Add a Read-Through Cache to the Catalog API

The catalog API serves product pages and checkout. It returns each product's details and its price in the shopper's currency. p99 read latency is 420 ms, and 80% of reads go to the top 2,000 products.

Bring p99 latency for product reads under 80 ms, and cut read load on the primary Postgres database by at least half.

We add a Redis cluster in front of Postgres. API responses are cached under the key product:{id}.

On a cache miss, the API reads the product from Postgres, writes the result to Redis, and responds.

When a product changes, the writer deletes the cache key and then updates Postgres. Keys expire after 24 hours.

The cache is flushed on every deploy, so new code never reads an entry written in an old format.

If Redis is unavailable, the API returns 503 so it never serves inconsistent data.

Redis gets 16 GB of memory, twice the size of the full catalog.

We'll alert on Redis CPU and memory.

Ship behind a flag at 5% of traffic, then raise it daily while watching p99 latency and Postgres read load.