Alternative Backends

A "backend" is simply a program which mdbook will invoke during the book rendering process. This program is passed a JSON representation of the book and configuration information via stdin. Once the backend receives this information it is free to do whatever it wants.

See Configuring Renderers for more information about using backends.

The community has developed several backends. See the Third Party Plugins wiki page for a list of available backends.

Setting Up

This page will step you through creating your own alternative backend in the form of a simple word counting program. Although it will be written in Rust, there's no reason why it couldn't be accomplished using something like Python or Ruby.

First you'll want to create a new binary program and add mdbook as a dependency.

$ cargo new --bin mdbook-wordcount
$ cd mdbook-wordcount
$ cargo add mdbook

When our mdbook-wordcount plugin is invoked, mdbook will send it a JSON version of RenderContext via our plugin's stdin. For convenience, there's a RenderContext::from_json() constructor which will load a RenderContext.

This is all the boilerplate necessary for our backend to load the book.

// src/main.rs
extern crate mdbook;

use std::io;
use mdbook::renderer::RenderContext;

fn main() {
    let mut stdin = io::stdin();
    let ctx = RenderContext::from_json(&mut stdin).unwrap();
}

Note: The RenderContext contains a version field. This lets backends figure out whether they are compatible with the version of mdbook it's being called by. This version comes directly from the corresponding field in mdbook's Cargo.toml.

It is recommended that backends use the semver crate to inspect this field and emit a warning if there may be a compatibility issue.

Inspecting the Book

Now our backend has a copy of the book, lets count how many words are in each chapter!

Because the RenderContext contains a Book field (book), and a Book has the Book::iter() method for iterating over all items in a Book, this step turns out to be just as easy as the first.


fn main() {
    let mut stdin = io::stdin();
    let ctx = RenderContext::from_json(&mut stdin).unwrap();

    for item in ctx.book.iter() {
        if let BookItem::Chapter(ref ch) = *item {
            let num_words = count_words(ch);
            println!("{}: {}", ch.name, num_words);
        }
    }
}

fn count_words(ch: &Chapter) -> usize {
    ch.content.split_whitespace().count()
}

Enabling the Backend

Now we've got the basics running, we want to actually use it. First, install the program.

$ cargo install --path .

Then cd to the particular book you'd like to count the words of and update its book.toml file.

  [book]
  title = "mdBook Documentation"
  description = "Create book from markdown files. Like Gitbook but implemented in Rust"
  authors = ["Mathieu David", "Michael-F-Bryan"]

+ [output.html]

+ [output.wordcount]

When it loads a book into memory, mdbook will inspect your book.toml file to try and figure out which backends to use by looking for all output.* tables. If none are provided it'll fall back to using the default HTML renderer.

Notably, this means if you want to add your own custom backend you'll also need to make sure to add the HTML backend, even if its table just stays empty.

Now you just need to build your book like normal, and everything should Just Work.

$ mdbook build
...
2018-01-16 07:31:15 [INFO] (mdbook::renderer): Invoking the "mdbook-wordcount" renderer
mdBook: 126
Command Line Tool: 224
init: 283
build: 145
watch: 146
serve: 292
test: 139
Format: 30
SUMMARY.md: 259
Configuration: 784
Theme: 304
index.hbs: 447
Syntax highlighting: 314
MathJax Support: 153
Rust code specific features: 148
For Developers: 788
Alternative Backends: 710
Contributors: 85

The reason we didn't need to specify the full name/path of our wordcount backend is because mdbook will try to infer the program's name via convention. The executable for the foo backend is typically called mdbook-foo, with an associated [output.foo] entry in the book.toml. To explicitly tell mdbook what command to invoke (it may require command-line arguments or be an interpreted script), you can use the command field.

  [book]
  title = "mdBook Documentation"
  description = "Create book from markdown files. Like Gitbook but implemented in Rust"
  authors = ["Mathieu David", "Michael-F-Bryan"]

  [output.html]

  [output.wordcount]
+ command = "python /path/to/wordcount.py"

Configuration

Now imagine you don't want to count the number of words on a particular chapter (it might be generated text/code, etc). The canonical way to do this is via the usual book.toml configuration file by adding items to your [output.foo] table.

The Config can be treated roughly as a nested hashmap which lets you call methods like get() to access the config's contents, with a get_deserialized() convenience method for retrieving a value and automatically deserializing to some arbitrary type T.

To implement this, we'll create our own serializable WordcountConfig struct which will encapsulate all configuration for this backend.

First add serde and serde_derive to your Cargo.toml,

$ cargo add serde serde_derive

And then you can create the config struct,

#![allow(unused)]
fn main() {
extern crate serde;
#[macro_use]
extern crate serde_derive;

...

#[derive(Debug, Default, Serialize, Deserialize)]
#[serde(default, rename_all = "kebab-case")]
pub struct WordcountConfig {
  pub ignores: Vec<String>,
}
}

Now we just need to deserialize the WordcountConfig from our RenderContext and then add a check to make sure we skip ignored chapters.

  fn main() {
      let mut stdin = io::stdin();
      let ctx = RenderContext::from_json(&mut stdin).unwrap();
+     let cfg: WordcountConfig = ctx.config
+         .get_deserialized("output.wordcount")
+         .unwrap_or_default();

      for item in ctx.book.iter() {
          if let BookItem::Chapter(ref ch) = *item {
+             if cfg.ignores.contains(&ch.name) {
+                 continue;
+             }
+
              let num_words = count_words(ch);
              println!("{}: {}", ch.name, num_words);
          }
      }
  }

Output and Signalling Failure

While it's nice to print word counts to the terminal when a book is built, it might also be a good idea to output them to a file somewhere. mdbook tells a backend where it should place any generated output via the destination field in RenderContext.

+ use std::fs::{self, File};
+ use std::io::{self, Write};
- use std::io;
  use mdbook::renderer::RenderContext;
  use mdbook::book::{BookItem, Chapter};

  fn main() {
    ...

+     let _ = fs::create_dir_all(&ctx.destination);
+     let mut f = File::create(ctx.destination.join("wordcounts.txt")).unwrap();
+
      for item in ctx.book.iter() {
          if let BookItem::Chapter(ref ch) = *item {
              ...

              let num_words = count_words(ch);
              println!("{}: {}", ch.name, num_words);
+             writeln!(f, "{}: {}", ch.name, num_words).unwrap();
          }
      }
  }

Note: There is no guarantee that the destination directory exists or is empty (mdbook may leave the previous contents to let backends do caching), so it's always a good idea to create it with fs::create_dir_all().

If the destination directory already exists, don't assume it will be empty. To allow backends to cache the results from previous runs, mdbook may leave old content in the directory.

There's always the possibility that an error will occur while processing a book (just look at all the unwrap()'s we've written already), so mdbook will interpret a non-zero exit code as a rendering failure.

For example, if we wanted to make sure all chapters have an even number of words, erroring out if an odd number is encountered, then you may do something like this:

+ use std::process;
  ...

  fn main() {
      ...

      for item in ctx.book.iter() {
          if let BookItem::Chapter(ref ch) = *item {
              ...

              let num_words = count_words(ch);
              println!("{}: {}", ch.name, num_words);
              writeln!(f, "{}: {}", ch.name, num_words).unwrap();

+             if cfg.deny_odds && num_words % 2 == 1 {
+               eprintln!("{} has an odd number of words!", ch.name);
+               process::exit(1);
+             }
          }
      }
  }

  #[derive(Debug, Default, Serialize, Deserialize)]
  #[serde(default, rename_all = "kebab-case")]
  pub struct WordcountConfig {
      pub ignores: Vec<String>,
+     pub deny_odds: bool,
  }

Now, if we reinstall the backend and build a book,

$ cargo install --path . --force
$ mdbook build /path/to/book
...
2018-01-16 21:21:39 [INFO] (mdbook::renderer): Invoking the "wordcount" renderer
mdBook: 126
Command Line Tool: 224
init: 283
init has an odd number of words!
2018-01-16 21:21:39 [ERROR] (mdbook::renderer): Renderer exited with non-zero return code.
2018-01-16 21:21:39 [ERROR] (mdbook::utils): Error: Rendering failed
2018-01-16 21:21:39 [ERROR] (mdbook::utils):    Caused By: The "mdbook-wordcount" renderer failed

As you've probably already noticed, output from the plugin's subprocess is immediately passed through to the user. It is encouraged for plugins to follow the "rule of silence" and only generate output when necessary (e.g. an error in generation or a warning).

All environment variables are passed through to the backend, allowing you to use the usual RUST_LOG to control logging verbosity.

Wrapping Up

Although contrived, hopefully this example was enough to show how you'd create an alternative backend for mdbook. If you feel it's missing something, don't hesitate to create an issue in the issue tracker so we can improve the user guide.

The existing backends mentioned towards the start of this chapter should serve as a good example of how it's done in real life, so feel free to skim through the source code or ask questions.