Distributed web crawler using Mesos and Go

OwlCrawler is a simple yet distributed web crawler I'm working onto learn Mesos. I'm using the mesos go binding to write the logic and use Iron MQ to trigger tasks that fetch pages. At this point I'm storing the url and html in etcd but I have plans to move that to CouchDB in the near future.

A bit of history.

Writing a web crawler is something I have been thinking about for several years now, when I first started using Scala many years ago, I thought actors would be a great tool to have workers go out and do the crawling. At that time Akka didn't have remote actors in place and even after they added them, it felt to me that setting up the different members of the cluster wasn't as simple as I wanted it to be, note that in my case, this is just a personal project, so I was looking for something simple and fun to use.

About a year and a half ago I started using Go and there goroutines looked like the way to go, but I still had to setup the different members of a cluster, and coordinate how they should all work together.

Finally, last week I was looking at mesos once again (I had been meaning to try it out for a while) and that's when I decided to take their sample framework and executor and see what it would look like to write a web crawler on top of it.

Mesos' rainbow

Getting mesos installed was pretty simple following their getting started page. After that, it was time to write the crawler's logic using mesos-go. For this I started with the framework part. I only had to modify the ResourceOffers method which is called by mesos when there are resources available on the cluster.

Framework's logic

The logic I have so far goes like this: I check if the cpu and memory resources mesos is giving me are enough for the task, if they are, I check if ironmq has any messages for this app, if there is a message, it would be the next url I'm supposed to fetch.

Once I read the url, I check if that url was already fetched, if etcd doesn't have it, I encode some data using go's encoding.gob and send the task to the mesos cluster.

You can see the complete framework code here. Note that while mesos calls it framework, it really doesn't have to be a lot of code.

I love how I don't need to tell the crawler which worker is going to get the next task, I just send it out there and mesos deals with it.

Executor's logic

On the executor's side, the main method I modified was LaunchTask. Here I decode the data I got from the framework, so far this is the data I pass from the framework to the executor:

type OwlCrawlMsg struct {
    URL       string //URL to fetch
    ID        string //IronMQ message id, so I can delete it once I'm done
    QueueName string //IronMQ queue name
    EtcdHost  string //Etcd host to look things up from

This is another place where the logic is simple, I use Go's http client to fetch data for the url, then I store the html in etcd and also parse the page looking for links. If the url looks like something I want to fetch, I then send it as a message to IronMQ and finally send an status update to the mesos master telling it we are done.

And the loop starts again, the framework gets a resource offer, and then it queries the queue, it will find this new url and it just keeps going until it got all the links from the site.


You can find this project's source code on github I'll be updating it to clean it up and fix any issues I find, it's in early stages, but I didn't want to wait any longer before putting it out there.


Mesos does exactly what I wanted, I setup a cluster of servers and then write my app without worrying about which server is going to run which task, or if I have enough resources for all of them. And using a message queue feels like the right choice here to communicate the source of the tasks.

And let's not forget that using Go here made a huge difference, as I was learning mesos, I had to do a lot of trial an error to get things working, and the fast compiler made this project a lot of fun to work on.

Random notes:


Using encoding/gob to pass binary data from one process to another is just dead simple:

var msgAndID bytes.Buffer
enc := gob.NewEncoder(&msgAndID)
err = enc.Encode(OwlCrawlMsg{
	URL:       msg.Body,
	ID:        msg.Id,
	QueueName: queueName,
	EtcdHost:  *etcHostAndport,
if err != nil {
	log.Fatal("encode error:", err)

And decoding it on another process:

var queueMessage OwlCrawlMsg
dec := gob.NewDecoder(payload)
err = dec.Decode(&queueMessage)
if err != nil {
	fmt.Println("decode error:", err)

In this case I'm using the same name for the struct OwlCrawlMsg but they are actually two different structs defined on two different files, as long as they have the same fields inside, they work.


Their api is almost too easy to use, from managing credentials by adding a .iron.json file in your $HOME, to the few method calls you have to do.

Thank you for reading and don't hesitate to leave a comment/question.