feat: add bandits support to precomputed client by sameerank · Pull Request #98 · Eppo-exp/eppo-ios-sdk

sameerank · 2026-03-12T06:31:53Z

🎟️ Fixes N/A - New feature
📜 Design Doc: N/A

Motivation and Context

Add contextual multi-armed bandit support to the iOS SDK's EppoPrecomputedClient, achieving feature parity with the Android and JS SDKs.

Description

This PR adds the getBanditAction() method and all supporting infrastructure for bandits in precomputed flags:

New Files:

Sources/eppo/dto/PrecomputedBandit.swift - Wire format and decoded bandit DTOs
Sources/eppo/dto/BanditResult.swift - Public return type with variation and optional action
Sources/eppo/dto/BanditEvent.swift - Logging model and BanditLogger type alias

Modified Files:

Precompute.swift - Added banditActions parameter for specifying available actions per flag
PrecomputedConfiguration.swift - Added bandits parsing and Base64 decoding
PrecomputedRequestor.swift - Added bandit_actions to request payload and bandits to response handling
PrecomputedConfigurationStore.swift - Added getDecodedBandit() method
EppoPrecomputedClient.swift - Added getBanditAction(), banditLogger, and deduplication via banditAssignmentCache

Key Features:

getBanditAction(flagKey:defaultValue:) returns BanditResult with variation and optional action
Bandits keyed by MD5-hashed flag key (matching existing flag lookup pattern)
Base64 decoding for obfuscated bandit data (banditKey, action, modelVersion, attributes)
BanditEvent logging with deduplication
Falls back to regular string assignment when no bandit exists for a flag
Thread-safe concurrent access

How has this been documented?

No external documentation changes required. The public API is straightforward:

getBanditAction(flagKey:defaultValue:) - similar to existing assignment methods
banditLogger parameter in initialization - similar to existing assignmentLogger

How has this been tested?

Unit tests: EppoPrecomputedClientBanditTests.swift - 11 tests covering:
- Basic bandit action retrieval
- Non-bandit flag fallback
- Missing flag handling
- Bandit logging with all event fields
- Subject attribute splitting (numeric/categorical)
- Logging deduplication
- Concurrent access safety
- Bandits without actions
Wire format tests: PrecomputedConfigurationWireTests.swift - 3 new tests using shared precomputed-v1.json test data:
- testBanditActionWithSharedTestData
- testBanditActionWithExtraLoggingSharedTestData
- testBanditActionForNonBanditFlagSharedTestData
End-to-end: Verified with EppoButtonClicker test app against live Eppo backend

        let precompute = Precompute(
            subjectKey: "test-user-123",
            subjectAttributes: [
                "country": .valueOf("US"),
                "age": .valueOf(25),
                "isPremium": .valueOf(true)
            ],
            banditActions: [
                "sameeran-test-1": [
                    "action-a": [
                        "price": .valueOf(9.99),
                        "category": .valueOf("electronics")
                    ],
                    "action-b": [
                        "price": .valueOf(19.99),
                        "category": .valueOf("clothing")
                    ],
                    "action-c": [
                        "price": .valueOf(29.99),
                        "category": .valueOf("home")
                    ]
                ]
            ]
        )

banditdemo.mov

All tests passing: swift test ✅

Add getBanditAction() method to EppoPrecomputedClient for contextual multi-armed bandit support. Key changes: - Add PrecomputedBandit DTO with wire format and decoded versions - Add BanditResult return type with variation and optional action - Add BanditEvent for logging bandit assignments - Update Precompute to accept banditActions parameter - Update PrecomputedRequestor to send bandit_actions in payload - Add banditLogger parameter to client initialization - Implement bandit logging with deduplication via banditAssignmentCache - Add comprehensive tests for bandit functionality

Copilot

Pull request overview

Adds contextual multi-armed bandit support to the iOS SDK’s EppoPrecomputedClient so precomputed assignments can return both a variation (model) and a selected action, with optional logging and deduplication.

Changes:

Introduces bandit DTOs (PrecomputedBandit + decoded variant), public API return type (BanditResult), and logging model (BanditEvent/BanditLogger).
Extends the precomputed fetch payload/response to send bandit_actions and parse bandits, including Base64 decoding in PrecomputedConfiguration.
Adds getBanditAction() to EppoPrecomputedClient with bandit-event logging + deduplication, plus comprehensive unit/wire tests.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
Sources/eppo/precomputed/EppoPrecomputedClient.swift	Adds `getBanditAction()` and bandit logging/deduplication infrastructure
Sources/eppo/precomputed/Precompute.swift	Extends `Precompute` to accept optional `banditActions`
Sources/eppo/precomputed/PrecomputedRequestor.swift	Adds `bandit_actions` to request payload and `bandits` to response decoding
Sources/eppo/precomputed/PrecomputedConfiguration.swift	Adds bandits to configuration model and Base64 decoding into decoded config
Sources/eppo/precomputed/PrecomputedConfigurationStore.swift	Adds decoded bandit lookup accessor
Sources/eppo/dto/PrecomputedBandit.swift	New wire-format and decoded bandit DTOs
Sources/eppo/dto/BanditResult.swift	New public return type for bandit action lookups
Sources/eppo/dto/BanditEvent.swift	New public logging event model and logger typealias
Sources/eppo/Constants.swift	Bumps SDK version constant
Tests/eppo/precomputed/EppoPrecomputedClientBanditTests.swift	New unit tests for bandit retrieval, logging, dedupe, concurrency
Tests/eppo/precomputed/PrecomputedConfigurationWireTests.swift	Adds wire-format tests for bandits using shared test data
Tests/eppo/precomputed/PrecomputedRequestorTests.swift	Updates request payload test to include new initializer arg
Tests/eppo/precomputed/PrecomputedTestHelpers.swift	Adds helpers to construct hashed/encoded bandit test fixtures

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-03-23T12:04:54Z

+                    subjectNumericAttributes[key] = doubleValue
+                }
+            } else {
+                if let stringValue = try? value.getStringValue() {


In logBanditAction, categorical subject attributes are extracted using getStringValue(), which throws for boolean values and arrays (and will silently skip them due to try?). Since Precompute.subjectAttributes can include booleans and string arrays (see EppoValue), this will drop valid attributes from bandit logging. Consider using EppoValue.toEppoString() for non-numeric values (and skipping only nulls) so booleans become "true"/"false" and arrays can be joined consistently.

Suggested change

if let stringValue = try? value.getStringValue() {

if let stringValue = try? value.toEppoString() {

aarsilv

Nice job bringing precomputed bandits to iOS 💪 I appreciate the clean data models, easy to understand code, and automated tests. I have mostly minor comments but there is one big one for which I'm requesting changes: we still need to log the "regular" string assignment when getting a bandit action. This is is needed for when A/B testing bandit against status quo.

aarsilv · 2026-03-23T12:03:05Z

+        actionProbability: Double,
+        optimalityGap: Double


Since action is optional, these should be too. If no action, these should be nil

aarsilv · 2026-03-23T12:03:37Z

+        actionNumericAttributes = try container.decodeIfPresent([String: String].self, forKey: .actionNumericAttributes)
+        actionCategoricalAttributes = try container.decodeIfPresent([String: String].self, forKey: .actionCategoricalAttributes)
+        actionProbability = try container.decodeIfPresent(Double.self, forKey: .actionProbability) ?? 0.0
+        optimalityGap = try container.decodeIfPresent(Double.self, forKey: .optimalityGap) ?? 0.0


Defaulting to 0.0 is a bit misleading for actionProbability and incorrect for optimalityGap, can we make these optional and default to nil?

aarsilv · 2026-03-23T12:03:59Z

+    let actionProbability: Double
+    let optimalityGap: Double


aarsilv · 2026-03-23T12:05:11Z

+
+        // First check if there's a bandit assignment for this flag
+        guard let bandit = configurationStore.getDecodedBandit(forKey: hashedFlagKey) else {
+            // No bandit, fall back to regular string assignment


aarsilv · 2026-03-23T12:05:43Z

+            // No action to log
+            return


aarsilv · 2026-03-23T12:13:37Z

+        XCTAssertEqual(result.variation, "hello world")
+        XCTAssertNil(result.action)


aarsilv · 2026-03-23T12:13:44Z

+        XCTAssertEqual(result.variation, "default")
+        XCTAssertNil(result.action)


aarsilv · 2026-03-23T12:15:48Z

+
+    // MARK: - Concurrent Access Tests
+
+    func testConcurrentBanditAccess() throws {


aarsilv · 2026-03-23T14:41:31Z

+        let hashedFlagKey = getMD5Hex(flagKey, salt: decodedConfig.decodedSalt)
+
+        // First check if there's a bandit assignment for this flag
+        guard let bandit = configurationStore.getDecodedBandit(forKey: hashedFlagKey) else {


prior to this, we'll want to log the "regular" assignment as well

aarsilv · 2026-03-23T14:46:51Z

+            return BanditResult(variation: variation, action: nil)
+        }
+
+        // Get the variation from the bandit


thinking we expand this comment to be even more clear, something like "if the bandit was invoked, the variation value is the same as the bandit key"

Then again if we have the variation (which we'll need to log for "regular" assignment log) maybe just use that to set to not confuse people.

sameerank force-pushed the sameerank/bandits-support branch from 0c65905 to 5b4f090 Compare March 12, 2026 06:38

sameerank marked this pull request as ready for review March 12, 2026 06:42

sameerank requested a review from aarsilv March 12, 2026 06:42

sameerank force-pushed the sameerank/bandits-support branch from 66efb74 to 17bdc1e Compare March 12, 2026 23:04

aarsilv requested a review from Copilot March 23, 2026 12:00

Copilot started reviewing on behalf of aarsilv March 23, 2026 12:01 View session

Copilot AI reviewed Mar 23, 2026

View reviewed changes

aarsilv requested changes Mar 23, 2026

View reviewed changes

aarsilv assigned sameerank Mar 23, 2026

	if let stringValue = try? value.getStringValue() {
	if let stringValue = try? value.toEppoString() {

		XCTAssertEqual(result.variation, "hello world")
		XCTAssertNil(result.action)

		XCTAssertEqual(result.variation, "default")
		XCTAssertNil(result.action)


		// MARK: - Concurrent Access Tests

		func testConcurrentBanditAccess() throws {

Conversation

sameerank commented Mar 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation and Context

Description

How has this been documented?

How has this been tested?

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

aarsilv left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

sameerank commented Mar 12, 2026 •

edited

Loading