agents-inc/api-database-upstash

Upstash Patterns

Quick Guide: Upstash provides a REST/HTTP-based Redis client (@upstash/redis) designed for serverless and edge runtimes where TCP connections are unavailable. Unlike ioredis/node-redis, every command is an HTTP request -- no persistent connections, no connection pools, no teardown. The client automatically serializes/deserializes JSON (objects stored via set come back as objects from get), which is convenient but has gotchas with large numbers and cross-client compatibility. Use redis.pipeline() to batch commands into a single HTTP request, redis.multi() for atomic transactions, and @upstash/ratelimit for pre-built rate limiting algorithms. For background jobs, use @upstash/qstash which pushes messages to your API via HTTP webhooks.

---

<critical_requirements>

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use Redis.fromEnv() for initialization in production code -- never hardcode UPSTASH_REDIS_REST_URL or UPSTASH_REDIS_REST_TOKEN values)

(You MUST handle the pending promise from @upstash/ratelimit responses in edge runtimes -- use context.waitUntil(pending) on Vercel Edge/Cloudflare Workers or analytics data is lost)

(You MUST use redis.pipeline() when issuing 3+ independent commands in a single handler -- each command is a separate HTTP round-trip without pipelining)

(You MUST NOT use Upstash for Pub/Sub, blocking commands (BRPOP, BLPOP, XREAD BLOCK), or Lua scripting -- REST API does not support these; use ioredis with a TCP connection instead)

</critical_requirements>

---

Examples

• Core Patterns -- Client setup, commands, auto-serialization, pipeline, transactions
• Rate Limiting -- @upstash/ratelimit algorithms, middleware, analytics
• QStash -- Background jobs, scheduling, message publishing

Additional resources:

• reference.md -- Command cheat sheet, constructor options, environment variables, eviction policies

---

Auto-detection: Upstash, @upstash/redis, @upstash/ratelimit, @upstash/qstash, Redis.fromEnv, UPSTASH_REDIS_REST_URL, UPSTASH_REDIS_REST_TOKEN, Ratelimit.slidingWindow, Ratelimit.fixedWindow, Ratelimit.tokenBucket, serverless Redis, edge Redis, REST Redis

When to use:

• Serverless functions (AWS Lambda, Vercel, Netlify) that cannot maintain TCP connections
• Edge runtimes (Cloudflare Workers, Vercel Edge, Fastly Compute) that only support HTTP
• Rate limiting API routes with pre-built algorithms (sliding window, fixed window, token bucket)
• Caching in serverless/edge where ioredis connection pooling is impractical
• Background job scheduling with QStash (push-based, no long-running consumers needed)
• Global read latency optimization via Upstash Global Database with read replicas

Key patterns covered:

• @upstash/redis client setup with Redis.fromEnv() and constructor options
• Automatic JSON serialization/deserialization behavior and gotchas
• Pipeline batching (redis.pipeline()) and atomic transactions (redis.multi())
• @upstash/ratelimit algorithms: sliding window, fixed window, token bucket
• @upstash/qstash for serverless background jobs and scheduling
• Global Database architecture (primary + read regions, eventual consistency)
• Edge runtime compatibility and context.waitUntil() patterns

When NOT to use:

• Long-running servers with persistent connections (use ioredis -- lower latency per command via TCP)
• Pub/Sub, blocking commands, or Lua scripting (REST API does not support these)
• Write-heavy workloads on Global Database (writes always go to primary region)
• Latency-critical paths where per-command HTTP overhead (~5-15ms) is unacceptable (use ioredis with TCP for <1ms per command)
• Large payloads (>1 MB) -- REST API has payload size limits

---

<philosophy>

Philosophy

Upstash exists because serverless and edge runtimes cannot maintain TCP connections. Traditional Redis clients (ioredis, node-redis) rely on persistent TCP sockets -- they fail in Cloudflare Workers, break in short-lived Lambda functions, and cannot run in browser/WebAssembly environments. Upstash replaces TCP with REST/HTTP, trading per-command latency (~5-15ms vs <1ms) for universal compatibility.

Core principles:

1. Connectionless by design -- Every command is a stateless HTTP request. No connection pools, no teardown, no connection limits. This is a feature, not a limitation.
2. Auto-serialization is default -- Objects go in, objects come out. No manual JSON.stringify/JSON.parse. This simplifies 90% of use cases but surprises developers who expect raw string behavior.
3. Pipeline for performance -- Without pipelining, N commands = N HTTP requests. Always batch independent commands with redis.pipeline() to reduce round-trips.
4. Rate limiting as a first-class citizen -- @upstash/ratelimit provides production-ready algorithms without writing Lua scripts. The library handles all the Redis plumbing internally.
5. Push-based messaging -- QStash delivers messages TO your API via HTTP webhooks. No long-running consumer processes needed -- perfect for serverless.

</philosophy>

---

<patterns>

Core Patterns

Pattern 1: Client Setup with Redis.fromEnv()

Initialize using environment variables for zero-config deployment. See examples/core.md for full examples including constructor options and timeout configuration.

// Good Example
import { Redis } from "@upstash/redis";

const redis = Redis.fromEnv();
// Reads UPSTASH_REDIS_REST_URL and UPSTASH_REDIS_REST_TOKEN automatically

export { redis };

Why good: Zero-config, environment variables injected by platform (Vercel, Fly.io), no secrets in code

// Bad Example
import { Redis } from "@upstash/redis";

const redis = new Redis({
  url: "https://us1-merry-cat-12345.upstash.io",
  token: "AXXXAAIgcDE...",
});

Why bad: Hardcoded credentials leak in version control, non-portable across environments

---

Pattern 2: Automatic JSON Serialization

Upstash auto-serializes objects with JSON.stringify on write and JSON.parse on read. See examples/core.md for type-safe patterns and disabling auto-serialization.

// Good Example -- objects round-trip automatically
interface UserProfile {
  name: string;
  email: string;
  loginCount: number;
}

const CACHE_TTL_SECONDS = 3600;

await redis.set<UserProfile>(
  "user:123",
  {
    name: "Alice",
    email: "[email protected]",
    loginCount: 42,
  },
  { ex: CACHE_TTL_SECONDS },
);

// Returns typed object -- no JSON.parse needed
const user = await redis.get<UserProfile>("user:123");
// user is UserProfile | null

Why good: TypeScript generics provide type safety, no manual serialization, TTL set via options object

// Bad Example -- unnecessary manual serialization
await redis.set("user:123", JSON.stringify({ name: "Alice" }));
const raw = await redis.get("user:123");
const user = JSON.parse(raw as string); // Double-serialized: "{\"name\":\"Alice\"}"

Why bad: Auto-serialization already calls JSON.stringify -- doing it manually results in double-encoded strings that return as escaped JSON

---

Pattern 3: Pipeline Batching

Batch multiple commands into a single HTTP request. Without pipelining, each command is a separate round-trip (~5-15ms each). See examples/core.md for typed pipeline results.

// Good Example -- single HTTP request for all commands
const USER_TTL_SECONDS = 3600;

const pipe = redis.pipeline();
pipe.set("user:123:name", "Alice", { ex: USER_TTL_SECONDS });
pipe.set("user:123:email", "[email protected]", { ex: USER_TTL_SECONDS });
pipe.incr("stats:signups");

const results = await pipe.exec<["OK", "OK", number]>();
// results[0] => "OK"
// results[1] => "OK"
// results[2] => 1 (incremented value)

Why good: Single HTTP round-trip for 3 commands, typed results with generics, named TTL constant

// Bad Example -- 3 separate HTTP requests
await redis.set("user:123:name", "Alice");
await redis.set("user:123:email", "[email protected]");
await redis.incr("stats:signups");
// 3 round-trips = ~15-45ms total vs ~5-15ms with pipeline

Why bad: Each await is a separate HTTP request, tripling latency in serverless where every millisecond of cold start matters

---

Pattern 4: Atomic Transactions

Use redis.multi() when commands must execute atomically. See examples/core.md for examples.

// Good Example -- atomic counter + flag update
const tx = redis.multi();
tx.incr("order:count");
tx.set("order:last-updated", Date.now());
const [count, status] = await tx.exec<[number, "OK"]>();

Why good: All commands execute atomically (no interleaving from other clients), typed results

When to use pipeline vs transaction:

• Pipeline (redis.pipeline()) -- Commands are independent, you want batching for speed, atomicity not required
• Transaction (redis.multi()) -- Commands must all succeed together, no interleaving allowed

---

Pattern 5: Rate Limiting with @upstash/ratelimit

Pre-built rate limiting that handles all Redis internals. See examples/rate-limiting.md for all algorithms, middleware integration, and analytics.

// Good Example
import { Ratelimit } from "@upstash/ratelimit";
import { Redis } from "@upstash/redis";

const MAX_REQUESTS = 10;
const WINDOW_DURATION = "10 s";

const ratelimit = new Ratelimit({
  redis: Redis.fromEnv(),
  limiter: Ratelimit.slidingWindow(MAX_REQUESTS, WINDOW_DURATION),
  analytics: true,
});

const { success, limit, remaining, reset, pending } =
  await ratelimit.limit("user:123");

// CRITICAL: In edge runtimes, handle the pending promise
// context.waitUntil(pending);

if (!success) {
  return new Response("Too Many Requests", {
    status: 429,
    headers: {
      "X-RateLimit-Limit": String(limit),
      "X-RateLimit-Remaining": String(remaining),
      "X-RateLimit-Reset": String(reset),
    },
  });
}

Why good: No Lua scripts needed, named constants for limits, analytics for monitoring, proper 429 response with standard headers

---

Pattern 6: QStash Background Jobs

Push-based messaging for serverless. See examples/qstash.md for scheduling, retries, and receiver verification.

// Good Example -- publish a background job
import { Client } from "@upstash/qstash";

const qstash = new Client({
  token: process.env.QSTASH_TOKEN!,
});

await qstash.publishJSON({
  url: "https://your-app.com/api/process-order",
  body: { orderId: "order-456", action: "fulfill" },
  retries: 3,
  delay: "10s",
});

Why good: Fire-and-forget from handler, automatic retries on failure, configurable delay, at-least-once delivery guaranteed

</patterns>

---

<decision_framework>

Decision Framework

Upstash vs ioredis/node-redis

Which Redis client should I use?
|-- Running in edge runtime (Cloudflare Workers, Vercel Edge)?
|   --> @upstash/redis (only option -- no TCP available)
|-- Running in serverless (Lambda, Vercel Serverless)?
|   |-- Short-lived functions with no connection reuse?
|   |   --> @upstash/redis (no connection management overhead)
|   |-- Long-lived functions with connection pooling?
|       --> ioredis (lower per-command latency)
|-- Running on a persistent server (Docker, EC2, K8s)?
|   --> ioredis (persistent TCP = <1ms latency vs ~5-15ms HTTP)
|-- Need Pub/Sub, blocking commands, or Lua scripts?
|   --> ioredis (REST API cannot support these)
|-- Need to run in browser or WebAssembly?
    --> @upstash/redis (HTTP works everywhere)

Which Rate Limiting Algorithm?

Which @upstash/ratelimit algorithm should I use?
|-- Need strict, evenly distributed limiting?
|   --> slidingWindow -- smoothest, no burst-at-boundary issues
|-- Need simple, low-overhead limiting?
|   --> fixedWindow -- cheapest computationally, allows boundary bursts
|-- Need to allow burst traffic up to a capacity?
|   --> tokenBucket -- smooths bursts, allows initial spike up to maxTokens
|-- Need multi-region rate limiting?
    --> fixedWindow (slidingWindow has high Redis command overhead in multi-region)

Pipeline vs Transaction vs Sequential

How should I batch these Redis commands?
|-- Commands are independent (no ordering dependency)?
|   --> Pipeline (redis.pipeline()) -- non-atomic but single HTTP request
|-- Commands must execute atomically (all-or-nothing)?
|   --> Transaction (redis.multi()) -- atomic, single HTTP request
|-- Only 1-2 commands?
    --> Sequential is fine -- pipeline overhead not worth it

Global Database vs Regional

Should I use Upstash Global Database?
|-- Read-heavy workload with users worldwide?
|   --> Global Database -- reads from nearest replica
|-- Write-heavy workload?
|   --> Regional Database -- writes always go to primary, replication doubles write cost
|-- Need strong consistency?
|   --> Regional Database -- Global is eventually consistent
|-- Latency-sensitive reads from multiple continents?
    --> Global Database -- sub-1ms reads from nearest region

</decision_framework>

---

<red_flags>

RED FLAGS

High Priority Issues:

• Using JSON.stringify() before passing objects to redis.set() -- auto-serialization already handles this, resulting in double-encoded strings like "{\"name\":\"Alice\"}" that break on read
• Ignoring the pending promise from ratelimit.limit() in edge runtimes -- analytics data and multi-region sync are lost silently; use context.waitUntil(pending)
• Issuing 5+ sequential await redis.get/set() calls without pipelining -- each is a separate HTTP request, adding 25-75ms of unnecessary latency
• Attempting Pub/Sub (redis.subscribe), blocking commands (BRPOP, BLPOP), or Lua scripting (eval) -- Upstash REST API does not support these; use ioredis with TCP

Medium Priority Issues:

• Missing TTL on cached keys -- same as any Redis: unbounded memory growth until eviction kicks in
• Using Global Database for write-heavy workloads -- writes always route to primary region and replication doubles command costs
• Not setting automaticDeserialization: false when interoperating with non-Upstash clients -- other clients store raw strings, Upstash will fail to parse them as JSON
• Creating a new Redis instance per request instead of reusing a module-level singleton -- while connectionless, the client still benefits from HTTP keep-alive and warm connections

Common Mistakes:

• Expecting redis.get() to return a string when an object was stored -- auto-deserialization returns the original object type, not a JSON string
• Assuming pipeline execution is atomic -- pipelines batch for network efficiency but other clients can interleave; use redis.multi() for atomicity
• Using Ratelimit.slidingWindow with MultiRegionRatelimit -- sliding window has high Redis command overhead in multi-region setups; use fixedWindow instead
• Storing values larger than 1 MB -- REST API has payload size limits; store references and fetch large data from object storage

Gotchas & Edge Cases:

• Large numbers become strings: JavaScript cannot safely handle numbers > 2^53 - 1 (Number.MAX_SAFE_INTEGER). Upstash returns these as strings even when the TypeScript type says number. Always validate large numeric values.
• Base64 encoding by default: The SDK requests base64-encoded responses to handle edge cases. If you see garbled output like dmFsdWU=, the response encoding is interfering -- check responseEncoding option.
• redis.get() returns null for missing keys, not undefined: This matters for TypeScript narrowing -- check result !== null, not truthiness.
• SET options use an object, not positional args: Upstash uses redis.set("key", "value", { ex: 300 }) not redis.set("key", "value", "EX", 300) -- the ioredis positional argument style does not work.
• Global Database is eventually consistent: A write followed immediately by a read from a different region may return stale data. Design for eventual consistency or use regional database for strong consistency.
• hgetall returns an empty object {} for non-existent keys: Check Object.keys(result).length === 0, not result === null.
• blockUntilReady() does not work on Cloudflare Workers: Cloudflare's Date.now() behaves differently; use limit() with manual retry logic instead.
• No WATCH command: Upstash REST API does not support WATCH for optimistic locking. Use redis.multi() for atomic operations or implement application-level optimistic concurrency.
• Auto-pipelining is available: The SDK can automatically batch commands issued during the same event loop tick via enableAutoPipelining: true in the constructor.

</red_flags>

---

<critical_reminders>

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use Redis.fromEnv() for initialization in production code -- never hardcode UPSTASH_REDIS_REST_URL or UPSTASH_REDIS_REST_TOKEN values)

(You MUST handle the pending promise from @upstash/ratelimit responses in edge runtimes -- use context.waitUntil(pending) on Vercel Edge/Cloudflare Workers or analytics data is lost)

(You MUST use redis.pipeline() when issuing 3+ independent commands in a single handler -- each command is a separate HTTP round-trip without pipelining)

(You MUST NOT use Upstash for Pub/Sub, blocking commands (BRPOP, BLPOP, XREAD BLOCK), or Lua scripting -- REST API does not support these; use ioredis with a TCP connection instead)

Failure to follow these rules will cause credential leaks, silent data loss in edge runtimes, unnecessary latency from sequential HTTP requests, and runtime errors from unsupported commands.

</critical_reminders>