james mckay dot net
because there are few things that are less logical than business logic

You have to tell AWS CLI that your EC2 instance is not in Virginia

Here’s a little gotcha with AWS that I keep running into time and time again. By default, the aws command line interface, and AWS API libraries such as boto3, will always use the us-east-1 (Virginia) region by default, even when running on EC2 instances in other regions.

This is not what you expect, and it is almost never what you want.

There is an issue on the awscli GitHub issue tracker to fix this, but it is still open four years after first being raised, with no indication when (or even whether) it will ever be addressed.

User @BradErz suggests including these lines in your user_data to set the default region:

region=$(curl|grep region|awk -F\" '{print $4}')
echo "[default]" > /root/.aws/config
echo "region = ${region}" >> /root/.aws/config

Note however that this will only set the default region for the root user; you will need to configure aws-cli separately for any other logins on your instance.

Annoying as this behaviour is, I would be surprised to see it fixed any time soon, as it would be a breaking change.

Design refresh

I updated the design of my blog over the weekend.

My main goal was to switch back to a new, responsive version of my original blog theme, with its orange and blue colour scheme. Since the start of this year I’d been trying out a variety of off-the-shelf WordPress themes to give me a responsive (and therefore more SEO-friendly) design, but I was never really satisfied with any of them, so orange and blue is back, with its first major design refresh since 2011.

I’d thought that making it responsive would be a massive undertaking, but in the end of the day it only took me a couple of hours on Saturday morning. I was helped greatly in this by the fact that I was already using Less CSS as a pre-processor. I was also able to use Google’s mobile friendly test tool to quickly identify and fix the issues that needed fixing.

If you shrink the window down to below 400 pixels, you’ll see that that the text in the header starts to shrink with it to fit. I’d seen a few other sites do this, and it turns out to be very simple to implement, using a combination of media queries and viewport units. In case you’re interested, here is the Less CSS mixin that I’m using to achieve this:

.responsive-font-size(@size, @resize-below) {
    font-size: @size;
    @media screen and (max-width: @resize-below) {
        font-size: unit(@size * 100 / @resize-below, vw);

I’ve also restored my old blog posts, together with most of the comments and attached pictures, from a backup that I’d forgotten that I had. In the process of doing so, I’ve updated all the internal hyperlinks and image locations to point to https:// URLs on my blog; this was needed to eliminate mixed-content security warnings in the browser address bar on older posts that contained images. I achieved this quite simply and elegantly by using sed on the output of mysqldump to perform an appropriate find and replace before reloading. The exact command to use is left as an exercise for the reader.

Behind the scenes, I’ve moved it onto a $5/month 512MB DigitalOcean droplet. This is all you need for a blog that only gets a hundred or so hits a day, and a bit of load testing with Apache JMeter suggested to me that it should be able to handle a spike from Hacker News if necessary — apparently hitting the HN home page can get you about 6,000 hits an hour. I’ve scripted the server setup using Terraform and I’ve also got a couple of scripts to backup and restore the data. This means that I can tear down and rebuild it very quickly if need be, in accordance with the modern best practice of treating your servers as cattle rather than pets.

If you have any problems with it, or if anything doesn’t look right, please let me know. If you want to be reminded what the old version looked like, here are some archive.org snapshots for 2007, 2010, 2011, 2012, and 2016.

Introducing Lambda Tools: a new framework for deployment to AWS Lambda

Lambda Tools is a new project that I’ve been working on over the past few weeks or so. It is a build and deployment toolkit, written in Python, to make it easier to work with AWS Lambda functions. It started out as a user story in our backlog at work but since then it’s grown into a full blown open source project with a life of its own.

The “serverless” model offered by AWS Lambda is a useful and potentially cost-effective one, especially for scheduled tasks that only need to run every so often and don’t require a lot of resources. The “free tier” doesn’t expire at the end of your initial twelve month trial, which is an added bonus.

The downside is that it can be tricky to work with. If your function requires additional libraries, it starts to get a bit more complex, and if any of these are written partly in C and require special compilation, things can get pretty messy. On top of that, you will want to write unit tests for your functions and set up some sort of Continuous Delivery pipeline for them.

Lambda Tools includes several features to make these things easier. For example, it gives you an option to build your function in a Docker container to avoid these messy “invalid ELF header” errors. You configure it simply by creating a YAML file called aws-lambda.yml, which might look like this for example:

version: 1

    runtime: python3.6
      source: src/hello_world
        - file: requirements.txt

      handler: hello.handler
      role: service-role/NONTF-lambda

      description: A basic Hello World handler
      memory_size: 128
      package: build/hello_world.zip
      region: eu-west-2
      timeout: 60

        Account: Marketing
        Application: Newsletters

Here are some of the other ideas that I’m thinking of implementing for it:

  • A unit test runner
  • Support for Python 2.7
  • Support for languages other than Python (.NET, Node.js, and so on)
  • Integration with Terraform:
    • The ability to plug it into Terraform’s external data source provider
    • The ability to read configuration from stdin
    • A scaffolding engine to generate Terraform modules on the fly
  • Integration with triggers such as CloudWatch cron jobs, API Gateway, and so on
  • A full-blown sample application
  • The ability to include or exclude specific files when building your package

At the moment it only supports Python 3.6 but nevertheless it is in a usable state. You can install it using pip install lambda_tools and it includes full instructions in the readme file on the GitHub repository. Additionally, if you’re interested in getting involved with its development, feel free to fork the project and send me a pull request or three. If it’s anything more complex, raise a ticket on the GitHub issue tracker and we’ll chat about it there.

Necessary and sufficient conditions

Take a look at these two statements. Are they both saying the same thing?

  1. “If you are using HTTPS, then your website is secure.”
  2. “If you are not using HTTPS, then your website is not secure.”

In actual fact, they are not. Furthermore, only the second statement is true: the first statement is false.

The first statement is an example of a sufficient condition. If it were true, all you would need to do to secure your website would be to install an SSL certificate and you’d be done.

The second statement, on the other hand, is an example of a necessary condition. There are, of course, other things you need to do to ensure that your website is secure: for example, take care to avoid SQL injection and cross-site scripting attacks, keep your servers patched and up to date, and so on. But you still need to use HTTPS in addition to all these. If you don’t, your site will be vulnerable to a man-in-the-middle attack.

You can see the difference if I draw up a truth table for a sufficient condition:

Sufficient condition Other stuff Secure?
No No No
Yes No Yes
No Yes Maybe
Yes Yes Yes

On the other hand, a necessary condition looks like this:

Necessary condition Other stuff Secure?
No No No
Yes No Maybe
No Yes No
Yes Yes Maybe

Some conditions can be both necessary and sufficient. In this case, the truth table looks like this:

Necessary and
sufficient condition
Other stuff Secure?
No No No
Yes No Yes
No Yes No
Yes Yes Yes

A necessary and sufficient condition can be written as “if and only if.” This is sometimes shortened to “iff.”

Insufficient does not mean unnecessary.

The most common misunderstanding that people have about necessary and sufficient conditions is the mistaken belief that one implies the other. Or that a lack of one implies a lack of the other.

  • It is possible for conditions to be sufficient but not necessary.
  • It is possible for conditions to be necessary but not sufficient.

Take, for example, this comment:

Google is just a bully because it is so big. It can go f*** itself. A standard webpage is not insecure and the use of SSL doesn’t make it secure either. Maybe everyone forgets that when SSL certs were comprised. I do work on e-commerce sites and I have seen clients who sites got hacked, not because of lack of SSL, but because of bad code on their backend. The hackers proceeded to add code so they would get emailed the credit card info after it was submitted. The user would never know, because the big green icon in the browser said it was secure. The whole thing is just a way for companies to make money.

This commenter correctly realised that SSL is insufficient but he then assumed that this means that SSL is therefore unnecessary. This is of course incorrect. SSL may be insufficient, but it is very, very necessary.

Unfortunately, in the world of IT security, there are plenty of necessary conditions. But there are no sufficient ones.

Programmatically starting an AWS instance with an encrypted EBS volume attached

I had to start some EC2 instances programmatically from inside an AWS lambda function. My code looked something like this:

import boto3

def handler(event, context):
    client = boto3.client('ec2')
    client.start_instances(InstanceIds=['i-0123456789abcdef0', ...])

This worked fine when I ran it from the command line, but when I ran it from inside the lambda, one particular instance stubbornly refused to start, even though the lambda ran without errors.

It turned out that the problem was a permissions issue. This particular instance had an additional encrypted EBS volume attached. The call to start_instances() was failing silently.

To fix this, make sure that the role under which your code runs is granted the kms:CreateGrant permission.

It took me a bit of trial and error to figure out which permission to add, but I wanted to make sure I got this one right. You should never give your code any more permissions than the bare minimum it needs in order to do what it needs to do. Unfortunately, figuring out exactly which permissions your code needs to run can sometimes be a bit of a challenge…