Skip to content

Performance Testing

Document ID: TEST-004
File Path: docs/15-testing/performance-tests.md
Version: 1.0.0
Status: Draft
Owner: Quality Engineering Team
Last Updated: 2026-06-27


This document defines performance testing for the Wovyr AI Platform — validating that services meet their latency and throughput non-functional requirements under realistic and extreme load.


TypeQuestion
LoadDoes it meet SLOs at expected load?
StressWhere does it break, and how?
SoakStable over hours/days (leaks, drift)?
SpikeDoes it absorb sudden surges?
CapacityHow much per node; how does it scale?

Performance tests assert documented NFRs, e.g.:

MetricTarget
API p95 latency< 200 ms
LLM Gateway overhead (non-cached)< 8 ms p95
Memory warm retrieval< 30 ms p95
Tool warm sandbox start< 20 ms p95
Throughputper-service NFR targets

Sources: LLM Gateway, Memory, Tool Runtime.


Define scenario + load profile
Run against a production-like environment (real backends, fake providers)
Measure p50/p95/p99 latency, throughput, error rate, saturation
Compare to targets + previous baseline (regression check)

Providers are faked with realistic latency so model variance does not skew results; a separate live-provider profile measures real end-to-end latency.


  • Agent runs (mixed tool/model usage)
  • Workflow executions (parallel branches, fan-out)
  • Memory retrieval at scale (millions–billions of records)
  • Tool execution fan-out across worker pools
  • LLM Gateway routing/failover under concurrency

Memory and tool subsystems are tested at platform-scale targets (e.g. billions of memories, thousands of concurrent executions) to validate sharding, indexing, and autoscaling.


Spike tests verify autoscaling reacts within target (e.g. add tool-worker capacity < 30 s) and that fair scheduling holds under contention.


  • Load generators (k6/Gatling/custom Rust harness).
  • Results correlate with metrics/traces captured during the run, using exemplars to find slow paths.

Baselines are stored; nightly/release perf runs fail if p95/throughput regress beyond a threshold versus baseline.




VersionDateDescription
1.0.02026-06-27Initial Performance Testing specification