In this post I’m going to describe, in non-technical terms, the basic structure of a web service like Facebook. Last week I wrote a post about how to get started in programming, and before that I wrote about some trends in system design. I intend that anyone who follows along with these posts will come away with enough understanding of computer systems to be conscious of the social, economic, and political issues they raise. But before we can talk about the social, economic, and political dimensions, we need to define some terms, and in this post we’re going to look at web service and web app.
I’m adopting a broad definition of web service here: a web service is any system where information has to travel between two or more computers for the system to function. I’m going to focus on web apps, which are a very common type of web service that provide a UI on a website as the main venue for interacting with the service. Web apps are usually understood to be interactive– a site like this one in September 2020, that does not provide any way for readers to interact with what’s here, would be a web service under my definition but it wouldn’t be a web app. By my definition any site where you can log in is automatically a web app. This includes highly interactive systems like Facebook as well as less interactive systems like Craigslist and Wikipedia.
Throughout this post, I’ll use a fictional “To-Do” app as an example. This type of app helps you keep track of your task list. You log in to the To-Do app through a website, and you’re able to add, edit, and check off items from your personal to-do list. For the purposes of this discussion, I’m also going to say that the app stores your data–that is, even if you close the browser window you started out in, you can open a new one, log back in, and your list will still be there. If you google “todo app,” you’ll find many apps that fit this description; I’m not going to link any because the specifics don’t really matter. I’m going to cover the parts of the app in a sort of barebones way–these are the things that make the app an app; whether it’s a good app is a different post.
An app like this can be divided into six areas: access control, user interface (UI), API and domain logic, application data, logging, and analytics. The access control consists of the user management and permission system–who can log in, and what permission they have to see and interact with data. The UI is all of the code and markup that tells your browser what to display (this includes assistive technologies–the way your website interacts with a screen reader for someone with a visual impairment is equally part of “UI”). The API and domain logic is the code on the server that handles interactions like “add a task” or “mark this task complete.” The application data is the actual data storage system used to save your todo items–probably a database, but other types of storage, like regular text files on a hard drive, could also do the same job. The logging consists of the records that the system keeps about what it does; logs are usually written to a text file full of lines like the following:
2020/09/13 03:30:50.789 [INFO] User testUser added Todo item with ID=a4th983e
The analytics system consists of the tools used by whoever is running the app to interpret the logs written by the system and answer questions like “when is the highest traffic period” or “what features do the people who use this app interact with most?” or “where do people who use this app encounter errors?”
Access control includes Authentication, Authorization, and Audit. Authentication (Authn) is the way that the system verifies that you are who you say you are–a username and password, a code sent to your phone, etc. Authorization (Authz) is the way the system decides what you’re allowed to do. For instance, if you’re logged in to Instagram, you are authenticated (Instagram trusts that you are the owner of your account) but unless you have 10k followers you are not authorized to include links in your stories–only users with more than 10k followers are allowed to do that. Audit (the ability to go back and look up who did what in the system) is also included in access control, but for the purposes of this discussion it’s included later as a specialized form of logging.
The UI includes both markup and code. Markup is text that includes hints about how it should be displayed–the text is “marked up” with formatting data. If you imagine an old-timey newspaper editor walking over to the typesetter’s desk with a pile of handwritten articles, the articles would have the actual text, but they would also have notes for the typesetter like “this goes on the front page” or “this headline needs to be really big” or “this ad should be business-card sized.” That whole thing–the text and the formatting information–is the markup. On the internet, HyperText Markup Language (HTML) is the dominant markup format, and it is augmented by Cascading Style Sheets (CSS), which carries additional formatting information.
API and Domain Logic
The code that runs on a web app’s UI needs to be able to talk to a server to do its job. The Application Programming Interface (API) is the code that runs on the server, listens for requests, and answers them. The most basic and generic part of this system is the web server–the part that sends the client a web page: the bundle of markup and code that makes up the UI. Since this bundle is usually the same for all users (everyone’s Facebook UI has the same structure, even though they see different content) there is sometimes a Content Delivery Network (CDN) that tries to speed up delivery of these standard parts.
Unlike the main markup-and-code bundle, any user-specific content (like the items in your feed or your profile information) is served by an application server. Rather than send whole web pages, the application server listens for requests like “mark this task as done” or “show me my outstanding tasks.” To answer these questions, the application server will read the request, check that the requester is allowed to perform the action (using the Authn and Authz system), and then it might get data out of the data store (e.g. the list of todo items) or it might modify the data (e.g. to mark a task as complete). Often when a programmer says that they work on “back-end” code, this is what they mean (the “front-end” means the UI).
The data that the application uses to do its job is what I define as application data. In the todo app, this would include information about users, the information about individual items in people’s task lists, etc. It would not include things like audit logs and server logs–those are data about the application, but not used by the application.
Data storage is a very deep subject, but it all boils down to what you want to store and how you plan to use it. Some types of databases are very fast at responding to requests to read data, but much slower at processing requests to add or modify data. Others are very good at handling large files like images or videos, but not as good at manipulating that data in complex ways. Different data storage systems also make different tradeoffs when things go wrong–some databases are designed to return errors unless several conditions can all be satisfied, while others are designed to tolerate certain types of failures. For instance, when a financial database needs to move money between two accounts, it is designed so that if either the withdrawal from the source or the deposit to the destination fails, the whole transaction gets reversed. However, a database that tells you how many likes a facebook post has might give you the most recent data that it can find, even if it’s unable to verify that it’s exactly up to date.
A single application might use more than one data store if it has to support different data access patterns.
Logging refers to the records that an application saves about what it does. At various
points within front-end and back-end code, the programmer writes instructions for the
program to output information. What this information looks like is up to the programmer
and varies based on the situation. An access control system’s audit log is a highly structured
document, making it possible for an analytics program to quickly determine things like
when a specific login or logout happened. An application server’s log might be much less
structured; each line might be a descriptive statement about something that the programmer
thought would be useful to know, like
TestUser signed in or
wtf foo=true?!?! or, in
certain exceptional cases
1. raph, remember to remove this log before merging.
Technically, most logging is actually part of one of the other subsystems–the UI code writes some logs, the API code writes others, the access control system writes others. But the logs as a body of information are so crucial to understanding how and if the system is working that they deserve to be treated separately.
Logs are also incredibly powerful tools for learning about people. Logs can tell system operators what pages people visited, how log they stayed, and sometimes what they were looking at. This makes log data the most important subject to consider when thinking about the privacy of people who use a system.
The analytics system is what tries to use the logs and other data generated about the application to give the operators insight about what the system is doing. The analytics might warn about errors, display traffic graphs, or help guide strategic decisions about what to build. A simple low-traffic app might not have any dedicated analytics at all; a system like Facebook would have much more investment and complexity in its analytics than in what the people who use it would describe as the “product.”
This is a pretty rough-and-ready description of the structure of an app as it would be deployed. I’ve left out important disciplines like testing and (general) security, which are not a specific part of the deployed application but which influence each part in a different way. But I would argue that this is also, in important ways, a complete definition– anything that has all these parts working together is definitively An App. It might be insecure, it might be full of bugs and crash all the time, it might not be useful, but it would have all the necessary parts for me to consider it a web app. And I find that it’s often very useful to have a list like this when I’m trying to start out designing something, so that I can make sure that I’ve at least thought through all of the individual aspects of the thing, even if I haven’t figured out all the details.