Skip to content

atdgen-tutorial.rst

Latest commit

 

History

History
1514 lines (1131 loc) · 39.9 KB

atdgen-tutorial.rst

File metadata and controls

1514 lines (1131 loc) · 39.9 KB

Tutorial

What is atdgen?

Atdgen is a tool that derives OCaml boilerplate code from type definitions. Currently it provides support for:

  • JSON serialization and deserialization.
  • Biniou serialization and deserialization. Biniou is a binary format extensible like JSON but more compact and faster to process.
  • Convenience functions for creating and validating OCaml data.

What are the advantages of atdgen?

Atdgen has a number of advantages over its predecessor json-static which was based on Camlp4:

  • produces explicit interfaces which describe what is available to the user (.mli files).
  • produces readable OCaml code that can be easily reviewed (.ml files).
  • produces fast code, 3x faster than json-static.
  • runs fast, keeping build times low.
  • same ATD definitions can be used to generate code other than OCaml. See for instance atdj which generates Java classes for JSON IO. Auto-generating GUI widgets from type definitions is another popular use of annotated type definitions. The implementation of such code generators is facilitated by the atd library.

Prerequisites

This tutorial assumes that you are using atdgen version 1.5.0 or above. The following command tells you which version you are using:

$ atdgen -version
1.5.0

The recommended way of installing atdgen and all its dependencies is with opam:

$ opam install atdgen

Getting started

From now on we assume that atdgen 1.5.0 or above is installed properly.

$ atdgen -version
1.5.0

Type definitions are placed in a .atd file (hello.atd):

type date = {
  year : int;
  month : int;
  day : int;
}

Our handwritten OCaml program is hello.ml:

open Hello_t
let () =
  let date = { year = 1970; month = 1; day = 1 } in
  print_endline (Hello_j.string_of_date date)

We produce OCaml code from the type definitions using atdgen:

$ atdgen -t hello.atd     # produces OCaml type definitions
$ atdgen -j hello.atd     # produces OCaml code dealing with JSON

We now have _t and _j files produced by atdgen -t and atdgen -j respectively:

$ ls
hello.atd  hello.ml  hello_j.ml  hello_j.mli  hello_t.ml  hello_t.mli

We compile all .mli and .ml files:

$ ocamlfind ocamlc -c hello_t.mli -package atdgen
$ ocamlfind ocamlc -c hello_j.mli -package atdgen
$ ocamlfind ocamlopt -c hello_t.ml -package atdgen
$ ocamlfind ocamlopt -c hello_j.ml -package atdgen
$ ocamlfind ocamlopt -c hello.ml -package atdgen
$ ocamlfind ocamlopt -o hello hello_t.cmx hello_j.cmx hello.cmx -package atdgen -linkpkg

And finally we run our hello program:

$ ./hello
{"year":1970,"month":1,"day":1}

Source code for this section

Inspecting and pretty-printing JSON

Input JSON data:

$ cat single.json
[1234,"abcde",{"start_date":{"year":1970,"month":1,"day":1},
"end_date":{"year":1980,"month":1,"day":1}}]

Pretty-printed JSON can be produced with the ydump command:

$ ydump single.json
[
  1234,
  "abcde",
  {
    "start_date": { "year": 1970, "month": 1, "day": 1 },
    "end_date": { "year": 1980, "month": 1, "day": 1 }
  }
]

Multiple JSON objects separated by whitespace, typically one JSON object per line, can also be pretty-printed with ydump. Input:

$ cat stream.json
[1234,"abcde",{"start_date":{"year":1970,"month":1,"day":1},
"end_date":{"year":1980,"month":1,"day":1}}]
[1,"a",{}]

In this case the -s option is required:

$ ydump -s stream.json
[
  1234,
  "abcde",
  {
    "start_date": { "year": 1970, "month": 1, "day": 1 },
    "end_date": { "year": 1980, "month": 1, "day": 1 }
  }
]
[ 1, "a", {} ]

From an OCaml program, pretty-printing can be done with Yojson.Safe.prettify which has the following signature:

val prettify : string -> string

We wrote a tiny program that simply calls the prettify function on some predefined JSON data (file prettify.ml):

let json =
"[1234,\"abcde\",{\"start_date\":{\"year\":1970,\"month\":1,\"day\":1},
\"end_date\":{\"year\":1980,\"month\":1,\"day\":1}}]"

let () = print_endline (Yojson.Safe.prettify json)

We now compile and run prettify.ml:

$ ocamlfind ocamlopt -o prettify prettify.ml -package atdgen -linkpkg
$ ./prettify
[
  1234,
  "abcde",
  {
    "start_date": { "year": 1970, "month": 1, "day": 1 },
    "end_date": { "year": 1980, "month": 1, "day": 1 }
  }
]

Source code for this section

Inspecting biniou data

Biniou is a binary format that can be displayed as text using a generic command called bdump. The only practical difficulty is to recover the original field names and variant names which are stored as 31-bit hashes. Unhashing them is done by consulting a dictionary (list of words) maintained by the user.

Let's first produce a sample data file tree.dat containing the biniou representation of a binary tree. In the same program we will also demonstrate how to render biniou data into text from an OCaml program.

Here is the ATD file defining our tree type (file tree.atd):

type tree = [
  | Empty
  | Node of (tree * int * tree)
]

This is our OCaml program (file tree.ml):

open Printf

(* sample value *)
let tree : Tree_t.tree =
  `Node (
    `Node (`Empty, 1, `Empty),
    2,
    `Node (
      `Node (`Empty, 3, `Empty),
      4,
      `Node (`Empty, 5, `Empty)
    )
  )

let () =
  (* write sample value to file *)
  let fname = "tree.dat" in
  Atdgen_runtime.Util.Biniou.to_file Tree_b.write_tree fname tree;

  (* write sample value to string *)
  let s = Tree_b.string_of_tree tree in
  printf "raw value (saved as %s):\n%S\n" fname s;
  printf "length: %i\n" (String.length s);

  printf "pretty-printed value (without dictionary):\n";
  print_endline (Bi_io.view s);

  printf "pretty-printed value (with dictionary):\n";
  let unhash = Bi_io.make_unhash ["Empty"; "Node"; "foo"; "bar" ] in
  print_endline (Bi_io.view ~unhash s)

Compilation:

$ atdgen -t tree.atd
$ atdgen -b tree.atd
$ ocamlfind ocamlopt -o tree \
    tree_t.mli tree_t.ml tree_b.mli tree_b.ml tree.ml \
    -package atdgen -linkpkg

Running the program:

$ ./tree
raw value (saved as tree.dat):
"\023\179\2276\"\020\003\023\179\2276\"\020\003\023\003\007\170m\017\002\023\003\007\170m\017\004\023\179\2276\"\020\003\023\179\2276\"\020\003\023\003\007\170m\017\006\023\003\007\170m\017\b\023\179\2276\"\020\003\023\003\007\170m\017\n\023\003\007\170m"
length: 75
pretty-printed value (without dictionary):
<#33e33622:
  (<#33e33622: (<#0307aa6d>, 1, <#0307aa6d>)>,
    2,
    <#33e33622:
      (<#33e33622: (<#0307aa6d>, 3, <#0307aa6d>)>,
        4,
        <#33e33622: (<#0307aa6d>, 5, <#0307aa6d>)>)>)>
pretty-printed value (with dictionary):
<"Node":
  (<"Node": (<"Empty">, 1, <"Empty">)>,
    2,
    <"Node":
      (<"Node": (<"Empty">, 3, <"Empty">)>,
        4,
        <"Node": (<"Empty">, 5, <"Empty">)>)>)>

Now let's see how to pretty-print any biniou data from the command line. Our sample data are now in file tree.dat:

$ ls -l tree.dat
-rw-r--r-- 1 martin martin 75 Apr 17 01:46 tree.dat

We use the command bdump to render our sample biniou data as text:

$ bdump tree.dat
<#33e33622:
  (<#33e33622: (<#0307aa6d>, 1, <#0307aa6d>)>,
    2,
    <#33e33622:
      (<#33e33622: (<#0307aa6d>, 3, <#0307aa6d>)>,
        4,
        <#33e33622: (<#0307aa6d>, 5, <#0307aa6d>)>)>)>

We got hashes for the variant names Empty and Node. Let's add them to the dictionary:

$ bdump -w Empty,Node tree.dat
<"Node":
  (<"Node": (<"Empty">, 1, <"Empty">)>,
    2,
    <"Node":
      (<"Node": (<"Empty">, 3, <"Empty">)>,
        4,
        <"Node": (<"Empty">, 5, <"Empty">)>)>)>

bdump remembers the dictionary so we don't have to pass the -w option anymore (for this user on this machine). The following now works:

$ bdump tree.dat
<"Node":
  (<"Node": (<"Empty">, 1, <"Empty">)>,
    2,
    <"Node":
      (<"Node": (<"Empty">, 3, <"Empty">)>,
        4,
        <"Node": (<"Empty">, 5, <"Empty">)>)>)>

Source code for this section

Optional fields and default values

Although OCaml records do not support optional fields, both the JSON and biniou formats make it possible to omit certain fields on a per-record basis.

For example the JSON record { "x": 0, "y": 0 } can be more compactly written as {} if the reader knows the default values for the missing fields x and y. Here is the corresponding type definition:

type vector_v1 = { ~x: int; ~y: int }

~x means that field x supports a default value. Since we do not specify the default value ourselves, the built-in default is used, which is 0.

If we want the default to be something else than 0, we just have to specify it as follows:

type vector_v2 = {
  ~x <ocaml default="1">: int; (* default x is 1 *)
  ~y: int;                     (* default y is 0 *)
}

It is also possible to specify optional fields without a default value. For example, let's add an optional z field:

type vector_v3 = {
  ~x: int;
  ~y: int;
  ?z: int option;
}

The following two examples are valid JSON representations of data of type vector_v3:

{ "x": 2, "y": 2, "z": 3 }  // OCaml: { x = 2; y = 2; z = Some 3 }

{ "x": 2, "y": 2 }          // OCaml: { x = 2; y = 2; z = None }

By default, JSON fields whose value is null are treated as missing fields. The following two JSON objects are therefore equivalent:

{ "x": 2, "y": 2, "z": null }
{ "x": 2, "y": 2 }

Note also the difference between ?z: int option and ~z: int option:

type vector_v4 = {
  ~x: int;
  ~y: int;
  ~z: int option;  (* no unwrapping of the JSON field value! *)
}

Here are valid values of type vector_v4, showing that it is usually not what is intended:

{ "x": 2, "y": 2, "z": [ "Some", 3 ] }
{ "x": 2, "y": 2, "z": "None" }
{ "x": 2, "y": 2 }

Smooth protocol upgrades

Problem: you have a production system that uses a specific JSON or biniou format. It may be data files or a client-server pair. You now want to add a field to a record type or to add a case to a variant type.

Both JSON and biniou allow extra record fields. If the consumer does not know how to deal with the extra field, the default behavior is to happily ignore it.

Adding or removing an optional record field

type t = {
  x: int;
  y: int;
}

Same .atd source file, edited:

type t = {
  x: int;
  y: int;
  ~z: int; (* new field *)
}
  • Upgrade producers and consumers in any order
  • Converting old data is not required nor useful

Adding a required record field

type t = {
  x: int;
  y: int;
}

Same .atd source file, edited:

type t = {
  x: int;
  y: int;
  z: int; (* new field *)
}
  • Upgrade all producers before the consumers
  • Converting old data requires special-purpose hand-written code

Removing a required record field

  • Upgrade all consumers before the producers
  • Converting old data is not required but may save some storage space (just read and re^write each record using the new type)

Adding a variant case

type t = [ A | B ]

Same .atd source file, edited:

type t = [ A | B | C ]
  • Upgrade all consumers before the producers
  • Converting old data is not required and would have no effect

Removing a variant case

  • Upgrade all producers before the consumers
  • Converting old data requires special^purpose hand^written code

Avoiding future problems

  • In doubt, use records rather than tuples because it makes it possible to add or remove any field or to reorder them.
  • Do not hesitate to create variant types with only one case or records with only one field if you think they might be extended later.

Data validation

Atdgen can be used to produce data validators for all types defined in an ATD file, based on user-given validators specified only for certain types. A simple example is:

type t = string <ocaml valid="fun s -> String.length s >= 8"> option

As we can see from this example, the validation function is specified using the annotation <ocaml valid="p">, where p is a predicate p : t -> bool, returning true when the value of type t is valid and false otherwise.

Calling atdgen -v on a file containing this specification will produce a validation function equivalent to the following implementation:

let validate_t path x =
  match x with
  | None -> None
  | Some x ->
      let msg = "Failed check by fun s -> String.length s >= 8" in
      if (fun s -> String.length s >= 8) x
      then None
      else Some {error_path = path; error_msg = msg}

Let's consider this particular example as an illustration of the general shape of generated validation functions.

The function takes two arguments: the first, path, is a list indicating where the second, x, was encountered. As specified by our example .atd code above, x has type t option.

The body of the validation function does two things:

1. it checks the value of x against the validation function specified in our .atd file, namely, checking whether there is Some s, and verifying that s is at least 8 characters long if so 2. in the event that the validation check fails, it constructs an appropriate error record.

In general, generated validation functions for a type t have a type equivalent to validate_t : path -> t -> error option, where the path gives the current location in a data structure and the error is a record of the location of, and reason for, validation failure.

A return value of None indicates successful validation, while Some {error_path; error_msg} tells us where and why validation failed.

Let's now consider a more realistic example with complex validators defined in a separate .ml file. We will define a data structure representing a section of a resume recording work experience. We will also define validation functions that can enforce certain properties to protect against errors and junk data.

In the course of this example, we will manually create the following 3 source files:

  • resume.atd: contains the type definitions with annotations
  • resume_util.ml: contains our handwritten validators
  • resume.ml: is our main program that creates data and checks it using our generated validation functions.

After generating additional code with atdgen, we will end up with the following OCaml modules:

  • Resume_t: generated into resume_t.ml by atdgen -t resume.atd, this provides our OCaml type definitions
  • Resume_util: written manually in resume_util.ml, this depends on Resume_t and provides validators we will use in resume.atd
  • Resume_v: generated into resume_v.ml by atdgen -v resume.atd, this depends on Resume_util and Resume_t and provides a validation function for each type
  • Resume_j: generated into resume_j.ml by atdgen -j resume.atd, this provides functions to serialize and deserialize data in and out of JSON.
  • Resume: written manually in resume.ml, this depends on Resume_v, and Resume_t, and makes use of the generated types and validation functions.

To begin, we specify type definitions for a data structure representing a resume in resume.atd:

type text = string <ocaml valid="Resume_util.validate_some_text">

type date = {
  year : int;
  month : int;
  day : int;
} <ocaml valid="Resume_util.validate_date">

type job = {
  company : text;
  title : text;
  start_date : date;
  ?end_date : date option;
} <ocaml valid="Resume_util.validate_job">

type work_experience = job list

We can now call atdgen -t resume.atd to generate our Resume_t module in resume_t.ml, providing our data types. Using these data types, we'll define the following handwritten validators in resume_util.ml (note that we've already referred to these validators in resume.atd):

open Resume_t

let ascii_printable c =
  let n = Char.code c in
  n >= 32 && n <= 127

(*
  Check that string is not empty and contains only ASCII printable
  characters (for the sake of the example; we use UTF-8 these days)
*)
let validate_some_text s =
  s <> "" &&
    try
      String.iter (fun c -> if not (ascii_printable c) then raise Exit) s;
      true
    with Exit ->
      false

(*
  Check that the combination of year, month and day exists in the
  Gregorian calendar.
*)
let validate_date x =
  let y = x.year in
  let m = x.month in
  let d = x.day in
  m >= 1 && m <= 12 && d >= 1 &&
  (let dmax =
    match m with
        2 ->
          if y mod 4 = 0 && not (y mod 100 = 0) || y mod 400 = 0 then 29
          else 28
      | 1 | 3 | 5 | 7 | 8 | 10 | 12 -> 31
      | _ -> 30
  in
  d <= dmax)

(* Compare dates chronologically *)
let compare_date a b =
  let c = compare a.year b.year in
  if c <> 0 then c
  else
    let c = compare a.month b.month in
    if c <> 0 then c
    else compare a.day b.day

(* Check that the end_date, when defined, is not earlier than the start_date *)
let validate_job x =
  match x.end_date with
      None -> true
    | Some end_date ->
        compare_date x.start_date end_date <= 0

After we call atdgen -v resume.atd, the module Resume_v will be generated in resume_v.ml, providing the function validate_work_experience . We can then use this function, along with the generated Resume_j in the following program written in resume.ml:

let check_experience x =
  let is_valid = match Resume_v.validate_work_experience [] x with
    | None -> true
    | _ -> false
  in
  Printf.printf "%s:\n%s\n"
    (if is_valid then "VALID" else "INVALID")
    (Yojson.Safe.prettify (Resume_j.string_of_work_experience x))

let () =
  (* one valid date *)
  let valid = { Resume_t.year = 2000; month = 2; day = 29 } in
  (* one invalid date *)
  let invalid = { Resume_t.year = 1900; month = 0; day = 0 } in
  (* two more valid dates, created with Resume_v.create_date *)
  let date1 = { Resume_t.year = 2005; month = 8; day = 1 } in
  let date2 = { Resume_t.year = 2006; month = 3; day = 22 } in

  let job = {
    Resume_t.company = "Acme Corp.";
    title = "Tester";
    start_date = date1;
    end_date = Some date2;
  }
  in
  let valid_job = { job with Resume_t.start_date = valid } in
  let invalid_job = { job with Resume_t.end_date = Some invalid } in
  let valid_experience = [ job; valid_job ] in
  let invalid_experience = [ job; invalid_job ] in
  check_experience valid_experience;
  check_experience invalid_experience

Output:

VALID:
[
  {
    "company": "Acme Corp.",
    "title": "Tester",
    "start_date": { "year": 2005, "month": 8, "day": 1 },
    "end_date": { "year": 2006, "month": 3, "day": 22 }
  },
  {
    "company": "Acme Corp.",
    "title": "Tester",
    "start_date": { "year": 2000, "month": 2, "day": 29 },
    "end_date": { "year": 2006, "month": 3, "day": 22 }
  }
]
INVALID:
[
  {
    "company": "Acme Corp.",
    "title": "Tester",
    "start_date": { "year": 2005, "month": 8, "day": 1 },
    "end_date": { "year": 2006, "month": 3, "day": 22 }
  },
  {
    "company": "Acme Corp.",
    "title": "Tester",
    "start_date": { "year": 2005, "month": 8, "day": 1 },
    "end_date": { "year": 1900, "month": 0, "day": 0 }
  }

Source code for this section

Modularity: referring to type definitions from another ATD file

It is possible to define types that depend on types defined in other .atd files. The example below is self-explanatory.

part1.atd:

type t = { x : int; y : int }

part2.atd:

type t1 <ocaml from="Part1" t="t"> = abstract
    (*
      Imports type t defined in file part1.atd.
      The local name is t1. Because the local name (t1) is different from the
      original name (t), we must specify the original name using t=.
    *)

type t2 = t1 list

part3.atd:

type t2 <ocaml from="Part2"> = abstract

type t3 = {
  name : string;
  ?data : t2 option;
}

main.ml:

let v = {
  Part3_t.name = "foo";
  data = Some [
    { Part1_t.x = 1; y = 2 };
    { Part1_t.x = 3; y = 4 };
  ]
}

let () =
  Atdgen_runtime.Util.Json.to_channel Part3_j.write_t3 stdout v;
  print_newline ()

Output:

{"name":"foo","data":[{"x":1,"y":2},{"x":3,"y":4}]}

Source code for this section

Managing JSON configuration files

JSON makes a good format for configuration files because it is human-readable, easy to modify programmatically and widespread. Here is an example of how to use atdgen to manage config files.

  • Specifying defaults is done in the .atd file. See section [Optional fields and default values] for details on how to do that.
  • Auto-generating a template config file with default values: a sample value in the OCaml world needs to be created but only fields without default need to be specified.
  • Describing the format is achieved by embedding the .atd type definitions in the OCaml program and printing it out on request.
  • Loading a config file and reporting illegal fields is achieved using the JSON deserializers produced by atdgen -j. Option -j-strict-fields ensures the misspelled field names are not ignored but reported as errors.
  • Reindenting a config file is achieved by the pretty-printing function Yojson.Safe.prettify that takes a JSON string and returns an equivalent JSON string.
  • Showing implicit (default) settings is achieved by passing the -j-defaults option to atdgen. The OCaml config data is then serialized into JSON containing all fields, including those whose value is the default.

The example uses the following type definitions:

type config = {
  title : string;
  ?description : string option;
  ~timeout <ocaml default="10"> : int;
  ~credentials : param list
    <ocaml valid="fun l ->
                    l <> [] || failwith \"missing credentials\"">;
}

type param = {
  name : string
    <ocaml valid="fun s -> s <> \"\"">;
  key : string
    <ocaml valid="fun s -> String.length s = 16">;
}

Our program will perform the following actions:

$ ./config -template
{
  "title": "",
  "timeout": 10,
  "credentials": [ { "name": "foo", "key": "0123456789abcdef" } ]
}

$ ./config -format
type config = {
  title : string;
  ?description : string option;
  ~timeout <ocaml default="10"> : int;
  ~credentials : param list
    <ocaml valid="fun l ->
                    l <> [] || failwith \"missing credentials\"">;
}

type param = {
  name : string
    <ocaml valid="fun s -> s <> \"\"">;
  key : string
    <ocaml valid="fun s -> String.length s = 16">;
}

$ cat sample-config.json
{
  "title": "Example",
  "credentials": [
    {
      "name": "joeuser",
      "key": "db7c0877bdef3016"
    },
    {
      "name": "tester",
      "key": "09871ff387ac2b10"
    }
  ]
}

$ ./config -validate sample-config.json
{
  "title": "Example",
  "timeout": 10,
  "credentials": [
    { "name": "joeuser", "key": "db7c0877bdef3016" },
    { "name": "tester", "key": "09871ff387ac2b10" }
  ]
}

This is our demo.sh script that builds and runs our example program called config:

#! /bin/sh -e

set -x

# Embed the contents of the .atd file into our OCaml program
echo 'let contents = "\' > config_atd.ml
sed -e 's/\([\\"]\)/\\\1/g' config.atd >> config_atd.ml
echo '"' >> config_atd.ml

# Derive OCaml type definitions from .atd file
atdgen -t config.atd

# Derive JSON-related functions from .atd file
atdgen -j -j-defaults -j-strict-fields config.atd

# Derive validator from .atd file
atdgen -v config.atd

# Compile the OCaml program
ocamlfind ocamlopt -o config \
  config_t.mli config_t.ml config_j.mli config_j.ml config_v.mli config_v.ml \
  config_atd.ml config.ml -package atdgen -linkpkg

# Output a sample config
./config -template

# Print the original type definitions
./config -format

# Fail to validate an invalid config file
./config -validate bad-config1.json || :

# Fail to validate another invalid config file (using custom validators)
./config -validate bad-config3.json || :

# Validate, inject missing defaults and pretty-print
./config -validate sample-config.json

This is the hand-written OCaml program. It can be used as a start
point for a real-world program using a JSON config file:

open Printf

let param_template =
  (* Sample item used to populate the template config file *)
  {
    Config_v.name = "foo";
    key = "0123456789abcdef"
  }

let config_template =
  (*
    Records can be conveniently created using functions generated by
    "atdgen -v".
    Here we use Config_v.create_config to create a record of type
    Config_t.config. The big advantage over creating the record
    directly using the record notation {...} is that we don't have to
    specify default values (such as timeout in this example).
  *)
  Config_v.create_config ~title:"" ~credentials: [param_template] ()

let make_json_template () =
  (* Thanks to the -j-defaults flag passed to atdgen, even default
    fields will be printed out *)
  let compact_json = Config_j.string_of_config config_template in
  Yojson.Safe.prettify compact_json

let print_template () =
  print_endline (make_json_template ())

let print_format () =
  print_string Config_atd.contents

let validate fname =
  let x =
    try
      (* Read config data structure from JSON file *)
      let x = Atdgen_runtime.Util.Json.from_file Config_j.read_config fname in
      (* Call the validators specified by <ocaml valid=...> *)
      if not (Config_v.validate_config x) then
        failwith "Some fields are invalid"
      else
        x
    with e ->
      (* Print decent error message and exit *)
      let msg =
        match e with
            Failure s
          | Yojson.Json_error s -> s
          | e -> Printexc.to_string e
      in
      eprintf "Error: %s\n%!" msg;
      exit 1
  in
  (* Convert config to compact JSON and pretty-print it.
    ~std:true means that the output will not use extended syntax for
    variants and tuples but only standard JSON. *)
  let json = Yojson.Safe.prettify ~std:true (Config_j.string_of_config x) in
  print_endline json

type action = Template | Format | Validate of string

let main () =
  let action = ref Template in
  let options = [
    "-template", Arg.Unit (fun () -> action := Template),
    "
          prints a sample configuration file";

    "-format", Arg.Unit (fun () -> action := Format),
    "
          prints the format specification of the config files (atd format)";

    "-validate", Arg.String (fun s -> action := Validate s),
    "<CONFIG FILE>
          reads a config file, validates it, adds default values
          and prints the config nicely to stdout";
  ]
  in
  let usage_msg = sprintf "\
Usage: %s [-template|-format|-validate ...]
Demonstration of how to manage JSON configuration files with atdgen.
"
    Sys.argv.(0)
  in
  let anon_fun s = eprintf "Invalid command parameter %S\n%!" s; exit 1 in
  Arg.parse options anon_fun usage_msg;

  match !action with
      Template -> print_template ()
    | Format -> print_format ()
    | Validate s -> validate s

let () = main ()

The full source code for this section with examples can be inspected and downloaded here.

Integration with ocamldoc

Ocamldoc is a tool that comes with the core OCaml distribution. It uses comments within (** and *) to produce hyperlinked documentation (HTML) of module signatures.

Atdgen can produce .mli files with comments in the syntax supported by ocamldoc but regular ATD comments within (* and *) are always discarded by atdgen. Instead, <doc text="..."> must be used and placed after the element they describe. The contents of the text field must be UTF8-encoded.

type point = {
  x : float;
  y : float;
  ~z
    <doc text="Optional depth, its default value is {{0.0}}.">
    : float;
}
  <doc text="Point with optional 3rd dimension.

OCaml example:
{{{
let p =
  { x = 0.5; y = 1.0; z = 0. }
}}}
">

is converted into the following .mli file with ocamldoc-compatible comments:

(**
  Point with optional 3rd dimension.

  OCaml example:

{v
let p =
  \{ x = 0.5; y = 1.0; z = 0. \}
v}
*)
type point = {
  x: float;
  y: float;
  z: float (** Optional depth, its default value is [0.0]. *)
}

The only two forms of markup supported by <doc text="..."> are {{ ... }} for inline code and {{{ ... }}} for a block of preformatted code.

Integration with build systems

OMake

We provide an Atdgen plugin for OMake. It simplifies the compilation rules to a minimum.

The plugin consists of a self-documented file to copy into a project's root. The following is a sample OMakefile for a project using JSON and five source files (foo.atd, foo.ml, bar.atd, bar.ml and main.ml):

# require file Atdgen.om
include Atdgen

# OCaml modules we want to build
OCAMLFILES = foo_t foo_j foo bar_t bar_j bar main

Atdgen(foo bar, -j-std)
OCamlProgram(foobar, $(OCAMLFILES))

.DEFAULT: foobar.opt

.PHONY: clean
clean:
  rm -f *.cm[ioxa] *.cmx[as] *.[oa] *.opt *.run *~
  rm -f $(ATDGEN_OUTFILES)

Running omake builds the native code executable foobar.opt.

omake clean removes all the products of compilation including the .mli and .ml produced by atdgen.

GNU Make

We provide Atdgen.mk, a generic makefile that defines the dependencies and rules for generating OCaml .mli and .ml files from .atd files containing type definitions. The Atdgen.mk file contains its own documentation.

Here is a sample Makefile that takes advantage of OCamlMakefile:

.PHONY: default
default: opt

ATDGEN_SOURCES = foo.atd bar.atd
ATDGEN_FLAGS = -j-std
include Atdgen.mk

SOURCES = \
  foo_t.mli foo_t.ml foo_j.mli foo_j.ml \
  bar_t.mli bar_t.ml bar_j.mli bar_j.ml \
  hello.ml
RESULT = hello
PACKS = atdgen
# "include OCamlMakefile" must come after defs for SOURCES, RESULT, PACKS, etc.
include OCamlMakefile

.PHONY: sources opt all
sources: $(SOURCES)
opt: sources
        $(MAKE) native-code
all: sources
        $(MAKE) byte-code

make alone builds a native code executable from source files foo.atd, bar.atd and hello.ml. make clean removes generated files. make all builds a bytecode executable.

In addition to native-code, byte-code and clean, OCamlMakefile provides a number of other targets and options which are documented in OCamlMakefile's README.

Ocamlbuild

There is an atdgen plugin for ocamlbuild.

Dune (formerly jbuilder)

Dune requires atdgen build rules to be specified for each .atd file.

Manual Approach

Given an example.atd, you can manually specify rules like:

(rule
 (targets example_j.ml
          example_j.mli)
 (deps    example.atd)
 (action  (run atdgen -j -j-std %{deps})))

(rule
 (targets example_t.ml
          example_t.mli)
 (deps    example.atd)
 (action  (run atdgen -t %{deps})))

You can refer to example_t.ml and example_j.ml as usual (by default, they will be automatically linked into the library being built in the same directory).

Note that any options atdgen supports can be included in the run atdgen section (-open, -deriving-conv, etc.).

You will need to write rules for each .atd file individually or ...

Automated Approach with Code Generation

While Dune does not yet support wildcard rules, you can work around this limitation using Dune's rule generation feature to automatically generate rules for all .atd files in a directory.

Step 1: Create a generator directory (e.g., gen/) with a dune file:

(executable
 (name gen))

Step 2: Create the generator OCaml program (gen/gen.ml):

let generate_atdgen_rules base =
  (* Generate -j rules *)
  Printf.printf
    {|(rule
 (targets %s_j.ml %s_j.mli)
 (deps %s.atd)
 (action
  (run atdgen -j -j-std %%{deps})))

|}
    base base base;
  (* Generate -t rules *)
  Printf.printf
    {|(rule
 (targets %s_t.ml %s_t.mli)
 (deps %s.atd)
 (action
  (run atdgen -t %%{deps})))

|}
    base base base;
  (* Generate -v rules *)
  Printf.printf
    {|(rule
 (targets %s_v.ml %s_v.mli)
 (deps %s.atd)
 (action
  (run atdgen -v %%{deps})))

|}
    base base base
;;

let collect_atd_files () =
  Sys.readdir "."
  |> Array.to_list
  |> List.sort String.compare
  |> List.filter_map (fun filename ->
    match Filename.chop_suffix_opt ~suffix:".atd" filename with
    | Some base -> Some base
    | None -> None)
;;

let () =
  let atd_files = collect_atd_files () in
  List.iter generate_atdgen_rules atd_files
;;

Step 3: Create an empty dune.inc file:

; Generated rules will be included here

Step 4: Update your main dune file to include and generate dune.inc:

; Include auto-generated rules for .atd files
(include dune.inc)

; Generate the dune.inc file from all .atd files
(rule
 (alias genatd)
 (mode promote)
 (deps (glob_files *.atd))
 (action
  (with-stdout-to
   dune.inc
   (run gen/gen.exe))))

Step 5: Generate the initial dune.inc file:

$ dune build @genatd

This only needs to be run once when you first create the empty dune.inc file.

Step 6: Build your project:

$ dune build

From this point forward, Dune will automatically regenerate dune.inc whenever you add, remove, or modify .atd files. The generated rules will be promoted to your source tree automatically.

This approach is especially useful when working with many .atd files, as it eliminates the need to manually maintain build rules for each file. You can also customize the generator to add project-specific atdgen flags (like -deriving-conv "eq,show" or -j-strict-fields) or generate additional stanzas (like library definitions with auto-discovered modules).

Dealing with untypable JSON

Sometimes we have to deal with JSON data that cannot be described using type definitions. In such case, we can represent the data as its JSON abstract syntax tree (AST), which lets the user inspect it at runtime.

Let's consider a list of JSON objects for which we don't know the type definitions, but somehow some other system knows how to deal with such data. Here is such data:

[
  {
    "label": "flower",
    "value": {
      "petals": [12, 45, 83.5555],
      "water": "a340bcf02e"
    }
  },
  {
    "label": "flower",
    "value": {
      "petals": "undefined",
      "fold": null,
      "water": 0
    }
  },
  { "labels": ["fork", "scissors"],
    "value": [ 8, 8 ]
  }
]

Hopefully this means something for someone. We are going to assume that each object has a value field of an unknown type, and may have a field label or a field labels of type string:

(* File untypable.atd *)

type obj_list = obj list

type obj = {
  ?label: string option;
  ?labels: string list option;
  value: abstract  (* requires ATD >= 2.6.0 *)
}

Until ATD 2.5, abstract could not be used as freely and would not stand for raw JSON by default. One had to write a dedicated type definition as shown below:

(* File untypable.atd *)

(* deprecated since ATD 2.6 *)
type json <ocaml module="Yojson.Safe"> = abstract
  (* uses type Yojson.Safe.t,
     with the functions Yojson.Safe.write_json
     and Yojson.Safe.read_json *)

type obj_list = obj list

type obj = {
  ?label: string option;
  ?labels: string list option;
  value: json
}

It is possible to give a different name than json to the type of the JSON AST, but then the name of the type used in the original module must be provided in the annotation, i.e.:

(* deprecated since ATD 2.6 *)
type raw_json <ocaml module="Yojson.Safe" t="json"> = abstract
  (* uses type Yojson.Safe.t,
     with the functions Yojson.Safe.write_json
     and Yojson.Safe.read_json *)

type obj_list = obj list

type obj = {
  ?label: string option;
  ?labels: string list option;
  value: raw_json
}

Compile either example with:

$ atdgen -t untypable.atd
$ atdgen -j -j-std untypable.atd
$ ocamlfind ocamlc -a -o untypable.cma -package atdgen \
    untypable_t.mli untypable_t.ml untypable_j.mli untypable_j.ml

Test the example with your favorite OCaml toplevel (ocaml or utop):

# #use "topfind";;
# #require "atdgen";;
# #load "untypable.cma";;
# Atdgen_runtime.Util.Json.from_channel Untypable_j.read_obj_list stdin;;
[
  {
    "label": "flower",
    "value": {
      "petals": [12, 45, 83.5555],
      "water": "a340bcf02e"
    }
  },
  {
    "label": "flower",
    "value": {
      "petals": "undefined",
      "fold": null,
      "water": 0
    }
  },
  { "labels": ["fork", "scissors"],
    "value": [ 8, 8 ]
  }
]
- : Untypable_t.obj_list =
[{Untypable_t.label = Some "flower"; labels = None;
  value =
  `Assoc
    [("petals", `List [`Int 12; `Int 45; `Float 83.5555]);
      ("water", `String "a340bcf02e")]};
{Untypable_t.label = Some "flower"; labels = None;
  value =
  `Assoc [("petals", `String "undefined");
          ("fold", `Null);
          ("water", `Int 0)]};
{Untypable_t.label = None; labels = Some ["fork"; "scissors"];
  value = `List [`Int 8; `Int 8]}]