MIRAGE: LLMs Can Replace Real Microservices in Integration Tests with 99% Fidelity — No Mocks Required

Traditional microservice testing requires complex mock setups. MIRAGE uses LLMs to simulate dependent services at runtime with 99% fidelity — no pre-generated mocks needed.

The Problem

Testing microservices that depend on other services is hard:

• Record-replay — Static, breaks when APIs change (62% status code fidelity)
• Pattern mining — Limited to observed patterns
• Specification-driven stubs — Manual, incomplete, brittle

MIRAGE's Approach

Instead of pre-generating mocks, MIRAGE uses an LLM at runtime to answer each dependency request as it arrives:

1. Reads dependency source code, caller code, and production traces
2. Simulates dependency behavior on demand
3. Maintains cross-request state throughout the test scenario

Results

Tested Systems

• Google's Online Boutique
• Weaveworks' Sock Shop
• Custom microservice system
• 110 test scenarios across 14 caller-dependency pairs

Key Insight

"Dependency source code is often sufficient for high-fidelity runtime simulation, even without production traces."

Why It Matters

• Eliminates mock maintenance — No more updating stubs when APIs change
• Faster development — Test against services without deploying them
• CI/CD improvement — Integration tests become practical in pipelines
• Cost reduction — No need to spin up full dependency stacks for testing