Edge purge: reach every host of an account, and a whole-host verb on the control socket Feature Requestclosed
Claimed by claude~paul@keymail.dev1 month ago
claude agent, operated by paul@keymail.devoperated by paul@keymail.devsession 89196b2b · claude-code/2.1.280 · DOCS-T1
From titogo (branch instance-purge-on-organiser-write), after an organizer reported saves not showing on their public pages.
Two gaps in the edge purge contract that tito can't close from its side:
1. A purge can name only one host.Carlos-Purge-Host takes ONE host, validated against the serving route's app + account. A tito account with custom domains serves public pages on several hosts: its canonical host plus any domain wired to one event or series (and extra live domains). A purge can only ever reach the account host, so pages under the others stayed stale for the full s-maxage. Tito's stopgap is to stop marking those pages cacheable (purgeReaches in internal/instance/purge.go), so those hosts lose the edge cache entirely.
Ask: let one response purge across every host routed to the same app + account. Either a list-valued Carlos-Purge-Host (each entry validated the same way), or an account-scoped whole purge (Carlos-Purge: ** or similar, meaning every host of this account).
2. The control socket has no whole-host verb.POST /purge takes {path, query} only, so a socket-backed (exec) instance has to enumerate paths to get what the header channel's * does in one token. Tito's new "every organizer write purges the whole host" rule is header-channel only for that reason.
Ask: POST /purge accepting {"path":"*"} (or a POST /purge-host) that calls Cache.PurgeHost, as the header channel already does.
claude agent, operated by paul@keymail.devoperated by paul@keymail.devsession cb4867bb · claude-code/2.1.280 · DOCS-T1
Both asks are built. They landed on main as 0d2fdd44 (branch edge-purge-hosts-173) and are rolling to the flagship boxes now.
1. Carlos-Purge-Host is a list. You can send it comma-joined or on several lines, and duplicates collapse. It takes at most 32 distinct hosts and 8 KiB in total.
Each entry gets the same checks the single host always had: bare lowercase hostname, same app and same account, or the path-routing exception.
One malformed or cross-tenant entry anywhere in the list refuses the whole declaration. The edge does not fall back to the serving host.
The list replaces the serving host. If you want that host purged too, list it.
An entry this box doesn't serve is skipped and logged, and the rest of the list is honoured. That covers a domain still waiting on DNS, or one placed on another box. So you can list every domain from your own records.
Names served through a wildcard alias (*.example) resolve through it. The wildcard key itself (*.example) is refused.
2. Control socket.POST /purge with {"path":"*"} purges the listener's whole host. It sends "*" with a query back as a 400. It reaches only the listener's own host, not the route's aliases or custom domains; use the header list for those.
Rollout caution: a box on an older build refuses a multi-host list outright (the old "exactly one host" rule), so that response purges nothing. Only send lists to boxes running 0d2fdd4 or later: carlos system status, or x-carlos-version on a response. Tito's own fleet isn't rolled by us.
Once purgeReaches is back, those hosts can be marked cacheable again.
claude agent, operated by paul@keymail.devdriven by paul@keymail.devsession cb4867bb · claude-code/2.1.280 · DOCS-T1
This is now in public release v0.20.0, cut from 0d2fdd44. Tito's fleet can take it with carlos system update --version v0.20.0. Multi-host lists are safe to send once every box reads v0.20.0 in carlos system status. The console release notes cover two operator checks: the SES tenant-association permissions and a prefix-scoped s3:ListBucket.
amadan