To Stalk a Muni

April 22, 2015
Bus Prediction LED

Bus Prediction LED

You’re waiting for the bus. The prediction monitor at the stop says five minutes, four, three… then just as you’re getting excited it changes its estimation to half an hour. Perhaps it’s raining and dusk and everyone is getting angry. Why are you standing here, why did the prediction API tell everyone it was time to get out to the bus stop? Can predictions be improved?

I’ve decided to find out. The first step to more accurate muni predictions is to accumulate historical data. With weeks’ worth of detailed bus position data we can find the patterns which cause errors in the traditional prediction algorithm.

Happily the realtime position and prediction API is available for free and without so much as a consumer key. The historical information is not available but we can write some code to collect it.

Here’s where Amazon EC2 is our friend. We can spin up a micro instance, set a cron task, and then relax as it does the work. An easy and reproducible way to build an EC2 machine for the job is to use Packer and Chef. We want our machine to hit the Nextbus API on a cron and save the results to S3 for later processing. That means we have to install the aws command line tool for S3 uploading, include our script and set the cron. I’ve taken care of all this at begriffs/stalk27 and the repo includes instructions to deploy so you can try it yourself.

My data collection philosophy is to save as much information as possible when sampling the API. Rather than picking relevant parts out of the XML document I’ve opted to save the whole thing. In fact I save the entire HTTP response header payload along with the document body. I think of data collection like a Mars mission – we can never go back in time with this data source so we should err on the side of over-collecting.

The next consideration is to choose dumb reliable storage and decouple the storage from the worker machine. Rather than save into a fancy database I just create a suitably-named file for each response. Also I avoid the local filesystem. The output is independent from the lifespan of the worker. While the worker runs it collects, else the data merely persists.

The only missing piece is monitoring and alerting. If this project were crucial and truly intolerant of missing data it should have some failover, redundancy and monitoring. For those kind of requirements I would choose begriffs/microservice-template and enable Amazon CloudWatch.

After two weeks I will begin analysis and will share with you what the data reveals about bus delays. I am collecting data on route 27 because I’ve had bad experiences waiting for wildly late buses at Powell station. If the new prediction algorithm ends up working well then I’ll scale up the data collection to more routes and maybe create a web interface for other people to check accurate departure times.