Swift Talk Backend
This is the source code of the Swift Talk backend: https://talk.objc.io
While we abstracted non-app-specific parts away into frameworks, this is not a web framework. Here's a minimal description of the structure:
SwiftTalkServerLib
This framework contains the application-specific code. There are a few main parts to it:
- The routes define the available endpoints and transform URLs to routes and routes to URLs
- The interpret methods contain the application logic for each route
- The views contain rendering logic
- The queries abstract away the database (but only a little bit)
- The third-party services communicate with JSON and XML APIs of third-party providers
Interpreting
For testability (and because we wanted to experiment), we wrote our route interpreter using the final-tagless style. This allows us to write a normal interpreter that does the usual web-server things: execute database queries, perform network requests, etc. It also allows us to have a test interpreter, so that we can write high-level flow tests (with easy network tests).
Database
We use PostgreSQL and write standard SQL queries to access the database. We represent each table with a struct and use Codable to help generate simple queries and to parse the results from PostgreSQL back into struct values (Episode #114).
Third-Party Services
Rather than depending on third-party frameworks, we decided to write our own wrappers around the REST endpoints of third-party services (e.g. GitHub, Recurly, Sendgrid, Vimeo) using our tiny networking library.
HTML
The HTML framework defines an enum to represent HTML/XML nodes. There is one special feature: a Node
is generic over some read-only state. This allows us to pass around "global" state like a CSRF token and session/user data without actually making that global, and without having to explicitly pass it around everywhere.
For an example, see HTMLExtensions.swift. We add multiple extension to our Node
type when the read-only state is of type STRequestEnvironment
.
Routing
For routing, we use a Router
struct that captures both parsing and generating a route in one. Our routes are defined as enums, and using the Router
we can write one description that converts the case into a URL and parses a URL, without having too worry too much about keeping them in sync.
We also use the enum cases to generate links, making sure that every link is well-formed and has all the necessary parameters.
Incremental
We use our Incremental programming library to transform and cache static data. For example, when the markdown file for an episode is changed, we recompute the highlighted version (highlighting is done using a SourceKitten
wrapper). Because this can take a little while, the results are cached.
NIOWrapper
This framework is a lightweight wrapper around SwiftNIO, which contains a few primitives to write data, send redirects, process POST data, etc. The wrapper depends only minimally on NIO, which makes it easy to test without NIO.
WebServer
The WebServer framework builds on top of the NIOWrapper, providing some higher level abstractions e.g. to write HTML or JSON responses. It also integrates the with the Database and Networking frameworks to provide response APIs to execute queries or call third-party network endpoints.
Installation Notes
Dependencies
If you want to run this locally (without Docker), you need to install the following dependencies:
- postgresql
- libpq-dev
- cmake
- cmark
- curl
PostgreSQL
You need PostgreSQL and libpq. To set up a local postgres instance:
initdb -D .postgres
chmod 700 .postgres
pg_ctl -D .postgres start
createdb swifttalk_dev
Compiling Assets
To build the stylesheets:
./build-css.sh
Deployment
We deploy to a heroku-based docker app (needs postgres as well).
If you get a "basic auth" error: heroku container:login
heroku container:push web
heroku container:release web
Running in Docker
For the docker container to be able to access PostgreSQL on the host, you have to allow access via TCP/IP.
Add host all all 0.0.0.0/0 trust
to pg_hba.conf (this opens up the PostgreSQL instance to everybody in your network, use something more finegrained if that's a problem) and add listen_addresses = '*'
to postgresql.conf.
docker run -a stdin -a stdout -i -t --env-file .env --env RDS_HOSTNAME=(ifconfig en1 | awk '/inet /{print $2}') -p 8765:8765 swifttalk-server
You could also set up a multi-container docker application. For example, like in this pull request.
Debugging Linux Bugs
You can run a docker container from one of the intermediate steps. Then install screen and vim, and you have a small linux dev environment.
https://medium.com/ihme-tech/troubleshooting-the-docker-build-process-454583c80665
Here are some steps, if you have a local postgres running (on en1
):
docker build -t "swifttalk-server-debug" -f Dockerfile.debug .
docker run --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --security-opt apparmor=unconfined -a stdin -a stdout -i -t --env-file .env --env RDS_HOSTNAME=(ifconfig en1 | awk '/inet /{print $2}') -p 8765:8765 swifttalk-server-debug bash
screen # this helps with having separate "windows"
swift build --configuration debug -Xswiftc -g
./build/debug/swifttalk-server &> output.log
# Wait for a crash to happen
./symbolicate-linux-fatal .build/debug/swifttalk-server output.log
Unfortunately, it doesn't seem like lldb is expected to work under Linux: https://bugs.swift.org/browse/SR-755