Assurance Report

How this extension is verified, with the actual numbers from the v5.0.0 validation campaign (2026-07-10, local x86-64 Linux, Bun 1.3, containers on bridge network). Every leg below is reproducible from this repository — commands in TESTING.md.

Correctness matrix

The full test suite (303 tests, 100% line coverage of src/) runs against real servers, not mocks:

Backend	Result
Dragonfly (latest)	303 / 303
Redis Stack 7.4	303 / 303
Redis 8.8 (native JSON)	303 / 303
@prisma/client peer floor (7.2.0, CI leg)	303 / 303

Plus, on every CI run: a node:test smoke suite against the built dist (CJS + ESM), compile-time type tests (tsc over the public surface, including the $transaction inference guard and rejected-input negatives), biome lint, publint, arethetypeswrong (node10 / node16-CJS / node16-ESM / bundler all green), and bun audit (zero findings). Releases are built on CI with npm provenance attestations.

Load

test/stress/load.ts — 50 bursts × 5,000 concurrent requests over 500 keys with 2s TTL / 1s stale windows:

Metric	Result
Requests	250,000 — zero failures
Database executions	2,500 (100:1 coalescing; ceiling was 25,500)
Throughput	~30,900 req/s sustained
Latency	p50 2.8ms · p95 6.7ms · p99 101ms (burst-start fills)
Heap	2MB → 5MB across the run

Soak (1 hour)

test/stress/soak.ts — 200 concurrent readers on 2s/1s TTL windows (constant fresh → stale → refresh churn), a writer overwriting hot keys, and a pattern-invalidation storm every 15s, sustained for 60 minutes:

Metric	Result
Requests	55,334,526 (~15.4k req/s sustained) — zero failures
Database executions	543,254 (~1 per key per TTL window throughout)
Invalidation storms	239 pattern sweeps, 7,199 entries removed mid-traffic
Heap	oscillated 7–18MB across the hour and ended below the post-warmup baseline (40MB → 3MB after final GC) — no growth trend, no leak

Chaos (dependency failure)

test/stress/chaos.ts — Redis is killed mid-traffic, then restarted:

• Outage: 100/100 requests still resolved (served from the database); healthCheck() reported unhealthy; every cache error surfaced through onError (201 counted), none thrown at callers.
• Recovery: after restart, 100/100 resolved, caching resumed, health returned to healthy. No process restart, no manual intervention.

Eviction pressure

test/stress/eviction.ts — 2,400 × 64KB entries written into a 64MB allkeys-lru cap (Redis Stack): the server evicted 1,603 entries underneath the extension; all 2,400 subsequent reads returned correct values (121 from cache, 2,279 as clean database misses that re-cached) — zero failures.

JSON-less server (Redis ≤ 7)

test/stress/plain-redis.ts against Redis 7.4 (no RedisJSON module):

• The startup probe detects the missing module and announces the remedy (console + onError).
• With type: 'JSON' misconfigured anyway: 50/50 queries still served correctly from the database; the actionable hint is attached to the first error exactly once.
• With type: 'STRING' on the same server: full caching (1 database call for 50 reads) and pattern invalidation work.

Adversarial edges

test/stress/adversarial.ts:

Scenario	Result
1MB payload round-trip (JSON and STRING types)	intact, served from cache
10,000-key pattern-invalidation storm	exactly 10,000 deleted in 25ms
Server clock skewed +1 hour	timestamps follow the server clock (0s drift); reads stay consistent
500 readers + 5 writers + invalidation storms, same keyspace	10,000/10,000 resolved, zero corruption, no deadlock

Client compatibility (live)

• iovalkey and ioredis: full RedisApi conformance suite against live servers.
• @upstash/redis: 10/10 live tests against a real Upstash REST endpoint (detection, JSON and STRING round-trips, pattern invalidation, conformance).
• Custom adapters: the same conformance kit ships in test/redisApiConformance.ts for your own implementation.

The published artifact

Installed prisma-extension-redis@5.0.0-next.1 from the npm registry into a fresh consumer: no dependencies field, no nested node_modules — zero runtime dependencies confirmed on the wire. ESM import and CJS require both verified, the full cache/read/invalidate cycle ran against a live server from the published build, and the v5 remedial error (bring your own client) reaches consumers. The dist imports nothing except @prisma/client/extension.

Findings and known limitations

• RedisJSON memory is invisible to maxmemory accounting (measured on both Redis Stack 7.4 and Redis 8.8: ~13MB of JSON documents registered as ~1.7MB of used_memory). Consequence: maxmemory + LRU cannot bound a JSON-typed cache. Every entry this extension writes carries a TTL (ttl + stale), so the keyspace is always time-bounded — but if you rely on byte-based eviction as your safety net, use type: 'STRING', where accounting is exact (see the eviction leg above).
• Redis 8 includes JSON natively — type: 'JSON' works on plain redis:8 images with no Stack/module install. The JSON-less scenario applies to Redis ≤ 7 without the module.
• Redis Cluster is not supported for SCAN-based utilities (pattern invalidation, maintenance) or multi-key UNLINK — see ADAPTERS.md.
• Write invalidation is same-model: cached results of model A that include model B rows are not purged by writes to B (details).
• On the ioredis-family MULTI write path, a JSON-less server surfaces EXECABORT rather than unknown command; the startup probe and read-path hints carry the remedy in both cases.