Dynamic Variables
Template substitutions¶
Dynaconf has 2 tokens to enable string substitutions, the interpolation works with @format
and @jinja
.
@format token¶
Dynaconf allows template substitutions for strings values, by using the @format
token prefix and including placeholders accepted by Python's str.format
method Dynaconf will call
it lazily upon access time.
The call will be like:
"<YOURVALUE>".format(env=os.environ, this=dynaconf.settings)
So in your string you can refer to environment variables via env
object, and also to variables defined int the settings object itself via this
reference. It is lazily evaluated on access it will use the final value for a settings regardless the order of load.
Example:
export PROGRAM_NAME=calculator
settings.toml
[default]
DB_NAME = "mydb.db"
[development]
DB_PATH = "@format {env[HOME]}/{this.current_env}/{env[PROGRAM_NAME]}/{this.DB_NAME}"
{env[HOME]}
is the same asos.environ["HOME"]
or$HOME
in the shell.{this.current_env}
is the same assettings.current_env
{env[PROGRAM_NAME]}
is the same asos.environ["PROGRAM_NAME"]
or$PROGRAM_NAME
in the shell.{this.DB_NAME}
is the same assettings.DB_NAME
orsettings["DB_NAME"]
so in your program
from dynaconf import settings
settings.DB_PATH == '~/DEVELOPMENT/calculator/mydb.db'
@format
token can be used together with tokens like @str
, @int
, @float
, @bool
, and @json
to cast the results parsed by @format
.
Example:
Casting to integer from @format
templated values
NUM_GPUS: 4
# This returns the integer after parsing
FOO_INT: "@int @format {this.NUM_GPUS}"
# This returns the string after parsing
FOO_STR: "@format {this.NUM_GPUS}"
@jinja token¶
If jinja2
package is installed then dynaconf will also allow the use jinja to render string values.
Example:
export PROGRAM_NAME=calculator
settings.toml
[default]
DB_NAME = "mydb.db"
[development]
DB_PATH = "@jinja {{env.HOME}}/{{this.current_env | lower}}/{{env['PROGRAM_NAME']}}/{{this.DB_NAME}}"
so in your program
from dynaconf import settings
settings.DB_PATH == '~/development/calculator/mydb.db'
The main difference is that Jinja allows some Python expressions to be evaluated such as {% for, if, while %}
and also supports calling methods and has lots of filters like | lower
.
Jinja supports its built-in filters listed in Builtin Filters Page and Dynaconf includes additional filters for os.path
module: abspath
. realpath
, relpath
, basename
and dirname
and usage is like: VALUE = "@jinja {{this.FOO | abspath}}"
@jinja
token can be used together with tokens like @str
, @int
, @float
, @bool
, and @json
to cast the results parsed by @jinja
.
Example:
Casting to integer from @jinja
templated values
NUM_GPUS: 4
# This returns the integer after parsing
FOO_INT: "@int @jinja {{ this.NUM_GPUS }}"
# This returns the string after parsing
FOO_STR: "@jinja {{ this.NUM_GPUS }}"
Example:
Casting to a json
dict from @jinja
templated values
MODEL_1:
MODEL_PATH: "src.main.models.CNNModel"
GPU_COUNT: 4
OPTIMIZER_KWARGS:
learning_rate: 0.002
MODEL_2:
MODEL_PATH: "src.main.models.VGGModel"
GPU_COUNT: 2
OPTIMIZER_KWARGS:
learning_rate: 0.001
# Flag to choose which model to use
MODEL_TYPE: "MODEL_2"
# This returns the dict loaded from the resulting json
MODEL_SETTINGS_DICT: "@json @jinja {{ this|attr(this.MODEL_TYPE) }}"
# This returns the raw json string
MODEL_SETTINGS_STR: "@jinja {{ this|attr(this.MODEL_TYPE) }}"
Name aliasing¶
In certain cases you need to reuse some variable value respecting the exact data type it is previously defined in another step of the settings load.
Example:
from dynaconf import Dynaconf
settings = Dynaconf(settings_files=["settings.toml", "other.toml"])
thing = "value"
number = 42
other_thing = "@get thing"
# providing default if thing is undefined
other_thing = "@get thing 'default value'"
## type casting
# Ensure the aliased value is of type int
other_number = "@get number @int"
# Ensure type and provide default if number is undefined
other_number = "@get number @int 12"
The @get
is lazily evaluated and subject to DynaconfFormatError
in case of
malformed expression, it is recommended to add validators.