Evaluation Algorithm

When your application calls /sdk/v1/evaluate, the service runs this algorithm for every flag in the requested environment. It is deterministic. The same user and flag always produce the same result given the same configuration.

The algorithm

Step by step

Step 1: Global switch

If enabled = false, return false immediately. No rules or rollout logic runs.

Step 2: Targeting rules

Active rules are sorted by priority descending. Each rule is evaluated against the user context:

• user_id — exact match against context.user_id
• user_email — exact match against context.user_email
• email_domain — suffix match against context.user_email

The first matching rule returns true. If no rules match, continue.

Step 3: Rollout percentage

The user is assigned to a bucket using FNV-1a hash of "<flag_key>:<user_identifier>". The result is mapped to a number from 0 to 99.

• If the bucket number is less than the rollout percentage, return true
• Otherwise return false

The user_identifier is user_id if provided, otherwise user_email. The same user always lands in the same bucket across requests, servers, and restarts.

Step 4: Default

If enabled = true and rollout_percentage is 0 (or not set), the rollout block is skipped entirely and the flag returns true for everyone with reason "Flag enabled globally". To disable a flag for all users, set enabled to false.

Why FNV-1a

FNV-1a is fast, deterministic, and produces a uniform distribution across the 0–99 bucket range. It requires no external state. The hash is seeded with the flag key, so the same user lands in a different bucket for different flags. Gradual rollouts across multiple flags are independent of each other.

The reason field

Every evaluation result includes a reason string:

Result	Reason string
Global switch off	"Flag is globally disabled"
Rule match	"Matched user_id targeting rule" / "Matched email_domain targeting rule" etc.
In rollout bucket	"User in 20% rollout"
Outside rollout bucket	"User not in 20% rollout"
Enabled, no rollout	"Flag enabled globally"

The reason is for debugging only. Branch on enabled, not reason.

Evaluation latency

Evaluation is typically under 10ms end-to-end. The service fetches all flags for the environment in one query and all rules in a second batch query. Evaluation logic runs in memory. Evaluation logs are written asynchronously and never block the response.