Skip to main content

Android Thin SDK

Last updated on

This guide covers the FME Android Thin SDK. Unlike the standard Android SDK, the thin SDK does not download flag definitions or segment data. It calls FME cloud to evaluate flags for a given target and caches the results locally for fast subsequent lookups.

Use the thin SDK when you want to keep rollout configuration server-side, minimize what is shipped on device, or restrict the targeting data exposed to clients.

All of our SDKs are open source. Go to our Android Thin SDK GitHub repository to see the source code.

Before you begin​

The Android Thin SDK is compatible with Android API 21 and later (5.0 Lollipop) and supports apps written in Kotlin or Java.

The library is published on Maven Central as io.split.client:android-thin-client.

Initialization​

To get started, set up FME in your code base with the following two steps.

1. Import the SDK into your project​

Add the SDK as a dependency in your app module's Gradle build script:

build.gradle.kts
dependencies {
    implementation("io.split.client:android-thin-client:1.0.0")
}

2. Instantiate the SDK and create a new SDK factory client​

You build the factory with an SdkKey and a default Target, then get the SDK client from it. The client is what you call getTreatment and track on.

import io.split.client.thin.SdkKey
import io.split.client.thin.SplitFactoryBuilder
import io.split.client.thin.Target

// Build the factory with an SDK key and a default target
val factory = SplitFactoryBuilder.build(
    context = applicationContext,
    sdkKey = SdkKey("YOUR_SDK_KEY"),
    defaultTarget = Target(
        // matchingKey identifies the entity being evaluated: an internal user id,
        // an account id, a device id, or a UUID for anonymous users.
        key = Key(matchingKey = "CUSTOMER_ID"),
        attributes = mapOf("plan" to "premium"), // optional, used for targeting rules
        trafficType = "user",
    ),
)

// Get the client instance bound to the default target
val client = factory.getClient()

We recommend instantiating the SDK factory once as a singleton and reusing it throughout your application. Avoid creating additional factories without a good reason (for example, using different SDK keys in the same app), since each factory opens its own sync connections and consumes additional resources.

SplitFactoryBuilder.build requires a Context (the SDK uses the application context internally), an SdkKey, and a default Target. SplitClientConfig is optional; if omitted, defaults are used.

If the SDK key or matching key fails validation, build returns a no-op factory that logs the error and returns fallback values from every call. Construct your SdkKey and Target carefully (see Configure fallback treatments to control what the no-op factory returns).

Configure the SDK with the SDK key for the FME environment that you would like to access. In legacy Split (app.split.io) the SDK key is found on your Admin settings page, in the API keys section. Select a client-side SDK API key. This is a special type of API token with limited privileges for use in browsers or mobile clients. See API keys to learn more.

A factory can serve more than one target. If your app evaluates flags for multiple users, accounts, or other entities, see Instantiate multiple clients under Advanced use.

Bucketing keys​

Key accepts a separate bucketingKey alongside the matching key, for example to keep a user in the same bucket after they sign in:

val target = Target(
    key = Key(matchingKey = "user-123", bucketingKey = "anon-cookie-abc"),
    trafficType = "user",
)

The matching key drives targeting-rule evaluation; the bucketing key drives the percentage rollout assignment. When bucketingKey is omitted, the matching key is used for both.

Use the SDK​

The following explains how to use this SDK.

Basic use​

After the SDK is instantiated it warms up by fetching the initial set of treatments for the default target from FME cloud. Until that completes, calls to getTreatment may return the configured fallback treatment (or control if no fallback is set).

Subscribe to SDK_READY before evaluating. Once the listener fires, getTreatment returns the latest result for the requested flag and target.

client.addEventListener(object : SplitEventListener() {
    override fun onReadyView(client: SplitClient, metadata: SdkReadyMetadata?) {
        when (client.getTreatment("FEATURE_FLAG_NAME").treatment) {
            "on" -> {
                // insert code here to show on treatment
            }
            "off" -> {
                // insert code here to show off treatment
            }
            else -> {
                // insert your control treatment code here
            }
        }
    }
})

Using attributes in flag evaluations​

Attributes live on the Target, not on individual evaluation calls. When you instantiate the SDK you provide a Target that includes the matching key, traffic type, and any attributes used for custom-attribute targeting. To change them later, call setTarget on the same client. If you need to evaluate for an entirely different identity in parallel, instantiate an additional client via getClient instead.

Whenever the matching key, bucketing key, or attributes change, the SDK refetches evaluations from FME cloud for the updated Target.

Valid attribute types are strings, numbers, booleans, and collections of those. null entries are dropped before the request is sent. Dates are expected as milliseconds since epoch (UTC); pass System.currentTimeMillis() or a precomputed Long.

val target = Target(
    key = Key(matchingKey = "CUSTOMER_ID"),
    trafficType = "user",
    attributes = mapOf(
        "plan_type" to "growth",
        "deal_size" to 10000,
        "paying_customer" to true,
        "permissions" to listOf("read", "write"),
        "registered_date" to System.currentTimeMillis(),
    ),
)

client.setTarget(target)

Multiple evaluations at once​

To evaluate several flags in a single call, use one of the variations of getTreatments:

  • getTreatments(flags) returns a result for each flag name you pass in.
  • getTreatmentsByFlagSets(flagSets) returns a result for every flag in the listed flag sets currently cached in the SDK.

Both methods return a List<EvaluationResult>.

// Get treatments for a list of flag names
val flagNames = listOf("FEATURE_FLAG_NAME_1", "FEATURE_FLAG_NAME_2")
val results = client.getTreatments(flagNames)

// Get treatments for the union of multiple flag sets
val flagSets = listOf("frontend", "client_side")
val bySets = client.getTreatmentsByFlagSets(flagSets)

// results / bySets have the following form:
// [
//   EvaluationResult(flag = "FEATURE_FLAG_NAME_1", treatment = "on", changeNumber = 1700000000000),
//   EvaluationResult(flag = "FEATURE_FLAG_NAME_2", treatment = "visa", changeNumber = 1700000000000),
// ]

Evaluation result​

Every evaluation method returns an EvaluationResult (or a List of them). The shape is:

data class EvaluationResult(
    val flag: String,         // the flag name evaluated
    val treatment: String,    // 'on', 'off', a custom treatment, or 'control'
    val config: String? = null,       // stringified JSON of the dynamic configuration, when present
    val changeNumber: Long? = null,   // version of the flag definition this result was produced from
    val flagSets: Set<String> = emptySet(), // the flag sets the flag belongs to
)

If a flag cannot be evaluated, treatment is the fallback treatment (or "control" if no fallback is set).

Dynamic configurations are opt-in. By default the SDK does not request them, so result.config is null and no configuration data is exposed in the response. Set configsEnabled = true at factory init to receive them. When enabled, config is a stringified JSON you parse with your library of choice before reading values.

val result = client.getTreatment("FEATURE_FLAG_NAME")
val config = result.config?.let { JSONObject(it) }

// result has the following form:
// EvaluationResult(
//   flag = "FEATURE_FLAG_NAME",
//   treatment = "on",
//   config = "{\"color\":\"red\",\"size\":\"L\"}",
//   changeNumber = 1700000000000,
//   flagSets = setOf("frontend", "client_side")
// )

Update the target's attributes or key​

Use setTarget to update the active target on an existing client when the session's identity or attributes change (the user logs in, switches accounts, an attribute flips). To run two identities in parallel during the same session, instantiate a second client via getClient instead.

The call returns synchronously. Any change to the matching key, bucketing key, or attributes triggers a background refetch of evaluations from FME cloud. Reads issued before the refetch lands operate on an empty cache for the new target, so getTreatment returns the configured fallback treatment for each flag (or control if no fallback is set). A change that only updates trafficType keeps the eval cache and skips the refetch.

Once the refetch lands, the SDK fires SDK_UPDATE on the same client. Subscribe to that event to refresh any UI that depends on flag values.

If the input is not a valid target, setTarget is a safe no-op: the current target is kept and the call does not throw.

client.setTarget(
    Target(
        key = Key(matchingKey = "user-456"),
        trafficType = "user",
        attributes = mapOf("plan_type" to "enterprise", "deal_size" to 50000),
    )
)

Shutdown​

The SDK exposes two lifecycle methods on the client:

  • flush() posts any tracked events and telemetry that are still buffered locally, without tearing down the client. Use it before a user-facing transition where you want pending data to land on FME cloud even if the client keeps running.
  • destroy() flushes pending data and tears down this client instance: it stops background sync, closes connections, releases resources, and removes the client from its parent factory. After destroy(), getTreatment* calls return the configured fallback (or control if none is set), track returns false, and manager.flagNames returns an empty list on a fully-torn-down factory.

SplitFactory also exposes destroy(), which tears down the factory and every client it created. There is no factory.flush(); flush at the client level.

All three methods are Kotlin suspend functions, so call them from a coroutine. From Java, the same operations are available as deprecated *Async(callback) overloads (flushAsync, destroyAsync).

import kotlinx.coroutines.runBlocking

runBlocking {
    client.flush()
    client.destroy()
    factory.destroy()
}

Track​

Use the track method to record any actions your customers perform. Each action is an event and corresponds to an event type. Calling track is the first step to getting experimentation data into Harness FME and lets you measure the impact of your feature flags on user actions and metrics. Learn more about using track events.

The thin client takes the matching key and traffic type from the Target bound to the client, so the track call itself takes only three arguments:

  • eventType: (String, required) The event type to record. Full requirements on this argument are:
    • Contains 80 characters or fewer.
    • Starts with a letter or number.
    • Contains only letters, numbers, hyphen, underscore, period, or colon.
    • This is the regular expression used to validate the value: ^[a-zA-Z0-9][-_.:a-zA-Z0-9]{0,79}$
  • value: (Optional) The value used when creating the metric. Pass null or 0 to use count-based metrics. The expected data type is Double.
  • properties: (Optional) A map of key-value pairs used to filter your metrics. Learn more about event property capture in the Events guide. FME supports three types of property values:
    • Strings: Use type String.
    • Numbers: Use type Number.
    • Booleans: Use type Boolean.

track returns a boolean indicating whether the event was successfully queued to be sent to Harness on the next flush. It returns false if the input failed validation or if the client has been destroyed. The thin client does not validate the traffic-type value against your FME project; an unknown traffic type still queues the event but the backend rejects it on ingest.

// Event type only
var queued: Boolean = client.track("page_load")

// Event type and value
queued = client.track("page_load_time", 83.334)

// Event type, value, and properties
val properties = mapOf(
    "package" to "premium",
    "admin" to true,
    "discount" to 50,
)
queued = client.track("checkout", 49.99, properties)

// Event type and properties only (skip value with a named argument)
queued = client.track("checkout", properties = properties)

Configuration​

The SDK has a small set of configuration knobs, each tuned to a reasonable default. Pass an optional SplitClientConfig as the fourth argument to SplitFactoryBuilder.build to override any of them.

Configuration is grouped into a few top-level fields alongside sdkKey and defaultTarget:

  • sync controls how the SDK keeps evaluations fresh (mode, polling, push rate, ready timeout, url overrides).
  • filters limits which flags are evaluated and synced.
  • storage configures the on-device cache. See Configure cache behavior.
  • fallbackTreatments defines what to return when the SDK cannot evaluate a flag. See Configure fallback treatments.
  • logLevel and configsEnabled are top-level flags.
ConfigurationDescriptionDefault value
logLevelSDK log level. One of NONE, ERROR, WARN, INFO, DEBUG, VERBOSE.NONE
configsEnabledWhether the SDK requests dynamic configurations along with treatments. Off by default to minimize payload; enable it if you read EvaluationResult.config.false
sync.modeSynchronization mode. STREAMING keeps a persistent connection open and applies updates as they arrive, falling back to polling when streaming is unavailable; POLLING refetches on a fixed interval; SINGLE_SYNC fetches once at startup and stops.STREAMING
sync.pollingRatePolling interval in seconds for the polling mode (and the fallback path when streaming is unavailable). Minimum 60.3600 (1 hour)
sync.pushRateInterval in seconds for posting tracked events and telemetry. Minimum 30.1800 (30 minutes)
sync.readyTimeoutTime in seconds to wait before firing SDK_READY_TIMEOUT. Set to -1 to disable the timeout.10
sync.serviceEndpointsUrl overrides for SDK services. Provide hosts only; the SDK appends paths internally. Supported keys: auth, evaluations, events, streaming.Default Harness FME hosts
storage.prefixPrefix prepended to the SDK's on-device database name. Useful for isolating storage between SDK instances sharing the same process. Must match ^[a-zA-Z0-9_]{1,80}$.null (no prefix)
filters.flagSetsRestrict evaluation and sync to flags that belong to one or more flag sets.null (all flags)
fallbackTreatmentsTreatments returned when the SDK cannot evaluate a flag (not ready, flag not found, parse error). See Configure fallback treatments.null

To set each of the parameters defined above, use the following syntax. The Kotlin DSL and the Java builder are equivalent; pick whichever fits your codebase.

import io.split.client.thin.FallbackTreatment
import io.split.client.thin.FallbackTreatmentsConfiguration
import io.split.client.thin.SplitClientConfig
import io.split.client.thin.splitClientConfig

val config = splitClientConfig {
    logLevel = SplitClientConfig.LogLevel.INFO
    configsEnabled = false

    sync {
        mode = SplitClientConfig.SyncMode.STREAMING
    }

    filters {
        flagSets = setOf("frontend")
    }

    fallbackTreatments {
        global(FallbackTreatment("off"))
        byFlag(mapOf("my-feature" to FallbackTreatment("on", "{\"color\":\"blue\"}")))
    }
}

Sync modes​

The SDK supports three synchronization modes, controlled by sync.mode. Pick the one that matches your runtime and freshness requirements.

  • STREAMING (default). The SDK keeps a long-lived connection open to FME cloud and applies rollout plan changes the moment they are published. Best for long-running app sessions where flag changes should take effect without a restart. If the streaming connection cannot be established or drops mid-session (network issues, restrictive proxies, captive networks), the SDK transparently falls back to polling at sync.pollingRate and retries streaming in the background, keeping flags in sync.

  • POLLING. The SDK refetches evaluations on a fixed interval set by sync.pollingRate (default 3600 seconds, minimum 60 seconds). Use this when streaming is blocked at the network layer, or when you want predictable, periodic refreshes instead of streaming push updates.

  • SINGLE_SYNC. The SDK fetches evaluations once at startup, fires SDK_READY, and stops syncing. Use this when you want a one-shot evaluation pass with no background updates over the client's lifetime.

val config = splitClientConfig {
    sync {
        mode = SplitClientConfig.SyncMode.POLLING
        pollingRate = 120
    }
}

Regardless of the mode, tracked events and telemetry are flushed on a separate cadence controlled by sync.pushRate.

Configure cache behavior​

The SDK persists evaluations on device so a fresh app launch can serve cached values immediately while the SDK refetches in the background. Storage is backed by a Room (SQLite) database the SDK manages internally; you do not need to configure a backend.

When persistent storage finds evaluations from a previous session, the SDK fires SDK_READY_FROM_CACHE so you can begin evaluating instantly against cached values while the SDK fetches fresh evaluations from FME cloud in the background.

The one user-facing knob is storage.prefix, prepended to the SDK's on-device database name. Use it to keep multiple SDK keys or apps isolated within the same process. The prefix must match ^[a-zA-Z0-9_]{1,80}$.

val config = splitClientConfig {
    storage {
        prefix = "my_app"
    }
}

The SDK pauses background sync automatically when the app goes to the background (via ProcessLifecycleOwner) and resumes when the app returns to the foreground. No additional configuration is required.

Manager​

Use the Manager to get the list of feature flag names the SDK currently has loaded. The thin Manager exposes a single accessor, flagNames; methods that return full flag definitions (split(name), splits()) are not part of the thin client surface. Get the Manager from the same factory you used for your client:

val manager: SplitManager = factory.getManager()

val flagNames: List<String> = manager.flagNames

flagNames returns the flag names currently cached on the SDK. Before the client is ready, the list is empty; wait for SDK_READY before calling, see Subscribe to events.

If you set filters.flagSets at factory init, flagNames returns only the names in those sets, since flags outside the configured sets are never fetched.

Logging​

The SDK writes log messages to Logcat with a SplitSDK tag, routing each level to its matching android.util.Log method (Log.e, Log.w, Log.i, Log.d, Log.v). Logging is off by default; turn it on by setting logLevel at factory init.

logLevel accepts the following values, in order of increasing verbosity:

  • NONE (default). No logs.
  • ERROR. Failures only, such as network errors and validation rejections.
  • WARN. Warnings and errors. Typically covers non-fatal issues like invalid input that the SDK accepted as a no-op or sanitized.
  • INFO. High-level lifecycle events such as factory init and readiness transitions, plus everything above.
  • DEBUG. Internal milestones useful when diagnosing a specific behavior, plus everything above.
  • VERBOSE. Maximum detail. Useful when assembling repro information for support; expect a high volume of Logcat output.
val config = splitClientConfig {
    logLevel = SplitClientConfig.LogLevel.DEBUG
}

Configure fallback treatments​

Fallback treatments let you define a treatment value (and optional dynamic configuration) to be returned when a flag cannot be evaluated, for example when the SDK is not ready, the flag does not exist, or a transient error occurs. By default, the SDK returns the reserved control treatment, but you can override this globally and/or per flag at the SDK level.

This is useful when you want to:

  • Maintain a predictable user experience during outages or evaluation failures (avoid unexpected control when not even local cache exists)
  • Protect critical user flows by returning a safe, stable treatment (for example, forcing off during an incident)
  • Customize behavior per flag so each evaluation inherits appropriate safe defaults if something goes wrong

Each entry, global or per flag, can be either a FallbackTreatment(treatment) or a FallbackTreatment(treatment, config), where config is a JSON string available on the evaluation result when configsEnabled is on. The builder also exposes global(String) and byFlagStrings(Map<String, String>) convenience overloads for the bare-string case.

Global fallback treatment​

Set a global fallback treatment when initializing the SDK factory. This value is returned whenever any flag cannot be evaluated.

val config = splitClientConfig {
    fallbackTreatments {
        global(FallbackTreatment("off", "{\"reason\":\"global-fallback\"}"))
    }
}

Flag-level fallback treatment​

You can also configure fallback treatments per flag, alongside global. A per-flag entry always takes precedence over the global fallback, so if both are defined the SDK returns the flag-level value when that specific flag cannot be evaluated.

val config = splitClientConfig {
    fallbackTreatments {
        global(FallbackTreatment("off"))
        byFlag(
            mapOf(
                "checkout_v2" to FallbackTreatment("control_v1", "{\"reason\":\"flag-level-fallback\"}"),
                "new_dashboard" to FallbackTreatment("off"),
            )
        )
    }
}

For more information, see Fallback treatments.

Advanced use cases​

This section describes advanced use cases and features provided by the SDK.

Instantiate multiple clients​

Use a second client when your app needs to evaluate for two identities at the same time, for example a logged-in user and the account that user belongs to, or two users on the same device. Each client is bound to its own target and runs its own readiness lifecycle, so the SDK can return treatments for both identities in parallel without one overwriting the other.

Call factory.getClient() with no arguments to get the client bound to the factory's default target. Pass a target to create or retrieve a client for a different identity.

The factory keys clients by matching key (and bucketing key when set). Calling getClient twice with the same key returns the same client; if the second call passes a different trafficType or different attributes, the SDK applies the change in the background via setTarget rather than creating a duplicate client.

Each client manages its own readiness, events, and pending writes. Wait for or check readiness on the specific client before evaluating; see Subscribe to events for the event listeners (useful when the client may already be ready from a prior getClient call).

val factory = SplitFactoryBuilder.build(
    applicationContext,
    "YOUR_SDK_KEY",
    Target(Key("CUSTOMER_USER_ID"), trafficType = "user"),
)

// Default target client (bound to CUSTOMER_USER_ID / "user")
val userClient = factory.getClient()

// Second client for the account that owns this user
val accountClient = factory.getClient(
    Target(Key("CUSTOMER_ACCOUNT_ID"), trafficType = "account"),
)

val userPoll = userClient.getTreatment("user-poll")
val accountPerms = accountClient.getTreatment("account-permissioning")

accountClient.track("account_created")

When an identity is no longer in scope, destroy its client to free its readiness listeners and stop syncing for that key. See Shutdown for the call shape: client.destroy() removes that one client from the factory's registry; factory.destroy() tears down the factory and every client it created. For sequential identity changes on a single client, prefer setTarget over destroying and recreating.

Subscribe to events​

You can listen for four different events from the SDK.

  • SDK_READY_FROM_CACHE. Fires when the SDK can serve cached evaluations from a previous session. If a cache from a previous session is present, this event fires almost immediately with stale-but-usable results. If no cache is available, the event still fires, together with SDK_READY, signaling that there were no cached results to load.
  • SDK_READY. Fires once the SDK has fetched a fresh set of evaluations for the current target from FME cloud and is ready to serve them.
  • SDK_READY_TIMEOUT. Fires if the SDK has not reached the SDK_READY state within the sync.readyTimeout window (10 seconds by default; set to -1 to disable). The SDK continues to fetch in the background and emits SDK_READY once it succeeds; calls made before then return the configured fallback treatment, or control if no fallback is set.
  • SDK_UPDATE. Fires when the SDK applies an update to its evaluations (for example, after a streaming push, a poll cycle, or a target switch).

Listeners are registered with client.addEventListener(listener). A listener extends SplitEventListener and overrides only the callbacks it needs. Each event exposes two callbacks: the base callback (onReady, onReadyFromCache, onUpdate, onTimeout) runs on a background thread, and the *View variant (onReadyView, onReadyFromCacheView, onUpdateView, onTimeoutView) runs on the main UI thread.

Pick the right thread

Touch UI from *View callbacks. Do work that may block (network calls, disk I/O, heavy computation) from the non-View callbacks. Blocking the main thread from a *View callback will jank or freeze your app.

import io.split.client.thin.SdkReadyMetadata
import io.split.client.thin.SdkUpdateMetadata
import io.split.client.thin.SplitClient
import io.split.client.thin.SplitEventListener

client.addEventListener(object : SplitEventListener() {
    override fun onReady(client: SplitClient, metadata: SdkReadyMetadata?) {
        // Fresh evaluations from FME cloud are available; safe to evaluate.
    }

    override fun onReadyFromCache(client: SplitClient, metadata: SdkReadyMetadata?) {
        // Cached evaluations restored from a previous session; results may be stale.
    }

    override fun onUpdateView(client: SplitClient, metadata: SdkUpdateMetadata?) {
        // Evaluations changed; re-evaluate the impacted flags and refresh affected UI.
        val updatedFlags = metadata?.names ?: return
        val results = client.getTreatments(updatedFlags)
    }

    override fun onTimeout(client: SplitClient) {
        // SDK did not become ready within sync.readyTimeout.
        // Calls return the configured fallback treatment (or `control`
        // if no fallback is set) until ready.
    }
})

Include metadata​

metadata provides additional context for events:

  • onReady / onReadyFromCache (and their *View variants): SdkReadyMetadata with isInitialCacheLoad (true when there were no cached results to load and the event was synthesized alongside SDK_READY) and lastUpdateTimestamp (milliseconds since epoch when the cache was last updated; null when isInitialCacheLoad is true).
  • onUpdate / onUpdateView: SdkUpdateMetadata with type (FLAGS_UPDATE or SEGMENTS_UPDATE) and names (list of impacted flag names; empty for segment-only updates).
client.addEventListener(object : SplitEventListener() {
    override fun onReady(client: SplitClient, metadata: SdkReadyMetadata?) {
        val initialCacheLoad: Boolean? = metadata?.isInitialCacheLoad
        val lastUpdateTimestamp: Long? = metadata?.lastUpdateTimestamp
    }

    override fun onUpdate(client: SplitClient, metadata: SdkUpdateMetadata?) {
        val type: SdkUpdateMetadata.Type? = metadata?.type
        val names: List<String>? = metadata?.names
    }
})