Launchdarkly-server-sdk

Added:
- The `all_flags_state` method now accepts a new option, `details_only_for_tracked_flags`, which reduces the size of the JSON representation of the flag state by omitting some metadata. Specifically, it omits any data that is normally used for generating detailed evaluation events if a flag does not have event tracking or debugging turned on.

Changed:
- The SDK previously contained a copy of code from the `expiringdict` package. This has been changed to use the current version of that package from PyPi.

Fixed:
- JSON data from `all_flags_state` is now slightly smaller even if you do not use the new option described above, because it omits the flag property for event tracking unless that property is true.

Fixed:
- In polling mode, if the client received an HTTP error from LaunchDarkly, it stopped polling. This has been fixed so it only stops polling if the error is 401 (indicating an invalid SDK key).
- When using a Redis feature store, if the `hgetall` method returned an invalid result, `all_flags` and `all_flags_state` would throw an exception. Instead, `all_flags` will now return an empty dict, and `all_flags_state` will return a state object with no flags and `valid==False`. (Thanks, [thieman](https://github.com/launchdarkly/python-client/pull/99)!)

Fixed:
- In Python 3, if the Redis feature store encountered a Redis exception, it would crash on trying to log the `message` property of the exception, which does not exist in Python 3. This has been fixed. (Thanks, [mattbriancon](https://github.com/launchdarkly/python-client/pull/96)!)

Added:
- The new `LDClient` method `variation_detail` allows you to evaluate a feature flag (using the same parameters as you would for `variation`) and receive more information about how the value was calculated. This information is returned in an `EvaluationDetail` object, which contains both the result value and a "reason" object which will tell you, for instance, if the user was individually targeted for the flag or was matched by one of the flag's rules, or if the flag returned the default value due to an error.

Fixed:
- When evaluating a prerequisite feature flag, the analytics event for the evaluation did not include the result value if the prerequisite flag was off.

Added:
- The new `LDClient` method `all_flags_state()` should be used instead of `all_flags()` if you are passing flag data to the front end for use with the JavaScript SDK. It preserves some flag metadata that the front end requires in order to send analytics events correctly. Versions 2.5.0 and above of the JavaScript SDK are able to use this metadata, but the output of `all_flags_state()` will still work with older versions.
- The `all_flags_state()` method also allows you to select only client-side-enabled flags to pass to the front end, by using the option `client_side_only=True`.

Deprecated:
- `LDClient.all_flags()`

Changed:
- In streaming mode, each connection failure or unsuccessful reconnection attempt logs a message at `ERROR` level. Previously, this message included the amount of time before the next retry; since that interval is different for each attempt, that meant the `ERROR`-level messages were all unique, which could cause problems for monitors. This has been changed so the `ERROR`-level message is always the same, and is followed by an `INFO`-level message about the time delay. (Note that in order to suppress the default message, the LaunchDarkly client modifies the logger used by the `backoff` package; if you are using `backoff` for some other purpose and _do_ want to see the default message, set `logging.getLogger('backoff').propagate` to `True`.) ([88](https://github.com/launchdarkly/python-client/issues/88))