Python Server SDK

Introduction

The Python Server SDK is a server-side SDK that evaluates configs and their targeting rules. Upon initialization, it retrieves configs and targeting rules from ConfigDirector services. From that point on, it evaluates configs locally for any context, and receives updates via server sent events (SSE) when configs are updated on the dashboard or via the admin API.

The minimum Python version supported is 3.10.

Installation

The SDK can be installed from PyPI: https://pypi.org/project/configdirector-server-sdk/

pip install configdirector-server-sdk

Configure and initialize the client

1. Create an instance of the client providing your server SDK key. You can retrieve a server SDK key for each environment under SDK keys in the dashboard's navigation panel.
2. Initialize the client to initiate its connection lifecycle.

from configdirector import create_client

client = create_client("YOUR-SERVER-SDK-KEY")

# Blocks until the client is initialized or it times out.
# If initialization times out, the client will continue attempting to
# initialize in the background.
client.initialize()

Additional configuration options

metadata

The metadata option allows you to provide your application's name and version. These values can be used in targeting rules conditionals. For example, if a certain feature should only be enabled starting with a certain version of your application.

from configdirector import Metadata, create_client

client = create_client(
    "YOUR-SERVER-SDK-KEY",
    metadata=Metadata(app_name="YOUR-APP-NAME", app_version="1.0.2"),
)

client.initialize()

log_level and logger

The SDK logs through the standard library to a logger named configdirector. Logging levels, formatting, and destinations can be controlled by accessing the logger:

import logging

logging.getLogger("configdirector").setLevel(logging.DEBUG)

The SDK sets no level of its own, so this logger follows the usual logging rules. With nothing configured at all, Python still surfaces warnings and errors on stderr. If you would rather not configure logging, the log_level parameter can be used to set a level on the logger:

client = create_client("YOUR-SERVER-SDK-KEY", log_level="DEBUG")

The logger can be overridden by passing any object implementing ConfigDirectorLogger (a different stdlib logger, or your own adapter) in the logger parameter:

client = create_client("YOUR-SERVER-SDK-KEY", logger=logging.getLogger("my_app.configdirector"))

connection

The connection parameter accepts a ConnectionOptions object with the following optional values:

• mode
  • The connection mode, one of streaming, polling or offline. It is recommended to use the default of streaming unless you have a specific need to use polling instead. In offline mode the client never connects to ConfigDirector and serves the values you give it; see Run without a connection.
  • Defaults to streaming
• file
  • The path of a JSON file holding the values to serve in offline mode, as a str or a pathlib.Path, resolved against the process working directory. Only one of file and values can be given, and only in offline mode. See Run without a connection.
• values
  • The values to serve in offline mode, keyed by config key. Only one of file and values can be given, and only in offline mode. See Run without a connection.
• polling_interval
  • Only used in polling mode. The interval, in seconds, to poll ConfigDirector services for updates. A value below the minimum of 60 seconds is raised to 60 seconds and a warning is logged. See Polling intervals.
  • Defaults to 5 minutes (300 seconds)
• timeout
  • The timeout, in seconds, to be used in initialization. This is how long the initialize method will wait for data from ConfigDirector services. If the timeout is reached, initialize will return but the client will still be in an unready status and returning in-code default values. The client will continue to attempt to connect and retrieve config values in the background. It has no effect in offline mode.
  • Defaults to 3 seconds
• url
  • The base URL used to connect to ConfigDirector services
  • This should only be provided if your environment requires you to configure a proxy server in order to connect to ConfigDirector services. It has no effect in offline mode.

from configdirector import ConnectionOptions, create_client

client = create_client(
    "YOUR-SERVER-SDK-KEY",
    connection=ConnectionOptions(mode="polling", polling_interval=600, timeout=5),
)

telemetry

Telemetry tuning. It is unlikely these settings need to be adjusted. However, in cases where your application has a large number of evaluations per second, you can adjust these settings to tune the memory footprint and frequency of telemetry requests.

Keep in mind that ConfigDirector relies on these telemetry events to provide insights and features related to the configs being used.

From version 1.9.0 of the SDK, each evaluation it reports says which variation was served and which percentage rollout and share served it. That is how ConfigDirector counts the contexts each variation of an experiment reached. Tracked events are sent with the same reports.

The telemetry object accepts the following optional values:

• event_queue_limit
  • The size limit of telemetry event queues. If the size limit is reached before the events are flushed to the network, older events will be dropped.
  • ConfigDirector keeps a count of dropped events. If the number of dropped events is higher than 50% of the total events, ConfigDirector will issue an alert in the dashboard.
  • A number between 100 and 100,000. Defaults to 5,000.
• flush_interval
  • How often events are flushed and sent over the network in seconds.
  • Decrease this number if your application consistently captures a large number of events in short periods of time in order to reduce memory footprint from a large event queue.
  • Defaults to 30 seconds.

from configdirector import TelemetryOptions, create_client

client = create_client(
    "YOUR-SERVER-SDK-KEY",
    telemetry=TelemetryOptions(event_queue_limit=10_000, flush_interval=15),
)

Hooks

Hooks are events you can subscribe to in order to be notified of some key actions from the client. In order to subscribe to hooks you can provide a handler in the options, or use the on method which returns a Subscription:

subscription = client.on("configs_updated", lambda event: print(event.keys))

from configdirector import ClientHooks, create_client

client = create_client(
    "YOUR-SERVER-SDK-KEY",
    hooks=ClientHooks(config_evaluated=lambda event: print(event.evaluation)),
)

The following hooks are available:

• client_ready: ClientReadyEvent
  • Emitted when the client is initialized. When this event is emitted it means the client has received a payload from the ConfigDirector servers and is ready to evaluate configs.
• configs_updated: ConfigsUpdatedEvent
  • Emitted when a payload is received from the ConfigDirector servers with config data. It is emitted during initialization when an entire config payload is received. After initialization, it is emitted when updates are pushed from the server (or discovered via polling if that connection mode is used).
  • keys: list[str] - The config keys that were included in the payload from the server, and the keys of configs whose targeting rules use a segment the payload included, sorted.
  • removed_keys: list[str] - The keys of configs the payload removed, sorted. A config is removed when a full payload no longer includes it. Empty when nothing was removed.
• config_evaluated: ConfigEvaluatedEvent
  • Emitted whenever a config is evaluated. This includes calls to get_value and evaluations delivered to watchers via watch.
  • evaluation: ConfigEvaluation - A ConfigEvaluation object containing the details of the evaluation:
    • key: str - The config key that was evaluated
    • value: str | int | float | bool | dict | list - The value the config evaluated to. It can be the default value provided to get_value or watch. For example, if the config was evaluated before the client was initialized.
    • value_id: str | None - The value ID, a stable hash of the value that you can send to a third party rather than the value. This can be useful for large values, like a JSON config, or to avoid disclosing the values themselves. It may be None if the evaluation had to fall back to the in-code default value. To report an experiment to an analytics tool, send the variation instead: the pages for Amplitude, Mixpanel, Google Analytics 4, Segment and PostHog name the event each one reads.
    • variation: Variation | None - The variation that was served, with its id and name. The id stays the same when the variation is renamed, and name is None for a variation without one. It is None when the config has no variations or the in-code default value is returned. Versions of the SDK before 1.9.0 do not report it.
    • is_default: bool - Whether or not the evaluation fell back to the default value provided in code.
    • reason: EvaluationReason - The reason for the evaluation resolution. In the case the config successfully evaluated based on server-provided targeting rules, it will be found-match. A context without an id that is served the default targeting rule value after skipping a percentage rollout reports rollout-skipped-no-identifier, with is_default set to False; see Contexts without an id. If the evaluation had to fall back to the default value, the reason will encode why the fallback was required:
      • client-not-ready - The evaluation happened before the client finished initialization
      • config-state-missing - The requested config key was not present in the payload received from the server. This could be due to an incorrect config key, or a config that was not enabled to be available to client SDKs.
      • invalid-number - The config was requested as a number but the value received from the server could not be converted to a number.
      • invalid-boolean - The config was requested as a boolean but the value received from the server could not be converted to a boolean.
      • invalid-json - The config was requested as a JSON object but the value received from the server could not be converted to a JSON object.
      • type-mismatch - The config was requested with a data type that did not match the type of the config and no reasonable type conversion was possible.
      • value-missing - The config key was present in the payload received from the server, but it carried no value.
    • context: Context | None - The context that was provided to the evaluation function, or None if no context was provided.

Run without a connection

Set the connection mode to offline to run your application without connecting to ConfigDirector, for example in a suite of UI tests where each run pins every flag to a known value. An offline client needs no server SDK key, opens no connection and sends nothing, so the application is invisible to the ConfigDirector dashboard: no evaluations are counted and no lifecycle state changes. Every context receives the same value and targeting rules are not applied.

Give the values in a JSON file, with one value per config key, or inline:

from configdirector import ConnectionOptions, create_client

client = create_client(
    connection=ConnectionOptions(mode="offline", file="./flags/new-checkout-on.json"),
)
client.initialize()

{
  "new-checkout": true,
  "max-items": 20,
  "greeting": "hello",
  "theme": { "color": "blue" }
}

from configdirector import ConnectionOptions, create_client

client = create_client(
    connection=ConnectionOptions(mode="offline", values={"new-checkout": True, "max-items": 20}),
)
client.initialize()

The file path is resolved against the process working directory. The config type of each value follows from the value itself: true or false (boolean), a whole number written without a decimal point or exponent (integer), any other number (float), a string, or an object or array (json). Inline values follow the same rule with Python types: a bool, an int, a float, a str, or a dict with str keys or a list. A value read with an in-code default value of another type returns the in-code default value, as it does against ConfigDirector. A watch registered before initialize is called with the offline value when initialize loads it. Offline values never change on their own, so the watch is called again only if initialize is called again, which loads the values, and reads the file, once more. polling_interval, timeout, url and the timeout argument of initialize have no effect.

With mode="offline" and neither file nor values, the client is ready at once and every config returns its in-code default value.

initialize raises a ConfigDirectorInitializationError, which the SDK exports, when the values cannot be loaded: the file is missing, cannot be read, is not a JSON object or holds a value such as null; file and values are both given; or either is given in streaming or polling mode. The client stays not ready, so a test run that starts from a misspelled path fails instead of passing on in-code default values. A later initialize that cannot load the values raises the same way and keeps serving the values loaded before.

If you come from LaunchDarkly, a file of values is what LaunchDarkly calls "flags from files", and offline with no values is its "offline mode".

Retrieve config values

Retrieving config values via get_value, and subscribe to updates via watch. The first argument is the config key, the second is the default value. The type of the default value provided determines the type the config is parsed as. The default value is returned if the config is not available or cannot be parsed as the given type:

# Retrieve config values
retries = client.get_value("max-retries", 3)
theme = client.get_value("theme", "light")
limits = client.get_value("rate-limits", {"per_minute": 60}) # JSON config

# Watch for config updates
def on_change(value: bool) -> None:
    print(f"new-checkout is now {value}")

subscription = client.watch("new-checkout", False, on_change)

# Call close when the subscription is not longer needed
subscription.close()

Evaluate config values with a context

get_value accepts three arguments, the first argument is the config key, the second one is the default value to be returned if the client is not yet initialized, and the third and optional argument is a context to be used for targeting rules evaluation:

from configdirector import Context

client.get_value("my-boolean-config-key", False, Context(id="user-id", name="Example User"))

To evaluate configs for an entity other than a user, such as a tenant, set the context kind with kind:

tenant = Context(id="acme", kind="tenant", name="Acme Inc.", traits={"plan": "enterprise"})

client.get_value("my-boolean-config-key", False, tenant)

Leaving kind out means user. A context kind is 1 to 64 lowercase letters, digits, dashes, or underscores. Contexts of different kinds are separate contexts even when they share an identifier, and a percentage rollout can place them in different shares. A context with an invalid context kind is ignored: the SDK logs an error, and configs are evaluated as if no context had been given. Context kinds are available from version 1.7.0 of the SDK.

watch accepts four arguments, the first is the config key, the second argument is the default value, the third argument is a function that will be executed when the config value is updated, and the fourth and optional argument is a context:

When a full payload no longer includes the watched config, the function is called with the default value, so it always agrees with what get_value would return.

def on_change(value: bool) -> None:
    print(f"new-checkout is now {value}")

subscription = client.watch(
  "new-checkout",
  False,
  on_change,
  Context(
    id="user-id",
    name="Example User",
    traits={"region": "North America"}
  )
)

Contexts without an id

A percentage rollout places each context in a share by its id, so a context without an id skips percentage rollouts: a rollout does not apply to it, and the targeting rules after the rollout are tried, as they are after a conditional rule the context does not match. The default targeting rule value is served when none of them applies. A rule that serves a single value still applies to it. To keep an anonymous visitor in a rollout, give the context a stable id, such as a session ID. Version 1.9.0 of the SDK ended the evaluation at the skipped rollout and served the default targeting rule value even when a later rule matched, and versions before 1.9.0 served every context without an id the first share of the rollout.

Track events

track reports to ConfigDirector that something happened in your application, such as a completed checkout. ConfigDirector shows tracked events on the Live feed and counts them for each variation of a running experiment; see Tracked events.

track accepts the name of the event and an optional context the event happened for. The event's value and details are optional keyword arguments:

from decimal import Decimal

from configdirector import Context

client.track(
    "checkout-completed",
    Context(id="user-id"),
    value=Decimal("49.90"),
    details={"plan": "pro", "seats": 3},
)

client.track("export-finished", Context(id="acme", kind="tenant"))

client.track("nightly-sync-finished")

• value is a number the event measures, such as an amount of money, a count or a duration. It can be an int, a float or a Decimal, and a Decimal is sent with every digit, so Decimal("49.90") arrives as 49.90.
• details is a mapping describing the event. Its values are strings, numbers, booleans, datetime values, or nested mappings and lists. A datetime is sent as ISO 8601 text, such as 2026-12-01T00:00:00+00:00, keeping its UTC offset if it has one.

An event counts for an experiment only when it is tracked for a context with an id, the same one the experiment's config was evaluated for, so pass the context you evaluate with when you track the outcome of an experiment. Without a context, or for a context without an id, the event is still reported and counts for no experiment. For a context marked anonymous, the event is reported without the id, so it counts for no experiment either. A context with an invalid context kind is ignored as it is in an evaluation: the SDK logs an error and reports the event without a context.

track returns nothing and never raises. The event is sent with the SDK's next telemetry report, every 30 seconds unless you changed flush_interval. In offline mode and after close, track does nothing. The SDK applies these limits, and logs one warning for each of them per telemetry report:

ConfigDirector applies limits of its own to the events it stores; see Limits.

track is available from version 1.9.0 of the SDK, which exports the TrackedEventDetails and TrackedEventDetailValue types. A datetime inside details is sent from version 1.9.1; version 1.9.0 left the details out.

Other useful client features

The client provides additional methods.

evaluate

Evaluates a config the way get_value does, with the same three arguments, and returns the ConfigEvaluation behind it — the value, whether it is the in-code default value, the reason, the value_id, the variation, and the context — instead of the value alone. It is the same object the config_evaluated hook reports, so use it where the reason matters at the point of the read, such as when reporting an evaluation to another system; get_value is enough everywhere else. Like every read, it is reported to config_evaluated and counted in telemetry. The SDK exports the ConfigEvaluation, EvaluationReason and Variation types.

evaluation = client.evaluate("my-config-key", False, Context(id="user-id"))
if evaluation.reason == "config-state-missing":
    print("No config with this key reached the SDK.")

evaluate is available from version 1.9.0 of the SDK.

get_all_configs

Returns every config evaluated for a context, as a dict of config keys to ConfigState, for example to hand the values to a page you render on the server. Pass config_keys to get only those keys. It returns an empty dict until the client is ready, and these reads are not counted in telemetry.

configs = client.get_all_configs(Context(id="user-id"), config_keys=["new-checkout"])

Each ConfigState carries the config's id, key and type, the value served as text, its value_id and its variation. When a percentage rollout served the value, it also carries the rollout's rule_id and share_id. variation, rule_id and share_id are available from version 1.9.0 of the SDK.

Test your code

Use the SDK's configdirector.testing module to test the code that reads your configs and flags. It creates a test client: the SDK's real client connected to an in-memory server that your test controls. Only the connection to ConfigDirector and the telemetry are replaced. Everything else is the same code that runs in production, so the code under test behaves exactly as it does against ConfigDirector, and no network connection is opened, no telemetry is sent, and no thread is started. The module ships in the package, so there is nothing else to install.

from configdirector.testing import create_test_client

with create_test_client({"new-checkout": True, "max-items": 20}) as test_client:
    service = CheckoutService(test_client.client)
    test_client.client.initialize()

    assert service.is_new_checkout_enabled("user-123")

    test_client.set_value("new-checkout", False)
    assert not service.is_new_checkout_enabled("user-123")

test_client.client is a ConfigDirectorClient, so it goes anywhere your code accepts one. It starts uninitialized, like a production client, because the code under test usually owns the call to initialize; with no hold or failure armed, initialize completes at once with the stored values. The test client is a context manager that closes its client on exit; unlike the client's own context manager, entering it does not initialize the client.

Module-level clients

An application that creates and initializes its client when a module loads, as the Flask sample did, runs the real client before any test can intervene, which opens a connection and starts threads. A monkeypatch also has to reach every module that bound the name: from configdirector_client import client binds its own client, so patching configdirector_client.client does not change what the importing module sees. Structure the application around a lazy getter instead, and replace that in tests:

import os
import threading

from configdirector import ConfigDirectorClient, create_client

_client: ConfigDirectorClient | None = None
_client_lock = threading.Lock()


def get_client() -> ConfigDirectorClient:
    global _client
    with _client_lock:
        if _client is None:
            _client = create_client(os.environ["CONFIGDIRECTOR_SERVER_SDK_KEY"])
            _client.initialize()
        return _client

The lock matters on a threaded server: without it, two first requests arriving together each build and initialize a client.

import pytest

import app
from configdirector.testing import create_test_client


def test_new_checkout(monkeypatch: pytest.MonkeyPatch) -> None:
    with create_test_client({"new-checkout": True}) as test_client:
        test_client.client.initialize()
        monkeypatch.setattr(app, "get_client", lambda: test_client.client)

        assert app.test_client().get("/checkout").json["new-checkout"] is True

Patch the name in the module that calls it (app.get_client when app.py does from configdirector_client import get_client), and initialize the test client where the application's getter would have.

Values

create_test_client, set_value, and replace_values take native values, and the config type follows from each value's type: a bool; an int (an integer config); a float (a float config); a str; or a dict with str keys or a list (a JSON config). A str is always a string config, so JSON text meant as a JSON config is given as json.loads(text). Every value is served as an unconditional config, so every context receives the same value; a test that needs different values per context is written as one test per value. None, an unsupported type such as a tuple or a set, or a dict with a non-str key raise ConfigDirectorTypeError; a blank key, a non-finite number, or an integer with too many digits to write out raise ConfigDirectorValidationError. In every case nothing changes.

Reads behave as they do in production: set_value("k", True) read as a string returns the in-code default value with the type-mismatch reason, set_value("k", 2.5) read with an int default returns the default with the invalid-number reason, and set_value("k", "") serves the in-code default value with the value-missing reason. Reasons arrive on the config_evaluated event. JSON objects keep their insertion order.

Controls

After a failed attempt, the next initialize succeeds and delivers the values stored in the meantime.

with create_test_client({"new-checkout": True}, timeout=0.5) as test_client:
    test_client.fail_initialization()

    test_client.client.initialize()

    assert not test_client.client.is_ready
    assert not service.is_new_checkout_enabled("user-123")

Options

create_test_client(values, timeout=..., logger=...) takes timeout (the client's connection timeout in seconds, which bounds how long a held initialize waits; defaults to the SDK's production timeout, and a timeout given to initialize itself takes precedence) and logger (defaults to the standard library logger named configdirector, so pytest's caplog captures it).

What to expect

• The first update is delivered while initialize connects, so watches registered before initialize run before it returns, and configs_updated fires before client_ready.
• Watches and handlers run on the thread that calls set_value, remove_value, replace_values, complete_initialization, or initialize, not on a transport thread.
• A held initialize that times out, or that is interrupted by close, leaves the client not ready until the next initialize, and logs the SDK's timeout warning; production streaming would keep retrying in the background. Readiness is sticky: once the client is ready, a later held, failed, or timed-out attempt leaves it ready with its previous values, and completing such an attempt fires no client_ready.
• Watches fire on every update carrying their key, whether or not the value changed, as in production. remove_value and replace_values deliver a full update, so they also fire the watches of every remaining key.
• An exception raised by a watch or handler is logged and does not propagate, as with the production transports.
• After close, is_ready is false, get_value serves the last values, get_all_configs returns an empty dict, initialize raises ConfigDirectorValidationError, every watch and event handler has been removed, the test client's controls are silent no-ops, and no SDK thread is left running.