence, matchRecogniseFirst){

                            var myObject = JSON.parse(row);
                            symbol = myObject.symbol;
                            price = myObject.price;
                            volume = myObject.volume;

                            if (price > 1.00){
                                return [symbol, price, volume];

                            return []}">>.

To debug a javascript row function use

ejsLog("/tmp/foo.txt", "Hello");

If a Row Function fails to match it must return an empty list.

The results of a Row Function must be a list. The order of the list is important as it dictates how the data is stored within the window results. Row Functions, searches and joins must respect this order.

When using joins Row Function code can be simplified by using a Join object.

Create a join object passing the joins parameter and optionally the name of the join feed / window if you want to restrict the join.

var j = new Join(joins);

Use the exists(value, position) function to Search the joins array for a value at position pos.

The source code for the javascript API can be found in /src/js. If you add any other javascript functions to this file / directory they will be loaded into the javascript vm each time a window starts.

Here's an example.

var rowFunction = function(parameters, joins, row, otherRow, sequence, matchRecogniseFirst){
    var j = new Join(joins);

    if (j.exists(parameters[0], 0)){
        return [\"Yes\"];

return [\"Nope\"]}

This example returns Yes if the value of parameters[0] is found at position 0.

Here's an example of a javascript Row Function with a join looking for data from a feed called feed1 and window called window1.

var rowFunction = function(parameters, joins, row, otherRow, sequence, matchRecogniseFirst){
    var j = new Join(joins, 'feed1', 'window1');

    if (j.exists(parameters[0], 0)){
        return [\"Yes\"];

return [\"Nope\"]}

Javascript "MatchRecognise" Row Functions

"MatchRecognise" Row Functions are used to find complex progressive patterns in data. Typically a MatchRecognise Row Function is called from two perspectives. Initially the "matchRecogniseFirst" parameter is set to true and the function acts as a threshold check. In this mode the Row Function is a filter restricting the data that can be added to the Window. On subsequent calls the matchRecogniseFirst parameter is set to false, and data stored from the previous match is passed into the "otherRow" variable so these values can be compared as part of a progressive patten match.

The matchRecogniseFirst parameter is set to false when data expires from the window, or a query fires (if resetStrategy is set to restart).

When the matchRecogniseFirst parameter is set to true the function must return an array of data.

When the matchRecogniseFirst parameter is set to false, the function is being called from the perspective of "matchRecognise" so only needs to return true or false.

An example javascript MatchRecognise Row Function

<<"var rowFunction = function(parameters, joins, row, otherRow, sequence, matchRecogniseFirst){

                            var myObject = JSON.parse(row);

                            symbol = myObject.symbol;
                            price = myObject.price;
                            volume = myObject.volume;

                            if (matchRecogniseFirst == true){

                            // This path is taken the first time that the row function runs

                                if (price > 0.50){
                                    return [symbol, price, volume];

                                    return [];


                                var prevObject = JSON.parse(otherRow);                              

                                if (volume > prevObject.volume){
                                    return true;

                                    return false;


Reduce Functions

Reduce Functions are called to aggregate results when a CEP Window Fires. They can be written in javascript or as erlang functions. The amount of data sent to the Reduce Function is equal to the numberOfMatches parameter set in the window configuration.

Javascript Reduce Functions

A Reduce Function takes one parameter which is a multidimensional array containing the results from all the Row Functions that have successfully matched (i.e the arrays returned from all successfully matched Row Functions).

In this javascript example the matches parameter holds the following array :-

[["Goog", 2.00, 20], ["Goog", 2.01, 21]]

<<"var reduceFunction = function(matches){
                        var sum = 0;

                        for(var i=0; i<matches.length; i++) {
                            var match = matches[i];
                            sum += match[2];

                    if (sum > 0){
                        return sum / matches.length;

                        return 0}">>.

The feed_api

The feed_api is used to :-

  1. Create new feeds and windows.
  2. Subscribe your application to be notified once Reduce Functions run.
  3. Set-up joins.

The important functions are :-

  1. feed_api:start_feed(). - Start a feed gen_server named after the atom.
  2. feed_api:start_window(, , RowFunction, ReduceFunction, QueryParameters, YourParameters). - Add a window to the feed.
  3. feed_api:subscribe_feed_window(, , Pid). - Subscribe a Pid to the feed / window.

The parameters of start_window are :-

  1. The name of the window as an atom.
  2. The name of the feed as an atom.
  3. A binary containing a javascript Row Function or a tuple containing an erlang Module and Function Name {RowFunctionModule, RowFunctionName}.
  4. A binary containing a javascript Reduce Function or a tuple containing an erlang Module and Function Name {ReduceFunctionModule, ReduceFunctionName}.
  5. A list of configuration parameters. View the Window Configuration section for more details.
  6. A list of parameters passed into the match function. This is often used to look for complex patterns.

The following code shows how to set-up a new feed, add a new window and subscribe a process to receive updates using javascript.

%% Create a list containing some Binary Json.  [<<" valid json ">>, <<" Some more ">>]  
Data = window_api:create_json([{"1.01", "14"}, {"1.02", "15"}, {"2.00", "16"}]),

%% Generate a test Row Function - In practice this could come from a database it just has to be valid javascript within an Erlang binary.
RowFunction = window_api:create_match_recognise_row_function(),

%% Generate a test Reduce Function - Again this could come from a database.
ReduceFunction = window_api:create_reduce_function(),

%% Set-up a four second time based window looking for three consecutive matchRecognise rows.
QueryParameterList = [{numberOfMatches, 3}, {windowSize, 4}, {matchType, matchRecognise}, {windowType, time}],

%% Register a new feed called stress

%% Create a new window called stressWin and apply it with out functions and parameters to the stress feed.
feed_api:start_window(stressWin, stress, RowFunction, ReduceFunction, QueryParameterList, []),

%% Spawn a callback function that will be messaged when the reduce function runs.
Pid = spawn(?MODULE, get_response, []),

%% Subscribe the callback function to the stressWin window.
feed_api:subscribe_feed_window(stress, stressWin, Pid),

lists:foreach(fun (Data) ->
                %% Add data to the stress feed
                feed_api:add_data(stress, DataElement)
              end, Data).

This code shows an example subscriber call back function.

get_response() ->
        Results ->
            io:format("Stress test fired ~p ~n", [Results]),

Row and Reduce Functions in erlang

Version 0.2 includes support for writing Row and Reduce functions in pure erlang.

An example MatchRecognise query written in erlang can be found within example_erlang_match_recognise.erl.

When using erlang make sure that the Row and Reduce Function parameters within the feed_api:start_window function are tuples containing the module and function where the Row and Reduce Queries are implemented.

You can find a full example of an erlang matchRecognise implementation in feed_api:erlang_match_recognise_test().


You can search within a window by calling

search_api:search_window(FeedName, WindowName, SearchParameter, Row, rowType).

Where :-

  1. FeedName is the name of the feed;
  2. WindowName is the name of the window;
  3. SearchName is the search parameters in the format of a list. Where the length of the list has to be the same size as the list returned from your row function.
  4. Row is the latest input data.
  5. rowType is an atom. Set this to hardCoded if you don't want to perform any substitution Set to json if you want to perform json variable substitiution.

For example :-

['', '', '_'] - List all of the results in the window.

["GOOG", '', ''] - List all of the results where the first value returned from the row function is "GOOG".

["GOOG", 1.01, 12] - Lists all of the results where the stock is GOOG, the price is 1.01 and the volume is 12.

["Obj.symbol", '', ''] - Extracts Obj.symbol from the latest row if the rowType atom is set to json

Some more examples

Please look in stress_test.erl (run_match_test() and run_every_test()) and examples.erl for more examples.

There are many good examples in the eunit tests within feed_api.erl.

Joining Windows

You can join windows by making use of the search api.

Joins need to be formatted as a list and added to a window using the feed_api:add_searches() function.

For example :-

    Search = [[joinFeedOne, joinWinOne, ['_', <<"Dan">>], hardCoded]],
    feed_api:add_searches(joinFeedOne, joinWinTwo, Search). 

This code sets up a hard coded search (i.e. one where search parameters do not need to be substituted) such that every time new data is added to the joinWinTwo window in the joinFeedOne feed, the search query is run on the joinWinOne window looking for the string Dan in the second results field.

The results of this search are applied to the join parameter within the Row Function in the format [ [FeedName, WindowName, [ Joined Rows ] ] ].

More complex searches can run over multiple windows. The results for a more complex multi-window search might look like this.

[ [FeedName1, WindowName1, [ FeedName1 WindowName1 Joined Rows ] ], [FeedName2, WindowName2, [ FeedName2 WindowName2 Joined Rows ] ] ].


See examples.erl.


  1. Spider Monkey
  2. erlang_js
  3. lager - An erlang logging framework

Compilation / run Instructions

chmod +x rebar

Get the dependencies

make deps

Clean the build environment

make clean

Build and deploy release

make deploy


make test

Start the application

make run

Make sure that you run make deploy before running make test!

You can speed things up a bit by disabling lager logging within the sys.config file within /erlang_cep/rel/files.

Logging / working out what has gone wrong.

I have included the excellent lager logging framework from Basho. You can change the logging levels by modifying the sys.config settings within the rel/files directory. By default the logs are outputted to the /erlang_cep/rel/erlang_cep/log directory.


Please note that this is a very early beta release, and I'm quite new to erlang..........

  1. Implement "Every" queries for standard windows, currently this functionality is only available for timed windows.
  2. Look at failure scenarios. Subscriptions and some config will not survive window process failures.
  3. Implement more advanced CEP algorithms such as Rete.
  4. Use Mnesia to try to make windows HA.
  5. Write a web based UI to allow users to create feeds and view results.
  6. Test, test and test again
  7. Some more stress tests.
  8. Support more window types.
  9. Each window currently has it's own Javascript VM. As the number of VM's increases this could cause issues.
  10. Enhance the API so that users can write Row and Reduce functions in erlang. Make sure that these functions are dynamically loaded.
  11. Create a REST API through which users can add / mutate windows, and row / reduce functions.

Please let me know if you find any bugs!


This software is open source. It is licensed under the Apache 2 license. Please feel free to email me at danmacklin10 at gmail.com if you have any issues or suggestions.


Change Log

19/05/2013 - V0.2 Allow Row and Reduce Functions to be written in pure erlang as well as Javascript