sort -k

TIL, when you call sort -k3, you’re not just sorting by the third field, but by whatever the value between the third field up to the end of the line is.
Not only that, in the case of ties, by default it will use also the first field.

Consider this example.

$ cat data
theta AAA 2
gamma AAA 2
alpha BBB 2
alpha AAA 3

Sorting with -k2 gives:

$ sort data -k2 --debug
sort: using simple byte comparison
gamma AAA 2
theta AAA 2
alpha AAA 3
alpha BBB 2 

Notice I’ve also added --debug, to show which parts are used in the comparisons.
So, first comes “AAA 2”, then “AAA 3”.
Also, for the two lines that have “AAA 2”, the first field is used, so “gamma” comes before “theta”.

Forget about the ties for now.
To consider field 2 only, rather than field 2 and all following fields, you need to specify a stop. This is done by adding “,2” to the -k switch. More in general, -km,n means “sort by field m up to n, boundaries included”.

$ sort data -k2,2 --debug
sort: using simple byte comparison
alpha AAA 3
gamma AAA 2
theta AAA 2
alpha BBB 2 

As you can see, field 2 only is taken into account at first.
“AAA 3” comes before “AAA 2” because, being a tie, the first field is used as a second comparison.

Taking this a step further, to actually only consider field 2 and resort to the original order in case of ties, that is, to have a stable sort, you need to pass the -s switch.

$ sort data -k2,2 -s --debug
sort: using simple byte comparison
theta AAA 2
gamma AAA 2
alpha AAA 3
alpha BBB 2 

This look similar to the first snippet, but actually the first two lines in the output are swapped. Here they appear in the original order.

Timezones and DST in Python

It’s incredible how fiddly it is to work with timezones.

Today, 14th of June—and this is important—I was trying to convert a made-up datetime from “Europe/London” to UTC.

I instinctively tried out this:
>>> almostMidnight =, minute=59, second=59, microsecond=999999, tzinfo=pytz.timezone('Europe/London'))
>>> almostMidnight
datetime.datetime(2017, 6, 14, 23, 59, 59, 999999, tzinfo=<DstTzInfo 'Europe/London' GMT0:00:00 STD>)

At this point you will notice it didn’t take into account the DST offset (it should read BST).

As a further confirmation, converting to UTC keeps the same time:
>>> pytz.UTC.normalize(almostMidnight)
datetime.datetime(2017, 6, 14, 23, 59, 59, 999999, tzinfo=<UTC>)

Notice this result would be fine during the winter, so depending how much attention you devote and when you write the code you might miss out on this bug – which is why I love having the same suite of tests always running on a system that lives right after the upcoming DST change.

Even more subtler, if you were to try and convert to a different timezone, a geographical timezone that observes DST, you would see this:
>>> almostMidnight.astimezone(pytz.timezone('Europe/Rome'))
datetime.datetime(2017, 6, 15, 1, 59, 59, 999999, tzinfo=<DstTzInfo 'Europe/Rome' CEST+2:00:00 DST>)

Interesting. Now DST is accounted for. So converting to geographical timezones might also mask the problem.

Long story short, the correct way *I believe* to convert the timezone of a datetime object to UTC is to create a naive datetime object (no timezone info attached) representing localtime, and then call the “localize” of the timezone of interest. In code:
>>> almostMidnight =, minute=59, second=59, microsecond=999999)
>>> almostMidnight
datetime.datetime(2017, 6, 14, 23, 59, 59, 999999)
>>> pytz.timezone('Europe/London').localize(almostMidnight).astimezone(pytz.UTC)
datetime.datetime(2017, 6, 14, 22, 59, 59, 999999, tzinfo=<UTC>)

There’s a very nice read on timezones by Armin Ronacher, which I recommend.

Remote tail

You might have wondered why I felt the urge to specify “local” in the title of last post. Well, fast forward a few days since then, for a similar set of tests I also needed to check a log file on a remote Linux machine – that is, I needed some kind of remote tail.

ssh tail -f

We already know select will be part of our tool set. On top of that, we’ll need to forward the command across the network – and a nice way of doing that is over SSH. In Python, this task is relatively simple if you choose to use paramiko, a third party library implementing the SSHv2 protocol.

A few caveats here, as well. The following snippet is a raw prototype to demonstrate the functionality. It fit my bills, but YMMV. Of course many aspects can be improved, starting from instance with isBeginningOfMessage, which is much better placed in a derived class, so that different BOM patterns can be handled. Closing the SSH channel cleanly is also something you might want to polish before using this class.

import paramiko
import re
import select
import Queue

class SSHTail(object):
    """Tail a remote file and store new messages in a queue
    READ_ONLY = select.POLLIN | select.POLLPRI | select.POLLHUP | select.POLLERR
    TIMEOUT = 1000  # milliseconds
    BUF_SIZE = 1024
    NEWLINE_CHARS = {'\n', '\r'}

    def __init__(self, host, path): = host
        self.path = path
        self.poller = select.poll()
        self.messageQueue = Queue.deque()

    def start(self):
        """Start the tail command and return the queue used to
        store read messages
        client = paramiko.SSHClient()
        self.client = client

        transport = self.client.get_transport()
        self.transport = transport

        channel = self.transport.open_session()
        channel.exec_command("tail -F %s" % self.path) = channel

        self.poller.register(, self.READ_ONLY)

        return self.messageQueue

    def isBeginningOfMessage(line):
        """Return True if the line starts with the hardcoded Beginning of
        Message pattern
        BOMPattern = ''
        return re.match(BOMPattern, line)

    def loop(self):
        """Whilst the SSH tunnel is active, keep polling for new
        content and call parseBuffer() to parse it into messages
        while self.transport.is_active():
            events = self.poller.poll(self.TIMEOUT)
            for fd, flag in events:
                if flag & (select.POLLIN | select.POLLPRI):
                    buf =

    def parseBuffer(self, buf):
        """Given a buffer buf, split it into messages and glue it together to
        previous messages, if buf is not the beginning of a message.
        Note: assumes each message is on its on line.
        if buf:
            messages = buf.splitlines()

            oldest = messages[0]
            if not self.isBeginningOfMessage(oldest):
                    messages[0] = self.messageQueue.popleft() + oldest
                except IndexError:

            for message in messages:

Local tail

tail -f

A while ago I needed, for one of my tests, to monitor a log file on a Linux system and store any new lines, so that I could access the added content at the end of the test. In a sense, I needed a kind of buffered tail -f on a local file.

A quick search led me to the select module.
Without further ado, here’s the code to watch one or more files, and to store anything added to those files in a message queue.

It’s a quick and dirty version which can be improved in many ways. For starters, the keys to access the message queues are the sockets themselves, pretty useless in general, but good enough in my case. Second, notice the file is never closed explicitly: definitely not ideal.

import Queue
import select

class Watcher(object):
    TIMEOUT = 1000
    READ_ONLY = select.POLLIN | select.POLLPRI | select.POLLHUP | select.POLLERR

    def __init__(self):
        """Initialize the Watcher"""
        self.poller = select.poll()

        self.fd_to_socket = {}
        self.message_queues = {}

    def addFile(self, path):
        """Add a file to monitor.
        :path: absolute path of the file, including the filename
        f = open(path)
        self.poller.register(f, self.READ_ONLY)
        self.fd_to_socket[f.fileno()] = f 
        self.message_queues[f] = Queue.deque()

    def start(self):
        """Start polling files"""
        while True:
            events = self.poller.poll(self.TIMEOUT)
            for fd, flag in events:
                s = self.fd_to_socket[fd]
                if flag & (select.POLLIN | select.POLLPRI):
                    lines = s.readlines()
                    if lines:

Filtered tail in less

Not too cryptic title for a simple post on how to tail a file in less, while also filtering only the lines you are interested in.

First, less +F the file. Next, CTRL+C to stop tailing and press & to call the filtering command. Type in your pattern, hit return and resume tailing by pressing Shift+F.

Append an item to an OrderedDict

I needed a way to append an item to an OrderedDict, without creating a new object (too consuming) and I stumbled upon this answer on StackOverflow.

The answer gives a solution to the inverse problem (that is, prepending an item), but was good enough to be modified for my situation, without me needing to delve too much into the details of the OrderedDict data structure (it’s basically a linked list, under the hood).

Enough said, here it is for future reference:

class MyOrderedDict(OrderedDict):
    def append(self, key, value):

    root = self._OrderedDict__root
    last = root[0]

    if key in self:
        raise KeyError
        root[0] = last[1] = self._OrderedDict__map[key] = [last, root, key]
        dict.__setitem__(self, key, value)

A web-app to give you your next bus arrival time

404 bus. Actually found!

You go to the bus stop and, by Murphy’s Law, either a) the bus has just gone or b) you’ll wait an endless amount of time for the bus to arrive.
Or at least this is what I experience most of the time.

That’s why I coded a small web-app that uses the TFL live stream data and HTML5 Geolocalization functionality to retrieve a list of the bus stops within 300 m from your location and the list of busses that should arrive from that moment on. And if it can’t find your location, you’re given the possibility to input a postcode, which, thanks to the data at OrdnanceSurvey and some good math, is converted into canonical coordinates and used to present with a similar result.

It was a very nice excuse for me to learn more about Flask (amazingly fast to use), HTML, CSS (first time I use Bootstrap!) and JS, and web-apps in general.

The solution is far from complete and several could be the improvements (handle errors with nice landing pages, allow the user to input a postcode regardless if geolocalization worked or not, add link to maps, …), but as it was more of an experiment, I’m happy with the result so far. It’s hosted at OpenShiftGive it a spin and check out the code.

Fear and loathing when compiling the kernel

Make menuconfig.

Carefully sift through the myriad of kernel options and select only the ones you need, dreaming of how fast your machine will be with the kernel you’re crafting.

Wait for your family to be asleep and spend the rest of the night on the interwebs to find out whether CONFIG_PPS_CLIENT_GPIO should be compiled as a module or built-in, if you should enable CONFIG_CAN as an excuse to buy that dongle you’ve been longing for a while and questioning every “if unsure, say N” – because you might well be unsure, but you certainly don’t want to miss out on that cool feature.

Be pleased with yourself when you have what you think is the perfect kernel for your box.

Save the config, but don’t use the default .config name: give it a different name for no particular reason, like kconfig, and exit.

make menuconfig again, load .config while you want to perform your last few tweaks. Be surprised when you find out CONFIG_HW_RANDOM_AMD, CONFIG_HW_RANDOM_INTEL and CONFIG_HW_RANDOM_VIA are all enabled, though you were sure all architecture-dependent options were already taken care of. Disable the two that don’t apply it and save as .config. Save it also as kconfig, just in case. Realise you’ve just overwritten the result of a dozen nights of lost sleep.

Count to ten, trying not to curse. Curse yourself when you reach 23. Exhale, slowly. Close the lid and get some sleep.

GOTO step 1 the day after.

Be even more pleased with yourself when you have another new shiny .config – also saved with different names to 8 other places, including a USB stick that you buried in the socks drawer and the SD card of your camera. Love the moment when you see you could shave off another handful of options and allow yourself to grin all day long when you go and come back to work. Remember you haven’t shaved since the first make menuconfig when a kid offers you a place on the bus, despite the fact you’re just 32.

make-kpkg your kernel and dpkg -i it. Smile and reboot.

Expect to select your kernel in GRUB and be a bit disappointed that you entirely missed the boot menu. Realise you’re still smiling anyway.

At the splash screen, wonder for a split second why neither keyboard nor touchpad work. Realise you can’t log in into the box and stop smiling suddenly.

Detach yourself and say to no one in particular “alright, I’ll just ssh in from the other box.” Try to remember whether sshd has started.

Start to realise a few things.

Realise in horror it’s not even installed.

Realise that you can’t reboot using the the older kernel, since the boot menu is hidden (you didn’t miss it: it really wasn’t there).

Realise you’re stuck. You’ve sort of locked yourself out your laptop. Realise this is the series of dumbest rookie errors you’ve done in ages.

Realise that, if there were an award for the stupidest thing, you’d win it hands down.

Give up on ssh and think about logging in using the on-screen keyboard. You still need a way to input characters.

Pray that you’ve enabled the support for a USB mouse. It’s kind of basic, but after what happened you’re doubting yourself about every single thing.

Locate in your head where you last saw a mouse in the house and run like there’s no tomorrow praying it’s still there. Find it and almost shed a tear of joy when you carry your mouse back to the laptop with both hands, like you just found the Graal.

Select each character of your 20-char password and be glad you’re alone and you won’t have to change your password(s) after these shenanigans.

Try to fire up the on-screen keyboard now that you’re logged in. Be lost wondering why this doesn’t work. Breathe slowly and resort to the char map. Admittedly not the best, decidedly better than nothing. Take three minutes to write sudo su and the password, char by char. As you can only find letters but no numbers or punctuation charaters in this stupid char map, open another shell and paste ll. Now you can copy paste numbers and some punctuation.

Can’t find the dash but you can write Devanagari in your shell

Open another shell and copy paste h i s t o r y. OK, now you’ve even got a selection of the most common commands. Realize you’re on the road to victory. Or at least, pretend so.

Now copy paste apt-get install sshd. Error. apt-cache search for it. Obviously it’s not sshd. It’s openssh-server. Install it. Error. Can’t find the repo. Move your head towards the top right corner and despair when you understand that wireless support is broken.

You will not be able to ssh in here. Ever.

Panic a bit and try a dpkg -P linux-image-4.1.6. Read a kind notice warning you you’re about to do a(nother) stupid(TM) thing. Wonder how you can select either ‘OK’ or ‘Cancel’ in this dialog, since you can’t paste control characters. Maybe this is a good thing after all.

Decide to edit the GRUB menu. Last time you checked, you needed to edit menu.lst. Learn that things have moved on and you now have to operate on /etc/defaults/grub. Open nautilus as root, thank the gods when you see that gedit is installed and regret all the times you said “I don’t need anything else outside vi“.

See that GRUB_HIDDEN_TIMEOUT_QUIET is set to true and GRUB_TIMEOUT is set to 0. Wonder why people think it’s always best when you hide things from sight. Change to true, 10. update-grub and reboot.

See the GRUB menu in all its uselessness, displaying a countdown from 10 with no other option.

You’re still on that crippled 4.1.6. Repeat all the weird stuff you just did, but at 10x speed. You’re now a semi-pro of on-screen keyboards, char maps and char copy/pasting.

Open /boot/grub/grub.cfg. Pause a second to thank for gedit again. See that the most-recent non-custom kernel is the third entry in the menu. So, without thinking twice set GRUB_DEFAULT from 0 to 2 in /etc/default/grub. Update-grub and reboot.

Accept that you messed up big time, as the system now enters the BIOS without even showing the useless countdown. You’re now on the path to desperation. Change boot settings and get yourself into a busybox environment. Feel like you’ve just seen the Matrix. At least now the keyboard works. <TAB> <TAB>. There are maybe 30 commands. vi is not one of them.

Think about mounting your partition, you’ll think how to modify grub.cfg once you get there.
Try to remember the layout of your HD. Screw it, just do mount -t ext4 /dev/sdaN /tmp/hd for N in 1 .. 6. Cry of joy at N=4. Try /tmp/hd/usr/bin/vi.tiny /tmp/hd/boot/grub/grub.cfg and be relieved watching vi compose the text of the file. Manually change every pointer to vmlinuz-4.1.6 to the last working kernel. Save. Reboot.

Back to normality.

Grep for ‘KEYBOARD‘ and ‘MOUSE‘ just for fun you kconfig (and all other 8 copies) and let your jaw drop when you see that neither CONFIG_KEYBOARD_ATKBD nor CONFIG_MOUSE_PS2 were selected.

GOTO beginning, again. Make menuconfig and go for another spin.

But not tonight.

Decoding RM4SCC for fun

I recently got curious about the bar code I could sometimes found on letters directed at me. I noticed there are just 4 symbols, begins and ends always in the same way, and is the same on all letters, regardless of the sender.

Armed with this basic information, after a bit of research I found out that the code is called Royal Mail 4-State Customer Code. Even more curious, I decided to write a simple decoder for it and, all of a sudden, all the knowledge in signal processing and telecommunications system retuned vivid in my mind, years after I took those classes (which I very much enjoyed, I must admit). Here is how I did it.

TL;DR: I put the code for the rm4sccdec (RM4SCC decoder) on GitHub. Use it at your own risk, as it’s not production-ready and needs some tweaking to reliably scan all types of image. I’ve used Python with OpenCV and numpy.

Step one: image pre-processing

The code does not include any information in the colour, which means we can simply get rid of the colour information and transform the image to greyscale.

Next, we want to maximise the “distance” between the information (the bars) and the noise (the background): this is usually done by thresholding the image. Using a global value for thresholding does not always give good results, especially when different areas of the image are characterised by different illumination. Some more advanced techniques, such as Otsu thresholding method (which I used in my decoder), are a better fit.

Finally, it is possible to have some residual noise, due to the thresholding process, whereby some white pixels are present in black areas and vice-versa. This is called salt’n’pepper noise and can effectively be filtered with median filters, which substitute the value of a pixel with the median of those around. The great advantage is that it preserves the edges of the image.

Step two: feature selection, extraction, and classification

Now we can start thinking about the features defining our symbols. We know we have 4 symbols, which we can call ascenderdescenderlong (Full Height, in the image), and short (Tracker, in the image).

The 4 symbols used in RM4SCC, from Wikipedia The 4 symbols used in RM4SCC, from Wikipedia

The first obvious feature we can select is the vertical position of each bar. After all, that’s the information we need to decode the codeword. However, if we choose the 4 points determining each bar, we’d probably end up complicating the decoding process too much.

An easy way out is to choose the centroid position (just its y-coordinate would be necessary) for each bar. Notice that, though, the long and short bars will share the same feature. If we go along this path, we need another feature to distinguish (at least) the long bar from the short bar. The second obvious feature is therefore the size or, more accurately, the area. This feature will allow us to distinguish easily long from short, but it will be pretty useless for the ascender and the descender.

For the extraction, we need to segment the image and find all the bars, and compute the so-called moments for each of them. The first three moments will be enough for us to get all the features we are interested in.

As a side note, as the segmentation function I have used does not return all segments in order, I had to extract the x-coordinate for each bar so as to be able and sort the vector of symbols.

If the code scanned is reasonably horizontal, we should be able to classify all four symbols pretty easily. For this bit I resorted to K-means clustering, although other classification methods can be used with similar results.

Step three: the actual decoding

If we don’t consider the starting and ending symbols, all symbols inbetween are grouped 4 by 4. For this reason we first need to build a dictionary that maps all valid combinations of 4-symbols group to the correct letter or number.

Finally, a bit of fun when computing the checksum. I translated the algorithm explained here, with the only difference that I wanted to avoid using yet another table to compute the final letter/number so instead I implemented the rule behind it (which boils down to ensuring ‘bit parity’).

Step four: enjoy it!

And possibly fork, improve and re-release :)