Preparing Assignment Source Tables for Warehouse Native Experimentation

Learn how to prepare your assignment source tables in your data warehouse for Warehouse Native Experimentation.

To prepare an Assignment Source for Warehouse Native Experimentation, transform your raw exposure or impression logs into a clean, standardized table that serves as the foundation for experimentation analyses.

This page describes the required fields, recommended fields, and best practices for preparing your assignment source tables.

Required columns

These fields are mandatory. Without them, Harness FME cannot map exposures to experiment results.

Every Assignment Source table must include the following columns:

Column	Type	Description
Unique Key	STRING	Unique identifier for the unit of randomization (for example, user_id, account_id, or a custom key). Must be stable across the experiment duration.
Exposure Timestamp	DATETIME / TIMESTAMP	The time the assignment occurred (for example, when an impression was logged, a flag evaluated, or getTreatment was called). Must be provided in epoch milliseconds or UTC format (e.g. 1714953600000 or 2026-05-05T00:00:00Z). Timestamps must be in UTC (epoch milliseconds or UTC datetime) and align with the format used in Metric Sources for accurate joins and analysis.
Treatment (Variant Group)	STRING	The assigned experiment variant (for example, control, treatment_a, variant_1).

Recommended columns

While not required, the following fields make debugging, filtering, and governance more efficient.

Column	Type	Description
Experiment ID / Name	STRING	Helps differentiate exposures when multiple experiments are logged in the same raw table.
Targeting Rule	STRING	Indicates which targeting rule or condition led to the assignment. Useful for audit and debugging. If you are using FME feature flag impressions, filter by a single targeting rule to ensure the experiment analyzes the intended population.
Environment ID	STRING	Allows filtering by environment (for example, production, staging). When configuring an assignment source in FME, you can map column values to a matching Harness environment or hard-code a single environment. When creating an experiment, it must be scoped to one environment.
Traffic Type	STRING	Distinguishes the unit type (for example, user, account, anonymous visitor). When configuring an assignment source, you can map column values or hard-code the environment. Each experiment must be scoped to one traffic type.

Common raw table schemas

Most organizations log impressions or exposures from feature flag evaluations, SDKs, or event pipelines. Below are common raw schemas and how to normalize them.

Feature Flag Evaluation Logs

Example Raw Schema	Transformations
user_id flag_name treatment impression_time environment rule_id	• Map flag_name values → experiment_id (if multiple flags correspond to the same experiment). • Cast evaluation_time to TIMESTAMP. • Deduplicate on (user_id, experiment_id) by keeping the earliest exposure.

A/B Test Impression Logs

Example Raw Schema	Transformations
experiment_id user_id bucket or arm impression_time	• Standardize bucket → treatment. • Standardize impression_time → exposure_timestamp. • Deduplicate to keep only the first exposure per user per experiment.

Event Logging Pipelines (Custom Analytics Events)

Example Raw Schema	Transformations
event_name event_time properties.experiment_id properties.variant properties.user_id	• Flatten nested fields (JSON → explicit columns). • Filter to only event_name = 'experiment_exposure'. • Standardize column names to match required schema.

Prepare your assignment table

Follow these best practices for preparing your assignment table in your data warehouse.

• De-duplication: Keep only the earliest exposure per user per experiment. For example:

  QUALIFY ROW_NUMBER() OVER (
    PARTITION BY user_id, experiment_id
    ORDER BY exposure_timestamp ASC
  ) = 1

• Consistent Variant Labels: Standardize variant naming (control, treatment, variant_1) across experiments. Avoid null or empty strings; default to control if needed.

• Timestamp Format: All exposure timestamps must be in epoch milliseconds or UTC (ISO 8601).

• Use UTC Consistently: Store all exposure and event timestamps in UTC to ensure consistent comparisons across regions.

• Stable Identifiers: Use the same user or account key across Assignment Source and Metric Source tables. If your system logs multiple IDs (for example, cookie_id and user_id), choose the most stable one.

• Environment Separation: If raw tables mix environments (for example, staging and production), add an environment_id column and filter accordingly. This prevents accidental inclusion of test data in production environments.

• Partitioning and Indexing: Partition large tables by DATE(exposure_timestamp) to optimize query performance. Cluster or index by experiment_id and user_id for faster lookups.

Example prepared table schema

Column	Type	Example
user_id	STRING	abc123
experiment_id	STRING	checkout_flow_v2
treatment	STRING	control
exposure_timestamp	TIMESTAMP	2025-03-14T12:45:00Z
environment_id	STRING	prod
traffic_type	STRING	user

Once your Assignment Source tables are prepared and validated, see Setting Up an Assignment Source to connect them in Harness FME.