Being prepared for traffic peaks with Gatling

I work in a ticketing company, where scaling is often difficult because of peaks in sales. For example, the traffic on ticket shops dramatically increases when tickets for famous singers go on sale or when large theaters sell their annual subscriptions.

Yesterday, tickets for a famous french band went on sale. Let’s have a look at the graph of the cumulative number of sales.

This is quite a typical distribution when the opening time for sales has been advertized beforehand, most of the tickets are sold quite quickly.
If we look at it more carefully, we can clearly see that there are two different parts in the graph:

in the first part, the number of sales increases very quickly and it seems that it follows a Heaviside function.
in the second part, the number of sales is linear.

Gatling 2-M1 was just released, with a completely new injection API. Now one can easily define some custom injection functions. Gatling also comes with some nice predefined functions which we are going to use to try to reproduce that distribution.

Let’s consider this code :

setUp(scn.inject(heaviside(2400 users).over(30 minutes),
                 ramp(340 users) over (30 minutes))
      .protocolConfig(httpConf))

It’s now time to compare the simulated load with what we observed earlier :

By trying to reproduce accuratly some patterns observed in production you definitely have the keys for making your website ready for high traffic.

24 January 2013

Photos from our round the world trip

It’s now almost two years since we came back from our trip around the world. Given the duration of the trip, I didn’t took so many photographs but still, there was about 60 rolls to develop, to scan and to go thru. So yes, it took me two years to process that but I’m quite thrilled to be able to share some exciting memory of that trip with you now.

See the photos here : http://nire.free.fr/rtw

27 November 2012

Gatling at the GenevaJUG

Last night was the second occasion for me to give a public presentation of Gatling ; an open-source load testing tool. I have been a contributer of the project for the last months after using it at work and have been very excited about it. I hope this presentation gave you a good idea of what the tool is capable of.

08 June 2012

How to install a driver for a Canon network printer on Ubuntu 12.04 64bits

Being able to print something on a Canon network printer (C5030i) from an Ubuntu 64 bits machine could be quite a pain. I spend a bit of time to have it working and I’m keeping here my notes.
The community has been asking for a packaging for the Canon UFR II driver and some of the tricks listed here have been found in the comments of the Launchpad issue¹.

Start by downloading the official canon drivers, a zip file called o1113enx_l_ufr220.zip at this URL, http://software.canon-europe.com/software/0040355.asp

As there is no 64bits deb package, we gonna have to use the rpm files. Proceed as follow:

# convert to deb files
sudo alien -d --scripts cndrvcups-ufr2-uk-2.2.20-1.x86_64.rpm
sudo alien -d --scripts cndrvcups-common-2.20-1.x86_64.rpm
# install the packages
sudo dpkg -i cndrvcups-common_2.10-2_amd64.deb
sudo dpkg -i cndrvcups-ufr2-uk_2.10-2_amd64.deb

Now, you can install the printer, go to SystemSettings > Printers > Add > NetworkPrinter and then, when you have to choose the driver, select Provide PPD file. Select a ppd file from the /usr/share/cups/model folder.
The issue is now that AppArmor is blocking the /usr/lib64 which is not standard on a Debian machine. So, edit /etc/apparmor.d/local/usr.sbin.cupsd and add these lines
```
/usr/lib64/cups/backend/cnusb Uxr,
/usr/lib64/cups/filter/pstoufr2cpca Uxr,
```
You should also do this so that the libraries can be found :
```
ln -s /usr/lib64/lib* /usr/lib/x86_64-linux-gnu/
```
The last problem will be fixed by installing some 32bits libraries as the driver is not entirely compiled in 64bits.
```
sudo apt-get install ia32-libs libjpeg62:i386
```

Try to print a test page and enjoy.

¹ https://bugs.launchpad.net/ubuntu/+bug/502920

14 November 2011

Instagram

The engineers at Instagram put forward a challenge for coders to unshred this chopped up image. See the original post on their engineering blog : Instagram Engineering Challenge: The Unshredder.

The problem

Here are the input and target output images:

The shredded version

The unshredded version

I wrote a solution in Python that resolves the problem; here is the basic idea of the algorithm :

define a correlation function for two pixel columns
start with a random shred
find the shred with the maximum correlation value that would fit on the left hand side or on the right hand side of the reconstructed picture
iterate until all the shreds have been used

The bonus question was to guess how wide the uniform shreds are. I proceeded this way :

Consider all shreds width that can divide the image correctly
Compute the difference score between all shreds – it’s the opposite of the correlation score used before and I also include a ratio of the shred_width
When we are on a edge between two sherds, the difference is at its maximum. So, take the maximum of all computed difference scores.

For more details, see the code in my repository on github