Cloudflare Workers in 2024: KV, D1, and the New Edge Stack

Actualizado: 2026-05-03

Cloudflare Workers moved from isolated edge function to complete edge platform during 2023-2024. The current stack combines Workers (compute) + KV (global key-value) + D1 (distributed SQLite) + R2 (S3-compatible object storage) + Durable Objects (stateful) + Queues + Workers AI (edge GPU inference). This article analyses when that stack seriously competes with AWS and when AWS remains the right answer.

Key takeaways

• V8 isolates with sub-5ms cold starts deliver sub-50ms p50 latency to users from any of 330+ PoPs.
• R2 eliminates egress fees — the most differentiated cost advantage versus S3.
• D1 (edge SQLite) reached GA in 2024 and covers small-to-mid CRUD apps well.
• Durable Objects are the right primitive for persistent WebSockets, global rate limiting, and collaborative apps.
• The 30-second request limit blocks long-running compute — that is the most important structural constraint.

The stack components

Workers — compute

• V8 isolates, cold start under 5 ms.
• Over 330 global PoPs.
• $5/mo base price, $0.30 per million requests.
• JavaScript, TypeScript, Rust (via Wasm), Python in beta.

KV — global key-value

• Eventual consistency; millisecond read latency.
• Natural use case: distributed cache, session tokens.

D1 — edge SQLite

• Replicated SQLite. Full SQL. GA since 2024.
• Pricing based on rows read and written.
• Use case: small-mid CRUD apps. Centralised writes, read replicas at each PoP.

R2 — object storage

• S3-API compatible. No egress fees — massive advantage versus AWS S3.
• Approximately $0.015/GB/mo.
• Use case: images, videos, static files.

Durable Objects

• Stateful edge compute. Strong per-object consistency. Persistent WebSockets.
• Use case: chat rooms, collaborative apps, global rate limiting.

Workers AI

• Edge GPU LLM inference (Llama, Mistral, others). Pay per token.
• Use case: chatbots, summarisation, image generation.

Cloudflare vs AWS for edge

Cloudflare wins on simple edge, latency and egress cost. AWS wins on enterprise ecosystem, mature managed databases, and compute-intensive or long-running workloads. The Fastly Compute article covers how another provider approaches the same space with more mature enterprise contracts.

Cases where Cloudflare fits

• Global edge APIs: user near any PoP, sub-20ms latency.
• Next.js / Astro / SvelteKit sites: see SvelteKit 1.0 and its real adoption for the adapter-cloudflare integration.
• Chat apps: Durable Objects with persistent WebSockets are exactly built for this.
• Edge image optimisation: Workers + R2 eliminates the origin round-trip.
• Global rate limiting: Durable Objects with strong per-object consistency.
• Rapid prototyping: setup in minutes, negligible cost at low volumes.

Cases where AWS still wins

• Intensive or long-running compute: Lambda accepts up to 15 minutes; Workers has a 30s limit.
• Mature managed DBs: RDS PostgreSQL versus D1, which is young.
• Streaming data: Kinesis, MSK.
• ML platforms: SageMaker.
• Complex IAM: AWS permission model is far richer.

Deployment with Wrangler

npm install -g wrangler
wrangler login

# Create worker
wrangler init my-app

# Deploy
wrangler deploy

# Tail logs
wrangler tail

Seconds-long deployment. No containers, no buckets to configure first.

Workers AI in practice

export default {
  async fetch(request, env) {
    const response = await env.AI.run("@cf/meta/llama-3-8b-instruct", {
      messages: [{ role: "user", content: "Hi" }],
    });
    return Response.json(response);
  },
};

LLM inference with under-1s edge latency. No GPU management. Pay per token. It complements architectures where the main model is accessible via LLM proxies like LiteLLM to manage multiple providers.

Honest limitations

• 30s max request: no long-running compute.
• Read-only D1 replicas: writes remain centralised.
• Durable Objects: limited concurrency per object.
• Python in beta: not production-ready without prior validation.
• Less rich debugging than serverful environments.
• Bundle size limit (~1-10MB depending on plan).

Real prices

Mid-size app (1M users/mo, 10M requests):

• Workers: $5 base + negligible.
• KV (1M reads): $0.50.
• D1 (1M reads): ~$1.
• R2 (10GB storage): $0.15.
• Approximate total: ~$10/mo.

The AWS equivalent typically runs $50-200/mo with similar workload.

Conclusion

The Cloudflare edge stack in 2024 is a real AWS alternative for many cases. For edge-native applications — global, latency-sensitive, medium complexity — it is simpler, cheaper, and faster to operate. For enterprise workloads with complex compliance, specific ecosystems, or intensive compute, AWS remains the primary option. The gap closes every month: it is worth evaluating Cloudflare for new projects before defaulting to AWS.