userver: Dynamic config
Loading...
Searching...
No Matches
Dynamic config

For schemas of dynamic configs used by userver itself:

See also
Dynamic config schemas

For information on how to write a service that distributes dynamic configs:

See also
Writing your own configs server

Dynamic config is a system of options that can be changed at runtime without restarting the service.

Dynamic config is distributed as a large JSON object where each direct member is called a dynamic config variable, or (somewhat confusingly) a dynamic config. For example:

{
// ...
"POSTGRES_DEFAULT_COMMAND_CONTROL": {
"network_timeout_ms": 750,
"statement_timeout_ms": 500
},
"USERVER_DEADLINE_PROPAGATION_ENABLED": true,
// ...
}

Why your service needs dynamic configs

Dynamic configs are an essential part of a reliable service with high availability. The configs could be used as an emergency switch for new functionality, selector for experiments, limits/timeouts/log-level setup, proxy setup and so forth.

See Production configs and best practices setup example.

Adding and using your own dynamic configs

Dynamic config values are obtained via the dynamic_config::Source client that is retrieved from components::DynamicConfig:

#include <userver/dynamic_config/value.hpp>
namespace myservice::smth {
Component::Component(const components::ComponentConfig& config,
: components::LoggableComponentBase(config, context),
config_(
// Searching for some component to initialize members
context.FindComponent<components::DynamicConfig>()
.GetSource() // getting "client" from a component
) {
// Reading config values from static config
[[maybe_unused]] auto url = config["some-url"].As<std::string>();
const auto fs_tp_name = config["fs-task-processor"].As<std::string>();
// Starting a task on a separate task processor from config
auto& fs_task_processor = context.GetTaskProcessor(fs_tp_name);
utils::Async(fs_task_processor, "my-component/fs-work", [] { /*...*/ }).Get();
// ...
}
} // namespace myservice::smth

To read the config, you first need to define a global dynamic_config::Key variable for it. Second, you should get the current value from the config using the Key:

namespace myservice::smth {
inline const dynamic_config::Key kMyConfig{"SAMPLE_INTEGER_FROM_RUNTIME_CONFIG",
42};
int Component::DoSomething() const {
// Getting a snapshot of dynamic config.
const auto runtime_config = config_.GetSnapshot();
return runtime_config[kMyConfig];
}
} // namespace myservice::smth

You can also subscribe to dynamic config updates using dynamic_config::Source::UpdateAndListen functions, see their docs for details.

What is needed to define a dynamic config

A dynamic_config::Key requires 3 main elements:

  1. The config name.
    • We use UPPER_CASE for config names.
  2. The way to parse the JSON value into a C++ object.
    • The default way that should be used in most scenarios is the normal JSON formats parsing.
      See also
      dynamic_config_parsing
    • For edge case scenarios, dynamic_config::Key constructors provide various other ways to parse configs, refer to its documentation for details.
  3. The default config value.

Summary of main dynamic config classes

dynamic_config::Snapshot

A type-safe map that stores a snapshot of all used configs. dynamic_config::Snapshot is cheaply copyable (and even more cheaply movable), has the semantics of std::shared_ptr<const Impl>. An obtained dynamic_config::Snapshot instance should be stored while its configs are used or referred to.

dynamic_config::Source

A reference to the configs storage, required for obtaining the actual config value and subscription to new values.

components::DynamicConfig

The component that stores configs in production and testsuite tests.

For unit tests, dynamic_config::StorageMock is used instead.

See also
dynamic_config_unit_tests

Recommendations on working with dynamic config

  1. Don't store and pass around components::DynamicConfig&. Most code should immediately call GetSource() on it, then use or store the dynamic_config::Source in a field.
  2. Don't use dynamic_config::Source::UpdateAndListen to take a piece of config in OnConfigUpdate and store it into a "comfy" class field. It creates a synchronization problem out of the blue, while it has already been solved in the dynamic config API. Just store dynamic_config::Source in a field and call dynamic_config::Source::GetSnapshot where you need to read the config.
  3. When using dynamic_config::Source::UpdateAndListen, be careful with the lifetime of the subscription handle (concurrent::AsyncEventSubscriberScope):
    • Put the subscription in the class, which method is specified as the callback, as the last field, or at least below all the fields that the callback needs to work. On class destruction, first the subscription object will be destroyed, guaranteeing that OnConfigUpdate will not be called afterward; then other fields of the class will be destroyed.
    • If the class has a non-default destructor (or might have in the future), call concurrent::AsyncEventSubscriberScope::Unsubscribe somewhere in the beginning of the destructor to ensure that OnConfigUpdate will not be called at the point when it is already unable to run safely.

Parsing dynamic configs

The following can be applied to parsing formats::json::Value in any context, but it frequently comes up in the context of defining configs.

If your dynamic config is any more complex than a trivial type, then you need to ensure that JSON parsing is defined for it.

Trivial types

JSON leafs can be parsed out of the box:

OpenAPI C++ type
boolean bool
integer integer types, e.g. uint32_t, size_t
numeric double
string std::string

If the whole config variable is of such type, then the default value for it can be passed to dynamic_config::Key directly (without using dynamic_config::DefaultAsJsonString).

Durations

Chrono durations are stored in JSON as integers:

  • json.As<std::chrono::seconds>() will parse 10 as 10 seconds;
  • json.As<std::chrono::milliseconds>() will parse 10 as 10 milliseconds;
  • same for minutes and hours.

To allow humans to more easily differentiate between them while looking at the JSON, we recommend ending the config name or the object label with one of the following suffixes: _MS, _SECONDS, _MINUTES, _HOURS (in the appropriate capitalization).

Enums

A string that may only be selected a finite range of values should be mapped to C++ enum class. Parsers for enums currently have to be defined manually. Example enum parser:

OverloadAction Parse(const formats::json::Value& value,
static constexpr utils::TrivialBiMap kMap([](auto selector) {
return selector()
.Case(OverloadAction::kCancel, "cancel")
.Case(OverloadAction::kIgnore, "ignore");
});
return utils::ParseFromValueString(value, kMap);
}

Structs

Objects are represented as C++ structs. Parsers for structs currently have to be defined manually. Example struct config:

struct SampleStructConfig final {
bool is_foo_enabled;
std::chrono::milliseconds bar_period;
};
extern const dynamic_config::Key<SampleStructConfig> kSampleStructConfig;
// Parser for a type must be defined in the same namespace as the type itself.
// In this example, JSON parser is only needed for dynamic config, so declaring
// it in hpp is not strictly necessary.
SampleStructConfig Parse(const formats::json::Value& value,
return SampleStructConfig{
/*is_foo_enabled=*/value["is_foo_enabled"].As<bool>(),
/*bar_period=*/value["bar_period_ms"].As<std::chrono::milliseconds>(),
};
}
const dynamic_config::Key<SampleStructConfig> kSampleStructConfig{
"SAMPLE_STRUCT_CONFIG", dynamic_config::DefaultAsJsonString{R"(
{
"is_foo_enabled": false,
"bar_period_ms": 42000
}
)"}};

Optionals

To represent optional (non-required) properties, use std::optional or As with default. For example:

  • json.As<std::optional<int>>() will parse a missing value as an empty optional;
  • json.As<int>(42) will parse a missing value as 42;
  • json.As<std::vector<int>>({}) will parse a missing value as a default-constructed vector;
  • to assign field from config if it exists, otherwise leave the default field value: field = json.As<T>(field).

To enable support for std::optional:

Containers

To represent JSON arrays, use containers, typically:

  • std::vector<T>
  • std::unordered_set<T>
  • std::set<T>

To represent JSON objects with unknown keys (OpenAPI's additionalProperties), use map containers, typically:

  • std::unordered_map<std::string, T>
  • std::map<std::string, T>

To enable support for such containers, use the following:

If the nested type is your custom type, make sure to define Parse for it:

// To parse this type
std::unordered_map<std::string, std::optional<MyStruct>>
// You need to define this parser
MyStruct Parse(const formats::json::Value& value,

Recommendations on parsing dynamic configs

  1. Do not repeat the work of defining parsers for containers themselves. Define the parser for the inner type instead.
  2. Do not hesitate to throw an exception if the provided config is invalid. There are multiple tiers of fallback mechanisms in place to make sure it will not break the whole service. Exceptions are the preferred mechanism to signal this failure to the dynamic config system.
  3. Make sure to check feasible invariants while parsing, e.g. minimum and maximum values of numbers. For that you occasionally may need to use dynamic_config::Key::Key(std::string_view name, JsonParser parser, DefaultAsJsonString default_json) Don't leave strings that are semantically enums as strings, parse them to C++ enum class.
  4. There is typically no need to be cautious about malicious input coming from the configs service. Still, in case of a human error, it's far better to fail a config update than to crash the service or to start silently ignoring the config.

Dynamic config defaults

Config defaults are specified in their C++ definition. They may be overridden in the static config of dynamic-config component:

See also
dynamic_config_defaults_override

If utils::DaemonMain is used, then the default dynamic configs can be printed by passing --print-dynamic-config-defaults command line option. It does not account for overrides in dynamic-config (components::DynamicConfig).

Dynamic config defaults are used in the following places:

  1. In unit tests and benchmarks
  2. In testsuite tests, defaults are the main source of configs (they can be overridden specifically for testsuite).
  3. Unconditionally if config updates are not set up (dynamic-config.updates-enabled: false)
  4. If dynamic-config-client-updater (components::DynamicConfigClientUpdater) (or a custom updater component) receives response from the configs service with some configs missing, then dynamic-config (components::DynamicConfig) fills in those configs from defaults.
  5. If dynamic-config component loads the config cache file, and some configs are missing, then those are filled in from defaults.

Setting up dynamic config

dynamic-config (components::DynamicConfig) itself is required for the functioning of various core userver components, so it is included in both components::MinimalComponentList and components::CommonComponentList.

By default, it is configured to function in the "static dynamic config" mode and may be omitted from the static config entirely.

Overriding dynamic configs defaults

If you rely on obtaining ground truth configs from a configs service (as it is advised for best experience; described below), then you may skip this section entirely and use defaults only for testing.

On the other hand, if you prefer to run without a configs service, then you may want to override some of the built-in userver configs.

Defaults can be overridden in the config of the dynamic-config component:

# yaml
dynamic-config:
defaults:
POSTGRES_CONNECTION_SETTINGS:
key-value-database:
max-ttl-sec: 60
recent-errors-threshold: 100000
POSTGRES_DEFAULT_COMMAND_CONTROL:
network_timeout_ms: 750
statement_timeout_ms: 500
POSTGRES_QUERIES_COMMAND_CONTROL:
sample_select_value:
network_timeout_ms: 70
statement_timeout_ms: 40
sample_transaction_insert_key_value:
network_timeout_ms: 200
statement_timeout_ms: 150

In this case YAML is automatically converted to JSON, then parsed as usual.

Alternatively, you can pass the path to a JSON file with defaults:

# yaml
dynamic-config:
defaults-path: $dynamic-config-defaults

It somewhat complicates the deployment process, though.

Setting up dynamic config updates

Suppose that we've managed to persuade you to wind up a configs service and start actively using dynamic configs.

You will need the following components to download configs:

Both are included in components::CommonComponentList.

Here is a reasonable static config for those:

# yaml
dynamic-config:
updates-enabled: true
fs-cache-path: $config-cache
fs-task-processor: fs-task-processor
dynamic-config-client:
config-url: $config-server-url
http-retries: 5
http-timeout: 20s
service-name: $service-name
dynamic-config-client-updater:
config-settings: false
first-update-fail-ok: true
full-update-interval: 1m
update-interval: 5s

Fallback mechanisms for dynamic configs updates

Suppose that the new revision of the current service released before the newly added config was accounted by the config service. In this case it should just return the configs it knows about. The current service will fill in the blanks from the defaults.

If the config service gives out invalid configs (which fail to be parsed), then the periodic config update will fail, and alerts will fire:

  1. dynamic-config.parse-errors metric (1) will be incremented, and dynamic-config.was-last-parse-successful metric (2) will be set to 1. You can combine those to safely detect parsing errors in the metrics backend:
    if ((1) != previous (1) || (2) == 0) show alert
  2. An alert will be registered in alerts::StorageComponent, which can be accessed from outside using the optional alerts::Handler component

If the config service is not accessible at this point (down or overloaded), then the periodic config update will also obviously fail.

After a config update failure, if the service has already started, then it will keep using the previous successfully fetched configs. Also, you can monitor config update health using

cache.any.time.time-from-last-successful-start-ms: cache_name=dynamic-config-client-updater

and related metrics.

See also
Service Statistics and Metrics (Prometheus/Graphite/...)

If the first config update fails, then the service will read the config cache file specified in dynamic-config.fs-cache-path, which is hopefully left since the previous service start.

If the first config update fails, and the config cache file is missing, then the service will fail to start, showing a helpful log message. Defaults are not used in this case, because they may be significantly outdated, and to avoid requiring the developer to always keep defaults up-to-date with production requirements. Another reason for such behavior is that the dynamic configs are used to fix up incidents, so such check of the dynamic configs service at first deployment prevents incident escalation due to unnoticed misconfiguration (missing routes to dynamic config service, broken authorization, ...).

If you still wish to boot the service using just dynamic config defaults, you can create a config cache file with contents {}, or bake a config cache file into the service's Docker image.

Using dynamic config in unit tests and benchmarks

Main testing page:

See also
Unit Tests and Benchmarks

dynamic_config::StorageMock stores configs for unit tests and benchmarks. It must be kept alive while any dynamic_config::Source or dynamic_config::Snapshot is accessing it.

Test code can obtain the default configs using dynamic_config::GetDefaultSnapshot or dynamic_config::GetDefaultSource, or override some configs using dynamic_config::MakeDefaultStorage.

Using dynamic config in testsuite

Main testsuite page:

See also
Functional service tests (testsuite)

Overriding dynamic config for testsuite globally

Dynamic config can be overridden specifically for testsuite. It can be done globally in the following ways:

  1. Providing a JSON file with --config-fallback option to pytest. When using userver_testsuite_add_simple to setup tests in CMake, it is enough to place the dynamic_config_fallback.json file next to the static config.
  2. Providing a patch directly in "Python JSON" format by overriding dynamic_config_fallback_patch fixture

The various config patches are applied in the following order, going from the lowest to the highest priority:

  1. In-code defaults from dynamic_config::Key
  2. dynamic-config.defaults option from static config , if any
  3. JSON file passed to --config-fallback, if any
  4. dynamic_config_fallback_patch, if any

Changing dynamic configs in testsuite per test

If the service has config updates disabled, then there is no way to change configs per-test. You can resort to creating multiple separate directories for testsuite tests and overriding the initial dynamic config (as shown above) in each of those directories in different ways.

If the service has config updates enabled, then you can change dynamic config per-test as follows:

@pytest.mark.config(MY_CONFIG_NAME=42, MY_OTHER_CONFIG_NAME=True)
async def test_whatever(service_client, ...):

Dynamic config can also be modified mid-test using dynamic_config fixture.

Such dynamic config changes are applied (sent to the service) at the first service_client request in the test, or manually:

await service_client.update_server_state()