Hangar, running · Two real runs
From nothing to production, for real
It's easy to draw a platform. These are two recordings of mine actually doing the work: a brand new service, from the first request to a canary on the prod cluster. Nothing is mocked. Waits are cut, cluster names are shown as dev and prod, and where something broke, it's left in.
gate-api, in the terminal
A small NodeJS service, from one claim to a canary on the prod cluster, with every command and every line of output shown as it really came back.
checkin-api, in Tower
A small Python service, the same path again through Tower: the create form, a staging environment, a crash loop of my own making, the App Configuration tab, and the release.
A demo that never breaks is a demo of the slides. The useful part is what the platform does when something is wrong.
What broke along the way
Running the whole path end to end, twice, found four real platform bugs that the individual tests hadn't. All four are fixed and written up in Glidepath's known gaps, and both videos end on the fixed path.
Four release gates ignored the new release file
Since the release-file split, a release lands in release.yaml, but sast, image-scan, sbom and provenance still only looked at values.yaml. Every new app's first release would have failed.
Image promotion waited forever
image-promotion waits for every gate, including the values check, and the values check never ran on a PR that only changed release.yaml.
The release relay address was wrong for a second machine
Every release wrote the dev node's own address for the prod cluster to call back on, which it can't reach. I'd been fixing it by hand. It's now a per-cluster setting in the cluster registry.
FixedA new app's release hung before it started
The release hooks run as a ServiceAccount that only exists when releaseTracking is set, and the ApplicationSet that creates it never read release.yaml. Every sync waited on a hook that couldn't start.
Two smaller things are still open. The /retest hint in the gate comments doesn't re-run a gate, and Tower showed my crash-looping pods as "Progressing" instead of saying why. Both are on the list.
Want the design behind what you just watched? The stack explains each tool, andCI/CD covers the delivery path in depth.