400M商品价格实时更新：MCP服务器数据新鲜度实战

When we launched BuyWhere MCP — a shopping data server for AI agents — we thought the hard part was getting 296M products into a single endpoint.

The hard part was keeping those prices fresh.

The Problem: AI Agents Recommend Stale Prices

Within the first month, users reported something uncomfortable: Claude, GPT, and local Llama agents were confidently recommending products with prices that were weeks old. Not because the LLM was hallucinating — our MCP server was actually returning them. The data was stale in our database.

For a price comparison tool, stale data isn't a bug. It's the product failing.

What We Tried First (And Why It Failed)

Approach 1: Crawl everything every 6 hours

With 296M products across Shopify, Amazon, and Shopee stores, re-crawling everything every 6 hours meant:

~4.9M requests per cycle

~163K requests per hour

Immediate IP blocks from every major marketplace

That lasted about 90 minutes before rate limits shut us down.

Approach 2: Crawl top 10% most-viewed products hourly

Better, but it created a popularity bias. Long-tail products — the exact queries AI agents are best at finding — stayed stale for days.

Users searching for "best adapter for Canon R5 to Fuji lens mount" got prices from last month.

The Architecture That Actually Works

We landed on a three-tier freshness system:

Tier 1: Hot Products (freshness target: < 1 hour)

Products appearing in agent conversations (search/get_deals calls)

Products in active price-drop alerts

~2M products, crawled every 45 minutes

Tier 2: Warm Products (freshness target: < 12 hours)

Products in comparison pages and deal roundups

~50M products, crawled on a rolling 6-hour window

Tier 3: Long-tail (freshness target: < 72 hours)

~348M products, crawled on a 3-day rolling window

Re-crawl triggered on access ("stale-read-through" pattern)

The key insight: crawl frequency should follow access patterns, not catalog size.

The Stale-Read-Through Pattern

When an agent queries a product that hasn't been refreshed recently:

1. Serve the cached price with a freshness: stale flag
2. Queue an async re-crawl for that product
3. If the product is queried again within the freshness window, serve the refreshed price

This means the products people actually look at get refreshed automatically, without pre-crawling the entire catalog.

Results After 3 Months

The MCP Server Implementation

For MCP servers serving real-time data, here's what matters:

Include freshness metadata in your tool responses. Our search_products and get_deals tools return lastUpdated timestamps so agents can decide whether to trust a price.

Don't block on freshness. A slightly stale price with a freshness flag is more useful than a 30-second timeout while you re-crawl.

Let the agent decide. We added a maxAge parameter so agents can request only fresh data when accuracy matters (e.g., before a purchase recommendation).

{
  "tool": "search_products",
  "params": {
    "query": "Sony WH-1000XM5",
    "country": "SG",
    "maxAge": 1440
  },
}

What I'd Do Differently

Start with the access log, not the catalog. We wasted 3 weeks crawling products nobody was looking at.

Add freshness flags from day one. Retroactive freshness tracking is painful.

Treat crawl failures as data. When Shopee blocks us, that's a signal that our crawl pattern needs rotation — not just a retry.

If you're building an MCP server that serves real-world data (not just static docs), freshness is your hardest problem. Happy to chat about what's worked for us.

BuyWhere MCP is live with 400M+ products if you want to test a shopping agent.