Milestone 3.9.2 — Context Assembly Load Test and Benchmark

Status: Planned
Goal: 3.9 — Phase 3 quality gate
Phase: 3 — Memory Engine and Operator Platform
Estimated effort: 2 days

Why This Milestone Exists

Phase 3 adds a gRPC call to the proxy's critical path. ARCHITECTURE.md commits to p95 < 50ms for context assembly. Under concurrent load, the context assembly service may queue requests and degrade. This benchmark validates the latency target holds under realistic load.

k6 load test

JavaScript

// benchmarks/k6/context_assembly_load.js
export const options = {
    stages: [
        { duration: "30s", target: 20 },   // ramp to 20 VUs
        { duration: "2m",  target: 50 },   // hold at 50 VUs (realistic load)
        { duration: "30s", target: 0 },    // ramp down
    ],
    thresholds: {
        http_req_duration: ["p(95)<50"],   // p95 < 50ms (context assembly only)
        http_req_failed:   ["rate<0.001"], // error rate < 0.1%
    },
};
 
export default function () {
    // POST to the context assembly gRPC-gateway endpoint
    // (or use grpc-load-test tool for native gRPC benchmarking)
    http.post(`${CONTEXT_URL}/v1/assemble`, JSON.stringify(SAMPLE_REQUEST), {
        headers: { "Content-Type": "application/json" },
    });
}

Acceptance Criteria

p95 context assembly latency < 50ms at 50 concurrent VUs
p99 < 100ms (degraded but not failing)
Error rate < 0.1% (context fallback responses are not errors)
Latency breakdown documented: embedding call, pgvector search, hot cache
Benchmark baseline committed to benchmarks/BASELINE_PHASE3.md