Public endpoints
Giving a workload a URL on the internet, and why that URL still asks who you are.
A workload can be published at a hostname on the internet. A project admin creates the endpoint; anyone in the project can then open it.
What you choose
A subdomain. You pick the label, and it becomes
your-label.<the gateway's domain>. It must be a valid DNS label — 1 to 63
characters of lowercase letters, digits and hyphens.
Some names are refused: www, api, login, admin, mail and the
standard role addresses, along with anything starting xn--. That is not
tidiness. Those names sit under the operator’s certificate, and a tenant able
to claim login.<domain> could put a convincing sign-in page on the
operator’s own domain.
A port. The container port the gateway should forward to. It has to be the port your workload actually listens on — the gateway does not guess, and does not rewrite paths or inject middleware. What your container serves is what a visitor receives.
One subdomain per endpoint. A workload can have more than one endpoint if it serves on more than one port.
It is published, not open
This is the part people expect to work differently.
The hostname resolves from anywhere on the internet, but the gateway will not
serve your workload to a stranger. Access is a signed link: the console mints
a URL of the form https://your-host/?kon=<token>, which sets a
host-scoped session cookie and redirects. The token is short-lived — under
fifteen minutes — and it identifies a real user of your project.
So “public endpoint” means publicly addressable, not publicly readable. There is no anonymous mode and no per-endpoint setting to turn authentication off. If you need something genuinely open to the world, this is not the feature for it.
The cookie is __Host- prefixed with SameSite=Strict, and unsafe methods
are additionally checked against their Origin and Referer — so a link that
leaks is scoped to one host and cannot be used to make a cross-site request
back into your workload.
When it does not resolve
An unknown host returns 404 from the gateway rather than an error page from your workload. If you get one, the endpoint row and the hostname you typed disagree.
DNS is the operator’s side of this. The gateway’s domain needs a wildcard record pointing at it. If a brand-new endpoint does not resolve at all while existing ones do, that is worth raising rather than debugging from inside the workload.
A 502 means the gateway reached the cluster and the workload did not answer — usually the wrong port, or a container that has not finished starting. The workload’s own logs will say which.