vercel-kv（Vercel KV 搜索）

Vercel KV

Last Updated: 2026-01-21 Version: @vercel/kv@3.0.0 (Redis-compatible, powered by Upstash)

Quick Start

# Create KV: Vercel Dashboard → Storage → KV
vercel env pull .env.local  # Creates KV_REST_API_URL and KV_REST_API_TOKEN
npm install @vercel/kv

Basic Usage:

import { kv } from '@vercel/kv';

// Set with TTL (expires in 1 hour)
await kv.setex('session:abc', 3600, { userId: 123 });

// Get
const session = await kv.get('session:abc');

// Increment counter (atomic)
const views = await kv.incr('views:post:123');

CRITICAL: Always use namespaced keys (user:123 not 123) and set TTL for temporary data.

Common Patterns

Caching (cache-aside):

const cached = await kv.get(`post:${slug}`);
if (cached) return cached;

const post = await db.query.posts.findFirst({ where: eq(posts.slug, slug) });
await kv.setex(`post:${slug}`, 3600, post); // Cache 1 hour
return post;

Rate Limiting:

async function checkRateLimit(ip: string): Promise<boolean> {
  const key = `ratelimit:${ip}`;
  const current = await kv.incr(key);
  if (current === 1) await kv.expire(key, 60); // 60s window
  return current <= 10; // 10 requests per window
}

Session Management:

const sessionId = crypto.randomUUID();
await kv.setex(`session:${sessionId}`, 7 * 24 * 3600, { userId });

Pipeline (batch operations):

const pipeline = kv.pipeline();
pipeline.set('user:1', data);
pipeline.incr('counter');
const results = await pipeline.exec(); // Single round-trip

Key Naming: Use namespaces like user:123, post:abc:views, ratelimit:ip:endpoint

Critical Rules

Always:

• ✅ Set TTL for temporary data (setex not set)
• ✅ Use namespaced keys (user:123 not 123)
• ✅ Handle null returns (non-existent keys)
• ✅ Use pipeline for batch operations

Never:

• ❌ Forget to set TTL (memory leak)
• ❌ Store large values >1MB (use Vercel Blob)
• ❌ Use KV as primary database (it's a cache)
• ❌ Store non-JSON-serializable data (functions, BigInt, circular refs)

Known Issues Prevention

This skill prevents 15 documented issues:

Issue #1: Missing Environment Variables

Error: Error: KV_REST_API_URL is not defined or KV_REST_API_TOKEN is not defined Source: https://vercel.com/docs/storage/vercel-kv/quickstart | GitHub Issue #759 Why It Happens: Environment variables not set locally or in deployment. In monorepos (Turborepo/pnpm workspaces), abstracting @vercel/kv into a shared package can cause Vercel builds to fail even though local builds work. Prevention: Run vercel env pull.env.local and ensure .env.local is in .gitignore. For monorepos, either (1) create client in consuming app not shared package, (2) use Vercel Environment Variables UI to set at project level, or (3) add env vars to turbo.json pipeline config: {"pipeline": {"build": {"env": ["KV_REST_API_URL", "KV_REST_API_TOKEN"]}}}.

Issue #2: JSON Serialization Error

Error: TypeError: Do not know how to serialize a BigInt or circular reference errors. Also, hset() coerces numeric strings to numbers. Source: https://github.com/vercel/storage/issues/89 | GitHub Issue #727 Why It Happens: Trying to store non-JSON-serializable data (functions, BigInt, circular refs). Additionally, when using hset() to store string values that look numeric (e.g., '123456'), hgetall() returns them as numbers, breaking type consistency. Prevention: Only store plain objects, arrays, strings, numbers, booleans, null. Convert BigInt to string. For hash fields with numeric strings, either (1) use non-numeric prefix like 'code_123456', (2) store as JSON string and parse after retrieval, or (3) validate and recast types: String(value.field) after hgetall().

Issue #3: Key Naming Collisions

Error: Unexpected data returned, data overwritten by different feature Source: Production debugging, best practices Why It Happens: Using generic key names like cache, data, temp across different features Prevention: Always use namespaced keys: feature:id:type pattern.

Issue #4: TTL Not Set

Error: Memory usage grows indefinitely, old data never expires Source: Vercel KV best practices Why It Happens: Using set() without setex() for temporary data Prevention: Use setex(key, ttl, value) for all temporary data. Set appropriate TTL (seconds).

Issue #5: Rate Limit Exceeded (Free Tier)

Error: Error: Rate limit exceeded or commands failing Source: https://vercel.com/docs/storage/vercel-kv/limits Why It Happens: Exceeding 30,000 commands/month on free tier Prevention: Monitor usage in Vercel dashboard, upgrade plan if needed, use caching to reduce KV calls.

Issue #6: Storing Large Values

Error: Error: Value too large or performance degradation Source: https://vercel.com/docs/storage/vercel-kv/limits Why It Happens: Trying to store values >1MB in KV Prevention: Use Vercel Blob for files/images. Keep KV values small (<100KB recommended).

Issue #7: Type Mismatch on Get

Error: TypeScript errors, runtime type errors. Generic kv.get<T>() sometimes returns null even when data exists. Source: Common TypeScript issue | GitHub Issue #510 Why It Happens: kv.get() returns unknown type, need to cast or validate. Additionally, there's a type inference bug where using generics like kv.get<T>() can cause the function to return null even when CLI shows data exists, due to serialization/deserialization issues. Prevention: Don't use generics with get(). Instead, retrieve without type parameter and cast after retrieval: const rawData = await kv.get('key'); const data = rawData as MyType | null;. Validate with Zod or type guards before using.

Issue #8: Pipeline Errors Not Handled

Error: Silent failures, partial execution Source: https://github.com/vercel/storage/issues/120 Why It Happens: Pipeline execution can have individual command failures Prevention: Check results array from pipeline.exec() and handle errors.

Issue #9: Scan Operation Inefficiency

Error: Slow queries, timeout errors. In v3.0.0+, cursor type changed from number to string. Source: Redis best practices | Release Notes v3.0.0 Why It Happens: Using scan() with large datasets or wrong cursor handling. Version 3.0.0 introduced a breaking change where scan cursor is now string instead of number. Prevention: Limit count parameter, iterate properly with cursor, avoid full scans in production. In v3.0.0+, use let cursor: string = "0" and compare with cursor!== "0" (not !== 0).

Issue #10: Missing TTL Refresh

Error: Session expires too early, cache invalidates prematurely Source: Production debugging Why It Happens: Not refreshing TTL on access (sliding expiration) Prevention: Use expire(key, newTTL) on access to implement sliding windows.

Issue #11: scanIterator() Infinite Loop (v2.0.0+)

Error: for await loop never terminates when using kv.scanIterator() Source: GitHub Issue #706 Why It Happens: Bug in v2.0.0+ where iterator doesn't properly signal completion. The iterator processes keys correctly but never exits, preventing the function from returning. Also affects sscanIterator(). Prevention: Use manual scan() with cursor instead of scanIterator().

// Don't use scanIterator() - it hangs in v2.0.0+
for await (const key of kv.scanIterator()) {
  // This loop never terminates
}

// Use manual scan with cursor instead
let cursor: string = "0"; // v3.x uses string
do {
  const [newCursor, keys] = await kv.scan(cursor);
  cursor = newCursor;
  for (const key of keys) {
    const value = await kv.get(key);
    // process key/value
  }
} while (cursor !== "0");

Issue #12: zrange() with rev: true Returns Empty Array

Error: kv.zrange(key, 0, -1, {rev: true}) returns empty array even though data exists Source: GitHub Issue #742 Why It Happens: SDK bug in reverse flag handling for certain key patterns. CLI always returns correct values. Removing the rev flag returns data correctly. Prevention: Omit rev flag and reverse in-memory, or use zrevrange() instead.

// This sometimes returns empty array (BUG)
const chats = await kv.zrange(`user:chat:${userId}`, 0, -1, { rev: true });
// [] - but CLI shows 12 items

// Workaround 1: Omit rev flag and reverse in-memory
const chats = await kv.zrange(`user:chat:${userId}`, 0, -1);
const reversedChats = chats.reverse();

// Workaround 2: Use zrevrange instead
const chats = await kv.zrevrange(`user:chat:${userId}`, 0, -1);

Issue #13: Next.js Dev Server Returns Null on First Call

Error: kv.get() returns null on first call after starting dev server, correct data on subsequent calls Source: GitHub Issue #781 Why It Happens: Next.js static rendering caches the first response. Happens on each compilation restart in Next.js dev server. Prevention: Force dynamic rendering with unstable_noStore(), use cache: 'no-store' in fetch calls, or add retry logic.

import { unstable_noStore as noStore } from 'next/cache';

export async function getData() {
  noStore(); // Force dynamic rendering
  const data = await kv.get('mykey');
  return data; // Now returns correct value on first call
}

// Or add retry logic
async function getWithRetry(key: string, retries = 2) {
  let data = await kv.get(key);
  let attempt = 0;
  while (!data && attempt < retries) {
    await new Promise(r => setTimeout(r, 100));
    data = await kv.get(key);
    attempt++;
  }
  return data;
}

Issue #14: Vite "process is not defined" Error

Error: Uncaught ReferenceError: process is not defined when importing @vercel/kv in Vite Source: GitHub Issue #743 Why It Happens: Vite doesn't polyfill Node.js process.env by default. @vercel/kv expects process.env to exist. Prevention: Use Vite's define to polyfill, use createClient with import.meta.env, or install vite-plugin-node-polyfills.

// Option 1: Vite config with define
// vite.config.ts
import { defineConfig, loadEnv } from 'vite';

export default defineConfig(({ mode }) => {
  const env = loadEnv(mode, process.cwd(), '');
  return {
    define: {
      'process.env.KV_REST_API_URL': JSON.stringify(env.KV_REST_API_URL),
      'process.env.KV_REST_API_TOKEN': JSON.stringify(env.KV_REST_API_TOKEN),
    },
  };
});

// Option 2: Use createClient with import.meta.env
import { createClient } from '@vercel/kv';

const kv = createClient({
  url: import.meta.env.VITE_KV_REST_API_URL,
  token: import.meta.env.VITE_KV_REST_API_TOKEN,
});

Issue #15: Next.js Server Actions Return Stale Data

Error: kv.get() in Server Actions returns stale/cached data even after KV values are updated Source: GitHub Issue #510 Why It Happens: Next.js static rendering caches Server Action responses. Console output appears yellow (cache hit indicator). Prevention: Use unstable_noStore() to force dynamic rendering, or use route handlers instead of Server Actions.

// In Server Actions
'use server'
import { unstable_noStore as noStore } from 'next/cache';

export async function logChat(text: string) {
  noStore(); // Force dynamic rendering
  let n_usage = await kv.get('n_usage');
  // Now returns fresh value, not cached
}

// Or use route handlers (automatically dynamic)
// app/api/chat/route.ts
export async function GET() {
  let n_usage = await kv.get('n_usage'); // Fresh data
  return Response.json({ n_usage });
}

Version Migration Guide

v3.0.0 Breaking Changes

Cursor Type Changed (Release Notes):

• Scan cursor now returns string instead of number
• Update comparisons from cursor!== 0 to cursor!== "0"

// v2.x
let cursor: number = 0;
do {
  const [newCursor, keys] = await kv.scan(cursor);
  cursor = newCursor;
} while (cursor !== 0);

// v3.x
let cursor: string = "0";
do {
  const [newCursor, keys] = await kv.scan(cursor);
  cursor = newCursor;
} while (cursor !== "0");

v2.0.0 Breaking Changes

Auto-Pipelining Enabled by Default (Release Notes):

• Commands now automatically batched for performance
• May cause timing issues in edge cases
• Disable with enableAutoPipelining: false if needed

// If auto-pipelining causes issues, disable it:
import { createClient } from '@vercel/kv';

const kv = createClient({
  url: process.env.KV_REST_API_URL,
  token: process.env.KV_REST_API_TOKEN,
  enableAutoPipelining: false // Disable auto-pipelining
});

scanIterator() Bug Introduced: v2.0.0+ has infinite loop bug with scanIterator(). See Issue #11 for workaround.

Known Limitations

Redis Streams Not Supported

Source: Vercel KV Redis Compatibility | GitHub Issue #278

@vercel/kv does not support Redis Streams (XREAD, XADD, XGROUP, etc.) even though underlying Upstash Redis supports them. The package lacks stream methods.

Workaround: Use @upstash/redis directly for streams:

// @vercel/kv doesn't have stream methods
// await kv.xAdd(...) // TypeError: kv.xAdd is not a function

// Use Upstash Redis client directly
import { Redis } from '@upstash/redis';

const redis = new Redis({
  url: process.env.KV_REST_API_URL!,
  token: process.env.KV_REST_API_TOKEN!,
});

await redis.xadd('stream:events', '*', { event: 'user.login' });
const messages = await redis.xread('stream:events', '0');

Hash Operations Return Objects Not Strings

Source: GitHub Issue #674

kv.hgetall() returns a JavaScript object, not a JSON string. This confuses developers expecting string serialization.

await kv.hset('foobar', { '1834': 'https://example.com' });

const data = await kv.hgetall('foobar');
console.log(typeof data); // "object" not "string"

// It's already an object - use directly
console.log(data['1834']); // 'https://example.com'

// If you need JSON string
const jsonString = JSON.stringify(data);

Advanced Patterns

Distributed Lock (prevents race conditions):

const lockKey = `lock:${resource}`;
const lockValue = crypto.randomUUID();

const acquired = await kv.setnx(lockKey, lockValue);
if (acquired) {
  await kv.expire(lockKey, 10); // TTL prevents deadlock
  try {
    await processOrders();
  } finally {
    const current = await kv.get(lockKey);
    if (current === lockValue) await kv.del(lockKey);
  }
}

Leaderboard (sorted sets):

await kv.zadd('leaderboard', { score, member: userId.toString() });

// Note: zrange with { rev: true } has a known bug (Issue #12)
// Use zrevrange instead for reliability
const top = await kv.zrevrange('leaderboard', 0, 9, { withScores: true });
const rank = await kv.zrevrank('leaderboard', userId.toString());

Last verified: 2026-01-21 | Skill version: 2.0.0 | Changes: Added 5 new issues (scanIterator hang, zrange rev bug, Next.js dev server null, Vite process error, Server Actions stale cache), expanded Issues #1, #2, #7, #9 with TIER 1-2 research findings, added Version Migration Guide and Known Limitations sections