How we built the CyberSift Attack Map

Recently we launched a small site called the “CyberSift Attack Map” hosted at http://attack-map.cybersift.ioAny one involved in the InfoSec industry will be instantly familiar with the site:


Screenshot from 2017-10-12 12-36-06

It’s basically a map of attacks which either trip some rule in a signature based IPS such as SNORT, or land in a honeypot. In this article we’ll list some of the libraries and techniques we used to build the site for any devs out there who are interested.


We used the python Flask microframework, work in unison with Gevents. This was a relatively straightforward operation, with the major point of note being our inclusion of support for HTML5 server side events [SSE]. SSE are WebSocket’s smaller brother. Whereas websockets allow for bi-directional communication between browser and server, SSE only allow for unidirectional communication from server to browser. This suited our needs perfectly since we don’t really need the browser to get back to our server, but we do need real-time updates to flow from our server back down to the browser to populate the attack details, gauges and so on. We based our SSE Flask implementation off the extremely useful snippet from Oskar Blom you can find here:

If you are going to go down this road, make sure you add some sort of “heartbeat” which instructs Flask to remove inactive clients from the subscriptions. As soon as our honeypot or twitter microservice got wind of a new attack, it’s pushed  as a message via SSE and updates the client browser in realtime


It’s amazing to see just how quickly you can build a useful site because of the incredible hard work of other talented devs… In this case we used the following libraries to build the site:

If the site really catches on, we would need to consider re-writing the frontend with something like ReactJS to improve rendering performance


Elasticsearch & Python: Tips for faster re-indexing

Some valuable lessons learned while going through an elasticsearch re-indexing exercise. (For the uninitiated, re-indexing data basically means getting large volumes of documents from elasticsearch, enriching or changing the data within each document, and then sending these back). The below assume python and the use of the python elasticsearch client (

1. When retrieving data, use the “scan” helper function

Most examples I encountered on the Internet used the “scroll” API. Nothing wrong with this, in fact the scan helper function is itself a higher level function based on the scroll API. However, the scan function abstracts everything away and frees the developer from having to worry about scroll IDs and so on. It’s much simpler to use, let’s see a quick example:

In line 4 above, we use the usual search API with a size of 0 to count the number of matching documents (this replaces the currently deprecated “count” API), which we print out. In line 10, we use the scan API to return a very large number of documents. The scan function returns an iterable which allows us to process the response (the res variable above) using loops and so on, which leads us to our next point

2. Use python’s multiprocessing capabilities to speed up the document processing

Since the object returned from the scan helper function is an iterable, this makes it ideal to split this over multiple processes that obviously use up processing power much more efficiently. The easiest way to do this that I found was:

  • Create a multiprocessor pool with the appropriate number of processes (usually one per CPU), and an array which will aggregate the results from the different processes
    import multiprocessing as mp
    pool = mp.Pool(processes=4) # start 4 worker processes
    result = []
  • Use python’s multiprocessing pool “map” functionality to split the iterable returned from the scan helper function over the processors
    N = 11
    while True:
        g2 =, itertools.islice(res, N))
        if g2:

    In the above code snippet, “processor” is a function which processes the documents. Each processor is fed a slice of the res iterator. Note how the “itertools.islice” function splits the res iterator returned from elasticsearch scan API into 11 slices. The results from each process is placed in the “results” array we previously initialized.

    At this stage, we have four processes in parallel munging through the documents, and emitting their results to a common array ready to be passed to the next step

3. Use the Python Elasticsearch “Parallel Bulk” helper function

The parallel bulk helper function again abstract a lot of work away from the developer. It allows you to very simply define the number of threads used to update elasticsearch and so on. As input, the parallel bulk API takes an array of “action items”, with each action being a python dictionary, an example of which can be seen below:

{'_op_type': "update",
        '_index': "filebeat-123",
        '_type': "syslog",
        '_id': "ABSAJSGA19872972_",
        'doc': { 'foo' : 'bar'}

In our case, this dictionary is what is returned by each of our multiprocessing processors, and appended to the results array. We now simply feed this array into the parallel bulk helper function as follows:

print 'updating results'
parallel_bulk(es, result)
print 'done updating results'

Where “es” is the elasticsearch client, and result is the array which the processors build, with each element being a dictionary as described above

Note: In my case, the above methods increased the throughput from approximately 2 documents/sec to about 10 documents/sec, but these figures came nowhere near as being as fast as a java implementation which gave me about 80-100 documents/sec. 

PS: the throughput numbers I quoted above are not meant as absolute benchmarks, I do some pretty heavy processing including HTTP requests, so you are likely to get faster results depending on your processing. The numbers are meant as relative benchmarks to show that multiprocessing can increase your throughput by 4x-5x as much.

P.S.S: I will post an article with the Java re-indexing lessons learned (hopefully soon)