Skip to main content
Version: 1.3

Changes to default composition order

Default composition order is changing in Hydra 1.1.

For this example, let's assume the following two configs:

config.yaml
defaults:
  - foo: bar

foo:
  x: 10
foo/bar.yaml
# @package _group_
x: 20



In Hydra 1.0, configs from the Defaults List are overriding config.yaml, resulting in the following output:

foo:
  x: 20

As of Hydra 1.1, config.yaml is overriding configs from the Defaults List, resulting in the following output:

foo:
  x: 10

Migration​

If your application uses hydra.main, the best way to verify that updating Hydra versions does not change your job configurations is to compare the output of python my_app.py --cfg job on both the new and old Hydra versions. If your application uses the Compose API, please make sure you have comprehensive unit tests on the composed configuration.

Primary config is a YAML file​

To ensure this change is not missed by people migrating from Hydra 1.0, Hydra 1.1 issues a warning if the Defaults List in the primary config is missing _self_, and there are config values in addition to the Defaults List.
To address the warning, add _self_ to the Defaults List of the primary config.

  • If the new behavior works for your application, append _self_ to the end of the Defaults List.
  • If your application requires the previous behavior, insert _self_ as the first item in your Defaults List.
config.yaml
defaults:
  - _self_
  - foo: bar

foo:
  x: 10
Output config
foo:
  x: 20




The Defaults List is described here.

Primary config is a Structured Config​

Structured Configs used as primary config may see changes as well. You should add _self_ to the defaults list to indicate the composition order. In such cases you will typically want _self_ to be the first item in the defaults list. 

defaults = [
    "_self_",
    {"db": "mysql"}
]

@dataclass
class Config:
    # this is unfortunately verbose due to @dataclass limitations
    defaults: List[Any] = field(default_factory=lambda: defaults)

    # Hydra will populate this field based on the defaults list
    db: Any = MISSING

Primary config is a config file with a Structured Config schema​

If you use Structured Config as a schema for your primary config, be sure to add _self_ after the schema in the Defaults List, otherwise the schema will override the config instead of the other way around.

my_app.py
@dataclass
class Config:
  host: str = "localhost"
  port: int = 8080

cs.store(name="base_config", 
         node=Config)
config.yaml
defaults:
 - base_config  # schema
 - _self_       # after schema

port: 3306


Output config
host: localhost # schema
port: 3306      # config.yaml





Compatibility with both Hydra 1.0 and 1.1​

If your config must be compatible with both Hydra 1.0 and 1.1, Insert _self_ as the first item in the Defaults List. Hydra 1.0.7 (or newer releases in Hydra 1.0) ignores _self_ in the Defaults List and Hydra 1.1 will compose the same config as Hydra 1.0 if _self_ is the first item.