The purge that took down a homepage — and the fix

The purge that took down a homepage — and the fix

An electronics retailer in Berlin pushed a price update and ran a full purge_everything. Cache-hit ratio cratered from 94% to 6% for 40 minutes; origin CPU pinned at 100% while the whole catalog re-warmed.

We rebuilt purging as a surgical procedure:

— Tagged every response with a Cache-Tag header by product ID and category
— On a price change, purged only that tag — one SKU, not 80,000 URLs
— Kept a denylist: never allow purge_everything from the CMS, only from a guarded admin tool
— Added a soft-purge option that served stale-while-revalidate instead of cold MISS

Next price update purged 1 product. Cache-hit stayed at 93%. Origin barely noticed. The rule we wrote on the wall: tag at write time, purge by tag, never nuke the lot.

The number that mattered: 79,999 URLs that no longer get evicted for a one-product price change.