Skip to content
Wovyr

GA Completion: Scale & Performance Validation

Document ID: GA-001 File Path: docs/18-roadmap/v1.0/A1-scale-performance.md Version: 1.0.0 Status: Planned — GA-completion work, not started Owner: Platform Team Last Updated: 2026-07-05

1. Purpose

Section titled “1. Purpose”

Turn the “Scale & Performance Validation” GA gap (PRD-002 §5.1, v1.0 §3 Scale row) into a concrete delivery plan: current state, work breakdown, exit criteria, and the environment dependencies that block validation today.

Unlike the Tier B research bets, this is committed intent — scoped, near-term, GA-blocking work, not an exploratory bet.

2. Current State

Section titled “2. Current State”
  • Perf tests exist but are deliberately toy-scale. Assertion-style p95 gates run under ordinary cargo test: crates/wovyr-provider/tests/perf.rs (gateway overhead) and crates/wovyr-memory/tests/perf.rs (warm retrieval), on the order of hundreds of records against the in-process mock provider. Large headroom keeps them stable in CI — they prove no regression, not scale.
  • The horizontal-scaling mechanics are built. Workflow queue partitioning (G6) shards executions across disjoint worker pools, and a scaling envelope with measured single-node baselines is published in distributed-execution §3.3.
  • The tiered memory backend exists (Postgres system-of-record + Qdrant ANN, behind the tiered feature) with capability-gated integration tests.

3. Gap

Section titled “3. Gap”

The NFR targets — billions of memories, thousands of concurrent runs (performance-tests) — are unvalidated against real capacity. The mechanics are in place; the proof at scale is not.

4. Scope & Requirements

Section titled “4. Scope & Requirements”

4.1 Functional / deliverables

Section titled “4.1 Functional / deliverables”
  • A reproducible load-generation harness (memory ingest/query, workflow submit/drive) parameterized by cardinality and concurrency.
  • Memory scale validation at target cardinality against a live TieredStore, publishing p50/p95/p99 retrieval latency and ingest throughput.
  • Workflow throughput validation across a real multi-worker pool using the existing partitioning (G6), under sustained load.
  • An updated, honest scaling envelope extending distributed-execution §3.3.

4.2 Non-functional

Section titled “4.2 Non-functional”
  • Measurements are reproducible (fixed dataset generators, recorded topology).
  • Results are published with methodology, not just headline numbers.

5. Exit Criteria

Section titled “5. Exit Criteria”

Documented, reproduced NFR numbers against live Postgres/Qdrant — or an honest, published statement of where the current architecture tops out and why (a real ceiling is an acceptable, GA-worthy outcome; an unmeasured claim is not).

This feeds the v1.0 exit criterion “meets published SLOs in production” (v1.0 §5).

6. Dependencies & Environment Caveats

Section titled “6. Dependencies & Environment Caveats”
  • Requires real cloud capacity (managed Postgres + Qdrant at scale) — absent in the current dev environment, so this cannot be validated in-house today. This is the primary blocker, and the reason the work is planned not in progress.
  • Should follow, or run alongside, an ADR recording the target topology and measurement method (per the graduation flow).

7. Risks

Section titled “7. Risks”
RiskMitigation
Numbers unverifiable in-houseFlag the environment dependency explicitly; never claim what wasn’t run
A real ceiling below targetTreat an honest, published ceiling as a valid outcome, not a failure to hide
Benchmark driftReproducible generators + recorded topology
Section titled “8. Related Documents”

9. Revision History

Section titled “9. Revision History”
VersionDateDescription
1.0.02026-07-05Initial GA-completion delivery doc for scale & performance validation