Purge — URL, surrogate key, or sledgehammer

Scene 07 · Purge — URL, surrogate key, or sledgehammer

Three purge flavors: URL (precise, slow at scale), surrogate-key (atomic, fast), zone-wide (sledgehammer that stampedes origin).

Previously

Purge is the escape hatch you reached for at the end of the last scene — but purging a copy in one POP doesn't help the other 299 POPs still serving it, so the question becomes WHICH POPs and WHICH entries one purge call actually invalidates.

Scene 07

Purge — URL, surrogate key, or sledgehammer

Watch

Diagram

A world map with 12 POPs; each POP holds a small grid of cached cells colored by surrogate-key tag (blue = product:42, orange = homepage, green = catalog, grey = other). The bottom panel offers three purge flavors: URL purge clears one cell per POP, surrogate-key purge clears every cell sharing the tag, zone-wide purge clears everything. After a deploy, you don't want to wait for TTL — you want to actively tell the cache to forget; that operation is called a purge. **purge** — invalidate cached entries across the CDN before their TTL expires. **surrogate key** — a tag the origin attaches to a response (Fastly term; Cloudflare calls them cache tags); one purge call invalidates every object sharing the tag. The propagation timeline shows p50 (~150 ms) and p99 (seconds) for the wave to reach all POPs; the origin RPS gauge spikes when zone-wide purge forces every POP to refill at once.

Sources

12 POPs, each with its own cached cells. Watch the purge wave radiate from the ingress POP.

A URL purge for /logo.png is fired from the Frankfurt ingress POP. Watch the wave ripple outward across all 12 POPs — only the /logo.png cell clears in each one. Origin RPS stays flat: each POP's next request will be a single miss.

Implementation

Edge.purge_handlers

three flavors — same control plane, different match rules

1def purge_url(url):
2    p = { kind: 'url', target: url }
3    broadcast_purge(p)  # one cell per POP
4 
5def purge_surrogate_key(tag):
6    p = { kind: 'surrogate', target: tag }
7    broadcast_purge(p)  # every cell sharing tag
8 
9def purge_zone():
10    p = { kind: 'zone', target: '*' }
11    broadcast_purge(p)  # every cell, every POP

POP.apply_purge

what each POP does when the purge message arrives

1def apply_purge(p):
2    for cell in self.cache:
3        if matches(cell, p):
4            cell.invalidate()  # next request = miss
5    ack(p.id, self.pop_id)
6 
7def matches(cell, p):
8    if p.kind == 'zone':       return True
9    if p.kind == 'url':        return cell.url == p.target
10    if p.kind == 'surrogate':  return p.target in cell.tags

Control.broadcast_purge

gossip the purge to every POP; track per-POP completion

1def broadcast_purge(p):
2    p.id = new_purge_id()
3    pending = set(all_pops)
4    for pop in all_pops:
5        pop.send(p)            # bimodal multicast / gossip
6    while pending:
7        pop_id = wait_for_ack(p.id)
8        pending.remove(pop_id)
9        # p50 ~150 ms; p99 = slowest POP in the fleet
10    return { p50: ack_p50(), p99: ack_p99() }

PreviousStale-while-revalidate — and the bug it ships NextCache key and Vary — when 'same URL' isn't