/config

Configuration library for JVM languages

Primary LanguageJava

Configuration library for JVM languages.

Overview

  • implemented in plain Java with no dependencies
  • supports files in three formats: Java properties, JSON, and a human-friendly JSON superset
  • merges multiple files across all formats
  • can load from files, URLs, or classpath
  • good support for "nesting" (treat any subtree of the config the same as the whole config)
  • users can override the config with Java system properties, java -Dmyapp.foo.bar=10
  • supports configuring an app, with its framework and libraries, all from a single file such as application.conf
  • parses duration and size settings, "512k" or "10 seconds"
  • converts types, so if you ask for a boolean and the value is the string "yes", or you ask for a float and the value is an int, it will figure it out.
  • JSON superset features:
    • comments
    • includes
    • substitutions ("foo" : ${bar}, "foo" : Hello ${who})
    • properties-like notation (a.b=c)
    • less noisy, more lenient syntax
    • substitute environment variables (logdir=${HOME}/logs)
  • API based on immutable Config instances, for thread safety and easy reasoning about config transformations
  • extensive test coverage

This library limits itself to config files. If you want to load config from a database or something, you would need to write some custom code. The library has nice support for merging configurations so if you build one from a custom source it's easy to merge it in.

Table of Contents generated with DocToc

Essential Information

License

The license is Apache 2.0, see LICENSE-2.0.txt.

Binary Releases

You can find published releases (compiled for Java 6 and above) on Maven Central.

<dependency>
    <groupId>com.typesafe</groupId>
    <artifactId>config</artifactId>
    <version>1.2.1</version>
</dependency>

Obsolete releases are here, but you probably don't want these:

Release Notes

Please see NEWS.md in this directory, https://github.com/typesafehub/config/blob/master/NEWS.md

API docs

Bugs and Patches

Report bugs to the GitHub issue tracker. Send patches as pull requests on GitHub.

Before we can accept pull requests, you will need to agree to the Typesafe Contributor License Agreement online, using your GitHub account - it takes 30 seconds. You can do this at http://www.typesafe.com/contribute/cla

Build

The build uses sbt and the tests are written in Scala; however, the library itself is plain Java and the published jar has no Scala dependency.

Using the Library

API Example

import com.typesafe.config.ConfigFactory

Config conf = ConfigFactory.load();
int bar1 = conf.getInt("foo.bar");
Config foo = conf.getConfig("foo");
int bar2 = foo.getInt("bar");

Longer Examples

See the examples in the examples/ directory.

You can run these from the sbt console with the commands project config-simple-app-java and then run.

In brief, as shown in the examples:

  • libraries should use a Config instance provided by the app, if any, and use ConfigFactory.load() if no special Config is provided. Libraries should put their defaults in a reference.conf on the classpath.
  • apps can create a Config however they want (ConfigFactory.load() is easiest and least-surprising), then provide it to their libraries. A Config can be created with the parser methods in ConfigFactory or built up from any file format or data source you like with the methods in ConfigValueFactory.

Immutability

Objects are immutable, so methods on Config which transform the configuration return a new Config. Other types such as ConfigParseOptions, ConfigResolveOptions, ConfigObject, etc. are also immutable. See the API docs for details of course.

Schemas and Validation

There isn't a schema language or anything like that. However, two suggested tools are:

  • use the checkValid() method
  • access your config through a Settings class with a field for each setting, and instantiate it on startup (immediately throwing an exception if any settings are missing)

In Scala, a Settings class might look like:

class Settings(config: Config) {

    // validate vs. reference.conf
    config.checkValid(ConfigFactory.defaultReference(), "simple-lib")

    // non-lazy fields, we want all exceptions at construct time
    val foo = config.getString("simple-lib.foo")
    val bar = config.getInt("simple-lib.bar")
}

See the examples/ directory for a full compilable program using this pattern.

Standard behavior

The convenience method ConfigFactory.load() loads the following (first-listed are higher priority):

  • system properties
  • application.conf (all resources on classpath with this name)
  • application.json (all resources on classpath with this name)
  • application.properties (all resources on classpath with this name)
  • reference.conf (all resources on classpath with this name)

The idea is that libraries and frameworks should ship with a reference.conf in their jar. Applications should provide an application.conf, or if they want to create multiple configurations in a single JVM, they could use ConfigFactory.load("myapp") to load their own myapp.conf. (Applications can provide a reference.conf also if they want, but you may not find it necessary to separate it from application.conf.)

Libraries and frameworks should default to ConfigFactory.load() if the application does not provide a custom Config object. This way, libraries will see configuration from application.conf and users can configure the whole app, with its libraries, in a single application.conf file.

Libraries and frameworks should also allow the application to provide a custom Config object to be used instead of the default, in case the application needs multiple configurations in one JVM or wants to load extra config files from somewhere. The library examples in examples/ show how to accept a custom config while defaulting to ConfigFactory.load().

For applications using application.{conf,json,properties}, system properties can be used to force a different config source:

  • config.resource specifies a resource name - not a basename, i.e. application.conf not application
  • config.file specifies a filesystem path, again it should include the extension, not be a basename
  • config.url specifies a URL

These system properties specify a replacement for application.{conf,json,properties}, not an addition. They only affect apps using the default ConfigFactory.load() configuration. In the replacement config file, you can use include "application" to include the original default config file; after the include statement you could go on to override certain settings.

Merging config trees

Any two Config objects can be merged with an associative operation called withFallback, like merged = firstConfig.withFallback(secondConfig).

The withFallback operation is used inside the library to merge duplicate keys in the same file and to merge multiple files. ConfigFactory.load() uses it to stack system properties over application.conf over reference.conf.

You can also use withFallback to merge in some hardcoded values, or to "lift" a subtree up to the root of the configuration; say you have something like:

foo=42
dev.foo=57
prod.foo=10

Then you could code something like:

Config devConfig = originalConfig
                     .getConfig("dev")
                     .withFallback(originalConfig)

There are lots of ways to use withFallback.

How to handle defaults

Many other configuration APIs allow you to provide a default to the getter methods, like this:

boolean getBoolean(String path, boolean fallback)

Here, if the path has no setting, the fallback would be returned. An API could also return null for unset values, so you would check for null:

// returns null on unset, check for null and fall back
Boolean getBoolean(String path)

The methods on the Config interface do NOT do this, for two major reasons:

  1. If you use a config setting in two places, the default fallback value gets cut-and-pasted and typically out of sync. This can result in Very Evil Bugs.
  2. If the getter returns null (or None, in Scala) then every time you get a setting you have to write handling code for null/None and that code will almost always just throw an exception. Perhaps more commonly, people forget to check for null at all, so missing settings result in NullPointerException.

For most situations, failure to have a setting is simply a bug to fix (in either code or the deployment environment). Therefore, if a setting is unset, by default the getters on the Config interface throw an exception.

If you want to allow a setting to be missing from application.conf in a particular case, then here are some options:

  1. Set it in a reference.conf included in your library or application jar, so there's a default value.
  2. Catch and handle ConfigException.Missing.
  3. Use the Config.hasPath() method to check in advance whether the path exists (rather than checking for null/None after as you might in other APIs).
  4. In your initialization code, generate a Config with your defaults in it (using something like ConfigFactory.parseMap()) then fold that default config into your loaded config using withFallback(), and use the combined config in your program. "Inlining" your reference config in the code like this is probably less convenient than using a reference.conf file, but there may be reasons to do it.
  5. Use Config.root() to get the ConfigObject for the Config; ConfigObject implements java.util.Map<String,?> and the get() method on Map returns null for missing keys. See the API docs for more detail on Config vs. ConfigObject.

The recommended path (for most cases, in most apps) is that you require all settings to be present in either reference.conf or application.conf and allow ConfigException.Missing to be thrown if they are not. That's the design intent of the Config API design.

Consider the "Settings class" pattern with checkValid() to verify that you have all settings when you initialize the app. See the Schemas and Validation section of this README for more details on this pattern.

If you do need a setting to be optional: checking hasPath() in advance should be the same amount of code (in Java) as checking for null afterward, without the risk of NullPointerException when you forget. In Scala, you could write an enrichment class like this to use the idiomatic Option syntax:

implicit class RichConfig(val underlying: Config) extends AnyVal {
  def getOptionalBoolean(path: String): Option[Boolean] = try {
     Some(underlying.getBoolean(path))
  } catch {
     case e: ConfigException.Missing =>
         None
  }
}

Since this library is a Java library it doesn't come with that out of the box, of course.

It is understood that sometimes defaults in code make sense. For example, if your configuration lets users invent new sections, you may not have all paths up front and may be unable to set up defaults in reference.conf for dynamic paths. The design intent of Config isn't to prohibit inline defaults, but simply to recognize that it seems to be the 10% case (rather than the 90% case). Even in cases where dynamic defaults are needed, you may find that using withFallback() to build a complete nothing-missing Config in one central place in your code keeps things tidy.

Whatever you do, please remember not to cut-and-paste default values into multiple places in your code. You have been warned! :-)

Understanding Config and ConfigObject

To read and modify configuration, you'll use the Config interface. A Config looks at a JSON-equivalent data structure as a one-level map from paths to values. So if your JSON looks like this:

  "foo" : {
    "bar" : 42
    "baz" : 43
  }

Using the Config interface, you could write conf.getInt("foo.bar"). The foo.bar string is called a path expression (HOCON.md has the syntax details for these expressions). Iterating over this Config, you would get two entries; "foo.bar" : 42 and "foo.baz" : 43. When iterating a Config you will not find nested Config (because everything gets flattened into one level).

When looking at a JSON tree as a Config, null values are treated as if they were missing. Iterating over a Config will skip null values.

You can also look at a Config in the way most JSON APIs would, through the ConfigObject interface. This interface represents an object node in the JSON tree. ConfigObject instances come in multi-level trees, and the keys do not have any syntax (they are just strings, not path expressions). Iterating over the above example as a ConfigObject, you would get one entry "foo" : { "bar" : 42, "baz" : 43 }, where the value at "foo" is another nested ConfigObject.

In ConfigObject, null values are visible (distinct from missing values), just as they are in JSON.

ConfigObject is a subtype of ConfigValue, where the other subtypes are the other JSON types (list, string, number, boolean, null).

Config and ConfigObject are two ways to look at the same internal data structure, and you can convert between them for free using Config.root() and ConfigObject.toConfig().

Using HOCON, the JSON Superset

The JSON superset is called "Human-Optimized Config Object Notation" or HOCON, and files use the suffix .conf. See HOCON.md in this directory for more detail.

After processing a .conf file, the result is always just a JSON tree that you could have written (less conveniently) in JSON.

Features of HOCON

  • Comments, with # or //
  • Allow omitting the {} around a root object
  • Allow = as a synonym for :
  • Allow omitting the = or : before a { so foo { a : 42 }
  • Allow omitting commas as long as there's a newline
  • Allow trailing commas after last element in objects and arrays
  • Allow unquoted strings for keys and values
  • Unquoted keys can use dot-notation for nested objects, foo.bar=42 means foo { bar : 42 }
  • Duplicate keys are allowed; later values override earlier, except for object-valued keys where the two objects are merged recursively
  • include feature merges root object in another file into current object, so foo { include "bar.json" } merges keys in bar.json into the object foo
  • include with no file extension includes any of .conf, .json, .properties
  • you can include files, URLs, or classpath resources; use include url("http://example.com") or file() or classpath() syntax to force the type, or use just include "whatever" to have the library do what you probably mean (Note: url()/file()/classpath() syntax is not supported in Play/Akka 2.0, only in later releases.)
  • substitutions foo : ${a.b} sets key foo to the same value as the b field in the a object
  • substitutions concatenate into unquoted strings, foo : the quick ${colors.fox} jumped
  • substitutions fall back to environment variables if they don't resolve in the config itself, so ${HOME} would work as you expect. Also, most configs have system properties merged in so you could use ${user.home}.
  • substitutions normally cause an error if unresolved, but there is a syntax ${?a.b} to permit them to be missing.
  • += syntax to append elements to arrays, path += "/bin"
  • multi-line strings with triple quotes as in Python or Scala

Examples of HOCON

All of these are valid HOCON.

Start with valid JSON:

{
    "foo" : {
        "bar" : 10,
        "baz" : 12
    }
}

Drop root braces:

"foo" : {
    "bar" : 10,
    "baz" : 12
}

Drop quotes:

foo : {
    bar : 10,
    baz : 12
}

Use = and omit it before {:

foo {
    bar = 10,
    baz = 12
}

Remove commas:

foo {
    bar = 10
    baz = 12
}

Use dotted notation for unquoted keys:

foo.bar=10
foo.baz=12

Put the dotted-notation fields on a single line:

foo.bar=10, foo.baz=12

The syntax is well-defined (including handling of whitespace and escaping). But it handles many reasonable ways you might want to format the file.

Note that while you can write HOCON that looks a lot like a Java properties file (and many properties files will parse as HOCON), the details of escaping, whitespace handling, comments, and so forth are more like JSON. The spec (see HOCON.md in this directory) has some more detailed notes on this topic.

Uses of Substitutions

The ${foo.bar} substitution feature lets you avoid cut-and-paste in some nice ways.

Factor out common values

This is the obvious use,

standard-timeout = 10ms
foo.timeout = ${standard-timeout}
bar.timeout = ${standard-timeout}

Inheritance

If you duplicate a field with an object value, then the objects are merged with last-one-wins. So:

foo = { a : 42, c : 5 }
foo = { b : 43, c : 6 }

means the same as:

foo = { a : 42, b : 43, c : 6 }

You can take advantage of this for "inheritance":

data-center-generic = { cluster-size = 6 }
data-center-east = ${data-center-generic}
data-center-east = { name = "east" }
data-center-west = ${data-center-generic}
data-center-west = { name = "west", cluster-size = 8 }

Using include statements you could split this across multiple files, too.

Optional system or env variable overrides

In default uses of the library, exact-match system properties already override the corresponding config properties. However, you can add your own overrides, or allow environment variables to override, using the ${?foo} substitution syntax.

basedir = "/whatever/whatever"
basedir = ${?FORCED_BASEDIR}

Here, the override field basedir = ${?FORCED_BASEDIR} simply vanishes if there's no value for FORCED_BASEDIR, but if you set an environment variable FORCED_BASEDIR for example, it would be used.

A natural extension of this idea is to support several different environment variable names or system property names, if you aren't sure which one will exist in the target environment.

Object fields and array elements with a ${?foo} substitution value just disappear if the substitution is not found:

// this array could have one or two elements
path = [ "a", ${?OPTIONAL_A} ]

Concatenation

Values on the same line are concatenated (for strings and arrays) or merged (for objects).

This is why unquoted strings work, here the number 42 and the string foo are concatenated into a string 42 foo:

key : 42 foo

When concatenating values into a string, leading and trailing whitespace is stripped but whitespace between values is kept.

Unquoted strings also support substitutions of course:

tasks-url : ${base-url}/tasks

A concatenation can refer to earlier values of the same field:

path : "/bin"
path : ${path}":/usr/bin"

Arrays can be concatenated as well:

path : [ "/bin" ]
path : ${path} [ "/usr/bin" ]

There is a shorthand for appending to arrays:

// equivalent to: path = ${?path} [ "/usr/bin" ]
path += "/usr/bin"

To prepend or insert into an array, there is no shorthand.

When objects are "concatenated," they are merged, so object concatenation is just a shorthand for defining the same object twice. The long way (mentioned earlier) is:

data-center-generic = { cluster-size = 6 }
data-center-east = ${data-center-generic}
data-center-east = { name = "east" }

The concatenation-style shortcut is:

data-center-generic = { cluster-size = 6 }
data-center-east = ${data-center-generic} { name = "east" }

When concatenating objects and arrays, newlines are allowed inside each object or array, but not between them.

Non-newline whitespace is never a field or element separator. So [ 1 2 3 4 ] is an array with one unquoted string element "1 2 3 4". To get an array of four numbers you need either commas or newlines separating the numbers.

See the spec for full details on concatenation.

Note: Play/Akka 2.0 have an earlier version that supports string concatenation, but not object/array concatenation. += does not work in Play/Akka 2.0 either. Post-2.0 versions support these features.

Miscellaneous Notes

Debugging Your Configuration

If you have trouble with your configuration, some useful tips.

  • Set the Java system property -Dconfig.trace=loads to get output on stderr describing each file that is loaded. Note: this feature is not included in the older version in Play/Akka 2.0.
  • Use myConfig.root().render() to get a Config printed out as a string with comments showing where each value came from.

Supports Java 6 and Later

Currently the library is maintained against Java 6. It does not build with Java 5.

Rationale for Supported File Formats

(For the curious.)

The three file formats each have advantages.

  • Java .properties:
    • Java standard, built in to JVM
    • Supported by many tools such as IDEs
  • JSON:
    • easy to generate programmatically
    • well-defined and standard
    • bad for human maintenance, with no way to write comments, and no mechanisms to avoid duplication of similar config sections
  • HOCON/.conf:
    • nice for humans to read, type, and maintain, with more lenient syntax
    • built-in tools to avoid cut-and-paste
    • ways to refer to the system environment, such as system properties and environment variables

The idea would be to use JSON if you're writing a script to spit out config, and use HOCON if you're maintaining config by hand. If you're doing both, then mix the two.

Two alternatives to HOCON syntax could be:

  • YAML is also a JSON superset and has a mechanism for adding custom types, so the include statements in HOCON could become a custom type tag like !include, and substitutions in HOCON could become a custom tag such as !subst, for example. The result is somewhat clunky to write, but would have the same in-memory representation as the HOCON approach.
  • Put a syntax inside JSON strings, so you might write something like "$include" : "filename" or allow "foo" : "${bar}". This is a way to tunnel new syntax through a JSON parser, but other than the implementation benefit (using a standard JSON parser), it doesn't really work. It's a bad syntax for human maintenance, and it's not valid JSON anymore because properly interpreting it requires treating some valid JSON strings as something other than plain strings. A better approach is to allow mixing true JSON files into the config but also support a nicer format.

Other APIs (Wrappers and Ports)

This may not be comprehensive - if you'd like to add mention of your wrapper, just send a pull request for this README. We would love to know what you're doing with this library or with the HOCON format.

Scala wrappers for the Java library

Ruby port