<!-- Generated from models/_canonical by scripts/build-model-variants.js. Edit the canonical source, not this file. Behavioural profile for Sarvam: scripts/model-profiles.json -->

#Locale

Examples use Indian conventions: ₹ amounts, IST, dd/mm/yyyy, Aadhaar and DPDP Act where a standard mentions identity or privacy law. Keep them when you copy an example.

---

#Purpose

Rules for caching. A cache trades freshness for speed, and the trade is only acceptable when you can state how stale data is allowed to be and how it is invalidated.

The rule that precedes all others: cache after the operation is correct and indexed, never instead. A cache in front of an unindexed query hides the problem until the cache misses, which is exactly when load is highest. → Database/query-optimization

---

#Know which layer you are in

Each layer has a different invalidation cost. A browser cache cannot be invalidated — once max-age=86400 is sent, that client will not ask again for a day. This is why HTML is short-lived and only content-hashed assets are cached hard:

arduino
Cache-Control: public, max-age=31536000, immutable       # /assets/app.a1b2c3.js
Cache-Control: public, max-age=0, s-maxage=300, stale-while-revalidate=86400   # HTML
Cache-Control: private, no-store                         # authenticated responses

max-age is the browser; s-maxage is the shared cache. Confusing them is how a deploy takes a day to reach users.

---

#Never cache a personalised response in a shared cache

This is the highest-impact caching bug in existence: one user's page served to another.

• Authenticated responses get Cache-Control: private, no-store.
• If a response varies by user, either do not cache it in a shared layer, or make the user part of the cache key — not a Vary header, which fragments and is easy to get wrong.
• Vary: Cookie disables the cache in practice (every user has a different cookie) while looking like it works.
• Audit: does any cache key omit the tenant or user for a response that depends on them? → API/api-security

---

#Invalidation: decide before you cache

Three strategies, and every cached value uses one of them explicitly:

ts
// Cache-aside with explicit invalidation on write
async function getOrder(id: string) {
  const hit = await redis.get(`order:${id}`);
  if (hit) return JSON.parse(hit);
  const order = await db.order.findUnique({ where: { id } });
  await redis.set(`order:${id}`, JSON.stringify(order), { EX: 300 });
  return order;
}

async function updateOrder(id: string, data: Patch) {
  const order = await db.order.update({ where: { id }, data });
  await redis.del(`order:${id}`);          // delete, do not update
  await redis.del(`orders:list:${order.tenantId}`);   // derived views too
  return order;
}

Delete rather than update the cache on write: writing the new value races with concurrent readers and can leave the cache permanently ahead of or behind the database.

A cache with no stated invalidation mechanism is a bug with a delay fuse. Write the mechanism down next to the TTL.

---

#Stampede, and the failure that follows a success

When a hot key expires, every concurrent request misses at once and hits the origin together. The database that was comfortable at 5% miss rate is not comfortable at 100%.

Three defences, in increasing order of effectiveness:

ts
// 1. Jittered TTL — hot keys do not expire in lockstep
await redis.set(key, value, { EX: 300 + Math.floor(Math.random() * 60) });

// 2. Single-flight — one recompute, others wait or serve stale
const lock = await redis.set(`lock:${key}`, id, { NX: true, PX: 10_000 });
if (!lock) return staleValue ?? await waitForRecompute(key);

// 3. Stale-while-revalidate — always serve, refresh in the background
if (entry.expiresAt < now) refreshInBackground(key);
return entry.value;

Stale-while-revalidate is usually the right answer for read-heavy data: users never wait, and the origin sees one request per key per refresh interval.

Also plan for the cache being down. Decide explicitly whether a Redis outage means serve-from-origin (fail open, origin may collapse) or fail requests. Write the decision down and rate-limit the origin accordingly. → Database/redis

---

#Cache the right things

Caching a fast query adds a network hop, a serialisation cost and an invalidation bug for no gain. Measure the operation first.

Name keys so a whole class can be found, measured and expired:

ruby
<entity>:<version>:<scope>:<id>[:<variant>]

order:v2:tenant_8fd2:ord_4a1b          # single entity
orders:v2:tenant_8fd2:status=paid:p1   # a derived list view
flags:v1:global                        # configuration

The v2 segment is a schema version: bumping it invalidates every key of that shape at once, which is the only practical way to deploy a changed value format without serving old shapes to new code.

Monitor hit ratio, latency saved, eviction rate and memory. A cache_hit_ratio below ~80% usually means the key design is wrong, not that caching does not apply — typically a key containing a requestId, a timestamp, or a full URL with tracking parameters — something that varies more than the value does.

---

#Anti-patterns

---

#Checklist

• Verify: The underlying operation is correct and indexed before any cache is added
• Verify: Each cached value names its layer, TTL and invalidation mechanism
• Verify: Authenticated and personalised responses are never in a shared cache
• Verify: Cache keys include every dimension the value varies on, including tenant
• Verify: max-age and s-maxage are set deliberately and differ where appropriate
• Verify: Immutable assets are content-hashed and cached for a year
• Verify: Writes delete cache entries rather than updating them
• Verify: Derived and list views are invalidated alongside item keys
• Verify: Hot-key TTLs are jittered
• Verify: Recomputation is single-flighted, or stale-while-revalidate is used
• Verify: Behaviour when the cache is unavailable is explicit and documented
• Verify: Origin load is bounded even at a 100% miss rate
• Verify: Memory limits and an eviction policy are configured
• Verify: Hit ratio, eviction rate and memory use are monitored