Skip to content

Local Evaluation

Server-side local evaluation runs evaluation logic on your server, saving you the overhead incurred by making a network request per user evaluation. The sub-millisecond evaluation is perfect for latency-minded systems which need to be performant at scale.

Client-side local evaluation experimentation diagram.

Exposure Tracking

Local evaluation doesn't automatically set experiment user properties. If you use local evaluation and you want to run experiments where success metrics are analyzed, you will need to implement exposure tracking (generally done on the client-side).

To more easily track exposures on the client-side, bootstrap the client-side SDK with the variants evaluated server-side and utilize automatic exposure tracking using one of the analytics SDK integrations.

Targeting Capabilities

Because local evaluation happens outside of Amplitude, advanced targeting and identity resolution powered by Amplitude Analytics isn't supported. That said, local evaluation allows you to perform consistent bucketing with target segments, which is often sufficient.

Feature
Remote Evaluation Local Evaluation
Consistent bucketing
Individual inclusions
Targeting segments
Amplitude ID resolution
User enrichment
Sticky bucketing

Implementation

Local evaluation is just evaluation--a function which takes a user and a flag as input, and outputs a variant.

Diagram of a local evaluation SDK.

The only non-local part of local evaluation is getting flag configurations from Amplitude Experiment, but this can happen at an interval, and flags can be cached in-memory on the server-side for zero latency access.

Edge Evaluation

The local evaluation Node.js SDK can be run in edge worker/functions which support JavaScript and a distributed store. Contact your representative or email experiment@amplitude.com to learn more.

SDKs

Local evaluation is only supported by server-side SDKs which have local evaluation implemented.

SDK Remote Evaluation Local Evaluation
Node.js
Ruby
JVM
Go
Python

Performance

The following results are for a single flag evaluation, and were collected over 10 executions of 10,000 iterations of evaluation with randomized user inputs evaluated for 1 flag configuration, selected at random out of 3 possible flag configurations.

SDK Average Median Cold Start
Node.js 0.025ms 0.018ms 3ms
Go 0.098ms 0.071ms 0.7ms
JVM 0.007ms 0.005ms 6ms

Still have questions? Ask them in the Community.