Collecting image data for training machine learning models can take precious time and lots of Google image searches.
I built a tool that generates labeled data for you. Choose your categories and download.
The first draft of Emoji Salad, our Emoji Pictionary bot, was a monolithic Node.js app. The server would respond only when a request came in via SMS. If there was a critical error, the whole server would fall on our heads, making it difficult to diagnosis the error and impossible to continue conversations. On top of that, deployments were nightmares that brought the possibility of disrupted conversations, if messages were received during a deployment.
Our current implementation uses a microservice based approach:
Moving to a services-based architecture has brought some big wins for us:
Isolating each service to its core functionality has made it easier to reason about what each one does. Since each service must interact through a defined API, we can drastically refactor one service without affecting any of the others.
For example, our first draft used an HTTP-like model, where server routes would be matched via regex against incoming messages. At a certain point in development, we realized we wanted the bot to be able to initiate messages to users, things like
Hey, we’re still waiting on that clue submission!
and gentle encouragements like:
This clue’s a doozy! Don’t forget, the emojis are 📽🤖❤️🤖
With a service-oriented architecture, we were able to rebuild how the Bot service processed messages, so it could spin itself to initiate messages to users, without affecting the other services (or even having to redeploy them).
Also, since we’ve done the work of defining how exactly the Bot expects to receive incoming messages, adding support for other messaging platforms (like Facebook Messenger) is a cinch.
With isolated services, it’s easier to write targeted tests.
By requiring each service to have a well-designed API for interaction, it becomes easier to make sure you’re testing the right things. Smaller services should have fewer dependencies, too. And as an added bonus, service-only test suites should take less time to run, which means you can run them more often.
We do have extension integration tests across all the services, which is a topic for another article.
It’s easier to deploy individual services, and fix them when they go down. For instance, our message queues rarely change; we have the confidence to know that if the Bot itself goes down, we’ll still be collecting incoming messages from users, and we’ll be able to handle those messages once the Bot comes back online.
Services do introduce overhead around managing multiple services and their dependencies, but Vagrant and Docker go a long way towards smoothing out that process.
There’s plenty more refactoring we could do; for instance, we’re eager to break apart our Bot service further, which is currently handling both the message parsing and the logic of the script driving the bot interactions.
Overall, moving to a services-based architecture has allowed us to iterate faster with much more overall stability, meaning a better bot experience from start to finish.