Statsig

Statsig helps you move faster with feature gates (feature flags), and/or dynamic configs. It also allows you to run A/B/n tests to validate your new features and understand their impact on your KPIs. If you're new to Statsig, check out our product and create an account at statsig.com.

Getting Started

Read through the Documentation or check out the Samples.

Development

Use the Node.js version in .node-version for repository tooling and local builds. For example, with nvm:

nvm install "$(cat .node-version)"

This is the repository toolchain version; it does not change the Node.js versions supported by the published SDK.

ID-list file routing and fallback

See ID-list loading for eager and lazy loading semantics, timeouts, concurrency, refresh behavior, and binding support.

download_id_list_file_api selects the preferred base URL for file downloads. Enable fallback_to_statsig_api to try the original file URL from the manifest after an overridden request fails with a retryable network error, provided that URL belongs to the trusted HTTPS OpenAI CDN origin:

let options = statsig_rust::StatsigOptions {
    download_id_list_file_api: Some("http://statsig-proxy:8080".to_string()),
    fallback_to_statsig_api: Some(true),
    ..Default::default()
};

Fallback accepts only https://statsigcdn.openai.com with the default port or explicit :443. Fallback sends no request or SDK credentials to other manifest origins; the original request error is returned. Fallback does not follow redirects. Custom network providers must implement send_without_redirects; otherwise the fallback returns an error without sending a request.

All ID-list file downloads, including direct CDN, proxy, and fallback requests, preserve the URL's path and query except for legacy range parameters, which are removed. Every request sends Range, statsig-id-list-file-size, and Accept-Encoding: identity headers, preserving the file identity and byte offset for incremental updates. Partial responses require a valid Content-Range matching the requested offset and received body length. They may use 206 or, for compatibility, 200. A 200 without Content-Range means the complete available file from byte zero; the SDK applies only bytes after the requested offset. Available bytes may be behind or ahead of the manifest: the SDK advances by bytes actually applied and keeps polling while below the advertised size. Responses that make no progress or have inconsistent ranges use the remaining retry budget without publishing the refresh. Partial bodies without Content-Range are unsupported; a malformed header is rejected even when the status is 200.

Set fallback_to_statsig_api to Some(false) or leave it unset to keep file downloads strictly on the selected destination. There is no fallback attempt when the override resolves to the original URL, no override is set, or the request fails with an authentication rejection (401/403) or another non-retryable error. Response parsing failures do not trigger fallback. This uses the existing option across language bindings; wrappers that disable it continue to opt out.

A successful proxy request makes no additional request. Primary and fallback requests share the same per-file attempt budget (three attempts by default). The next synchronization starts with the configured override again.

GNU/Linux heap trimming

By default, the SDK calls glibc's malloc_trim(0) after eligible config updates and mmap writes. To skip these calls across all SDK instances and native bindings in a process, set NO_MALLOC_TRIM in the worker environment.

export NO_MALLOC_TRIM=1

The SDK checks the variable before each trim. Only 1 and lowercase true skip the call. An unset variable or any other value keeps trimming enabled. For Flink, set the variable in the TaskManager environment and restart the workers so their Python processes inherit it. To restore the default, unset the variable and restart the workers.

Read-only data stores

Rust data stores can override DataStoreTrait::is_read_only(); Python DataStore subclasses can override the synchronous is_read_only() method:

def is_read_only(self) -> bool:
    return not self.is_leader

The default is False, preserving existing writable stores. Return True for a follower that only consumes shared cache entries. Bootstrap/polling reads and config hydration/evaluation remain enabled, but full network responses no longer build or compress hydrated cache snapshots or call the datastore's write methods. Delta responses continue to skip snapshot writes. Keep the role stable during an update. A failing Python capability check is treated as read-only.

Remote SDK telemetry control

Set StatsigOptions.disable_remote_telemetry to Some(true) to suppress Statsig-side diagnostics and SDK exception uploads. ObservabilityClient metrics/errors, local logs, console capture, custom events, and exposures remain active as configured.

Only boolean true enables suppression; false or unset/null preserves existing reporting. The option is captured per SDK/adapter at construction. Configure it before creating either; roll back by recreating them with false or unset. No process restart is needed. Environment variables and remote config cannot override it, and opposite settings remain isolated even under the same SDK key/instance ID.

Packages

Protobuf-typed reads and publication callbacks are described in Typed dynamic configs, with Python, Node, and Go API guides. See Python mmap preload for optional returnable hash precomputation when loading writer-published artifacts. See Verified blob reuse from mmap for automatic reuse of writer-verified large config values during SDK hydration.

Bindings

• Node [npm] [source] [docs]

• Python [pypi] [source] [docs]

• Java [internal Maven release guide] [source] [docs]

• Rust [crates.io] [source] [docs]

Community

If you need any assitance or just have a question, feel free to reach out to us on Slack.