WFGY/ProblemMap/GlobalFixMap/OpsDeploy/cache_warmup_invalidation.md

Cache Warmup and Invalidation: OpsDeploy Guardrails

🧭 Quick Return to Map

You are in a sub-page of OpsDeploy.
To reorient, go back here:

• OpsDeploy — operations automation and deployment pipelines
• WFGY Global Fix Map — main Emergency Room, 300+ structured fixes
• WFGY Problem Map 1.0 — 16 reproducible failure modes

Think of this page as a desk within a ward.
If you need the full triage and all prescriptions, return to the Emergency Room lobby.

Warm caches stop cold-start spikes and prevent stampedes. Correct invalidation prevents stale or cross-index answers. Use this page to design keys, warmup order, and safe evictions for RAG and agent pipelines.

---

Open these first

• Readiness gate: rollout_readiness_gate.md
• Version locking: version_pinning_and_model_lock.md
• Index build and swap: vector_index_build_and_swap.md
• Canary staging: staged_rollout_canary.md
• Traceability and contracts: retrieval-traceability.md, data-contracts.md
• Live ops: live_monitoring_rag.md, debug_playbook.md

---

When to use

• After index rebuilds or alias flips.
• When prompts, models, rerankers, or analyzers change.
• If first calls after deploy are slow or inconsistent.
• If users see stale citations or mixed-arm results.
• If retry storms or thundering herd appear under load.

---

Acceptance targets

• First production call warms within 5 minutes of release.
• ΔS(question, retrieved) ≤ 0.45 on three paraphrases after warmup.
• Coverage ≥ 0.70 to the target section.
• λ convergent across two seeds with warmed headers.
• Duplicate side effects = 0 during warmup window.

---

60-second warmup plan

1. Partition the cache by version
   Include these in every cache key: INDEX_HASH, PROMPT_VER, MODEL_VER, RERANK_CONF, and optionally TENANT_ID, LOCALE.
   See: version_pinning_and_model_lock.md

2. Prime the hot paths
   Run a 20–40 item gold set to fill retrieval and rerank caches. Verify citations and offsets.
   See: retrieval-traceability.md

3. Enable single-flight
   Coalesce identical misses so only one worker computes each key. Stampede protection is mandatory.

4. Set layered TTLs
   Short TTL for query→top-k, slightly longer for rerank→final set, shortest for negative cache. Use jitter.

5. Wire safe evictions
   Do not scan and delete. Flip namespace by changing the versioned prefix. Old space dies out naturally.

---

Key design: scope and structure

• Two-level cache
  L1: query normalization → candidate ids.
  L2: candidate ids → ranked snippets with citation fields.

• Key template

L1: q:{hash(query\_norm)}:{INDEX\_HASH}:{PROMPT\_VER}:{RERANK\_CONF}
L2: r:{hash(candidates)}:{INDEX\_HASH}:{PROMPT\_VER}:{RERANK\_CONF}

If locale, tokenizer, or analyzer can change results, add LOCALE, TOK_VER, ANALYZER_CONF.

• What never to cache
  Tool call side effects or write steps. Use idempotency fences.
  See: idempotency_dedupe.md

---

Invalidation strategy

• Prefer namespace rotation over delete
  Keys carry the current INDEX_HASH and PROMPT_VER. A swap creates a new namespace. Old keys expire on TTL.

• Trigger list that must rotate namespace

• Index alias flips or index rebuilds.

• Prompt pack changes or header reordering.

• Reranker cutoff or model version changes.

• Analyzer or tokenizer changes.

• Negative cache discipline
  Cache misses for a short time only. TTL 10–30 seconds with jitter to avoid synchronized expiry.

• Per-tenant isolation
  If tenants use different corpora or flags, put TENANT_ID into the key prefix. Never share warmed entries across tenants.

---

Warmup ordering

1. Secrets and boot order pass.
   See: bootstrap-ordering.md
2. Retriever reports INDEX_HASH ready.
3. Run the gold set to fill L1 and L2.
4. Prime common user queries and help topics.
5. Verify ΔS, coverage, λ on warmed paths.
6. Open canary at 5 percent.
   See: staged_rollout_canary.md

---

Stampede control

• Single-flight lock for each key with a short lease.
• Request collapsing by query hash.
• Backoff and queueing on provider 429/5xx.
• Circuit breakers for downstream stores.

---

Store-specific notes

• Redis
  Avoid SCAN DEL. Use versioned prefixes. For atomic set-if-absent plus short lock, use SET key val NX PX=ms.

• CDN or edge
  Only cache read-only pages. For dynamic RAG responses, keep cache at app or Redis with versioned keys.

• Vector result caches
  Key by INDEX_HASH and metric. If metric or normalization changes, treat as a new namespace.
  See: metric_mismatch.md, normalization_and_scaling.md

• Fragmented stores
  If hit distributions vary wildly across partitions, caching will mask deeper issues. Fix the store first.
  See: pattern_vectorstore_fragmentation.md

---

Example: single-flight and versioned key (pseudo)

ver = f"{INDEX_HASH}:{PROMPT_VER}:{RERANK_CONF}"
k1  = f"q:{hash(query_norm)}:{ver}"

if redis.set(f"lock:{k1}", "1", nx=True, px=3000):
  try:
      val = compute_topk(query_norm)
      ttl = jitter(60, 15)  # 60s ±15%
      redis.set(k1, serialize(val), ex=ttl)
  finally:
      redis.delete(f"lock:{k1}")
else:
  sleep(0.05); val = redis.get(k1) or wait_poll(k1)
return val

---

Observability fields to log

• Version pins: INDEX_HASH, PROMPT_VER, MODEL_VER, RERANK_CONF, TOK_VER, ANALYZER_CONF.
• Cache events: hit, miss, fill time, lock wait, evict reason.
• Quality: ΔS(question, retrieved), ΔS(retrieved, anchor), coverage, λ states.
• Latency p50 and p95 for retrieve, rerank, reason.

---

Common pitfalls

• Deleting keys in place which causes long stalls. Rotate namespace instead.
• Mixed caches across regions or tenants. Prefix by region and TENANT_ID.
• Cache keys without INDEX_HASH, leading to stale blends after a swap.
• Negative cache with long TTL, causing false “no answer” pockets.
• Warming only L1 but not L2, so reranker work still stampedes.

---

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

---

🧭 Explore More

Module	Description	Link
WFGY Core	WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack	View →
Problem Map 1.0	Initial 16-mode diagnostic and symbolic fix framework	View →
Problem Map 2.0	RAG-focused failure tree, modular fixes, and pipelines	View →
Semantic Clinic Index	Expanded failure catalog: prompt injection, memory bugs, logic drift	View →
Semantic Blueprint	Layer-based symbolic reasoning & semantic modulations	View →
Benchmark vs GPT-5	Stress test GPT-5 with full WFGY reasoning suite	View →
🧙‍♂️ Starter Village 🏡	New here? Lost in symbols? Click here and let the wizard guide you through	Start →

---

👑 Early Stargazers: See the Hall of Fame ⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.