Ready

mirendev/runtime#1447

This is ready for a human to look at. Both of my earlier threads are fixed in the text, and the two new upgrade notes match the server docs.

Since the last review

I also checked a few entries against code. app disable requires the deploy permission (servers/app/authorization_test.go:69), so "anyone who can deploy can disable" holds. The cold-start page shows after 1.5s (servers/httpingress/httpingress.go:472).

Still open: confirming the release covers everything since v0.17.0 needs git log v0.17.0..main, which I can't run here.

I'm not recommending approval. The v0.10.0 cutoff, the hey flags and that completeness check all need someone with the git history.

Replies to existing threads

Posted to the PR's comment threads when you submit.


Verdict: ready

Data

{
  "owner": "mirendev",
  "repo": "runtime",
  "number": 1447,
  "verdict": "ready",
  "event": "comment",
  "summary": "This is ready for a human to look at. Both of my earlier threads are fixed in the text, and the two new upgrade notes match the server docs.\n\n**Since the last review**\n- **Etcd authorization note (`changelog.md:17`):** it now has its own Upgrade Notes bullet with three points:\n  - check every runner's certificate name before upgrading the coordinator;\n  - re-join mismatched runners with `remove`/`join`, keeping the same runner ID;\n  - roll back through the pre-upgrade etcd snapshot, not by swapping in the old binary.\n\n  All three match the `[etcd]` section of `server-config.md`, and the `#etcd` anchor resolves. \"Networking may stop\" also fits the docs' caveat that a legacy name can sometimes be mapped but networking isn't guaranteed.\n\n  One point needs someone with the history. The bullet says only runners that joined **before v0.10.0** are at risk, because they got a shortened 8-character name. `server-config.md` tells operators to check every runner, and I couldn't find the v0.10.0 cutoff anywhere in the tree. If that version is wrong, operators with affected runners will skip the check. The CommonName drift warning in `runner start` (`cli/commands/runner_start.go:321`) catches some of these cases either way.\n- **`distributed` mode removal (`changelog.md:18`):** this matches the Embedded Components section. That covers which settings are refused, the server refusing to start, the rollback during an upgrade, and leftover addresses being ignored. The anchor resolves.\n- **`service` claim:** it now has the caveat and points to the limits. I've resolved that thread.\n- **TTL default:** `pr-environments.md` now says 7 days, which matches the deploy flag's default.\n- **`hey` flags:** the entry now only says the short flags `-n`/`-c`/`-z` match. That's a smaller claim and less likely to mislead, but I still can't check it because the old command isn't in the tree.\n\nI also checked a few entries against code. `app disable` requires the deploy permission (`servers/app/authorization_test.go:69`), so \"anyone who can deploy can disable\" holds. The cold-start page shows after 1.5s (`servers/httpingress/httpingress.go:472`).\n\n**Still open:** confirming the release covers everything since v0.17.0 needs `git log v0.17.0..main`, which I can't run here.\n\nI'm not recommending approval. The v0.10.0 cutoff, the `hey` flags and that completeness check all need someone with the git history.",
  "replies": [
    {
      "thread_id": "PRRT_kwDONZ__ss63RIx2",
      "body": "Fixed by the new Upgrade Notes bullet at line 17. It covers the pre-upgrade runner check and re-joining with the same runner ID. It also says that only the managed rollback, not swapping in the old binary, turns the new checks off. All three points match `server-config.md#etcd`. I couldn't confirm the \"before v0.10.0\" cutoff in the tree, so I've asked a human to check it in the summary.",
      "resolve": true,
      "resolve_reason": "addressed"
    },
    {
      "thread_id": "PRRT_kwDONZ__ss63RIyN",
      "body": "The entry now says it only separates services when no role can deploy or exec into the app, and it points to the limits. That's the caveat I was after.",
      "resolve": true,
      "resolve_reason": "addressed"
    }
  ],
  "posted_to_pr": true,
  "draft": true,
  "head_sha": "5b5999568b056dbe49768efa9ca8aeed20a3a9a2"
}
Previous revisions (1)
Revision 1 · 2d ago

mirendev/runtime#1447

This is ready for a human to look at, but one upgrade note looks incomplete. Since this PR's job is to get v0.18 ready to cut, that gap should be settled before it merges.

The sealing-key upgrade note leaves out the etcd authorization steps (changelog.md:16) server-config.md describes etcd certificate-identity authorization as something the coordinator turns on "before importing coordinator credentials". That import is #1349's feature. The same section gives operators two things to act on:

  • Before upgrading: check that every runner certificate's CommonName is runner-<runner_id>. Re-provision any runner that doesn't match, or its etcd (Flannel) networking may stop once auth is on.
  • Downgrading: an in-place binary downgrade leaves authorization enabled. To go back you have to restore the pre-upgrade etcd snapshot through the managed rollback path.

The v0.17.x notes never mention etcd authorization. As far as I can tell from the docs, it first ships in this release. If so, the runner pre-check belongs in Upgrade Notes, because a distributed cluster can lose overlay networking on a release whose notes only say "back up a key". If authorization actually landed earlier, or sits behind the distributed-runner work you left out on purpose, ignore this.

Other things a reviewer should check

  • Completeness: "everything merged since v0.17.0" needs a git log v0.17.0..main to confirm, and I can't run one here. The 1365–1440 range has gaps (for example #1375–1377, #1393–1396, #1413–1414 and #1427–1429). Someone should confirm those are all CI, tests, docs or work that isn't switched on yet. Leaving out the v0.17.x patch PRs is correct.
  • hey flags: I couldn't check "It takes the same flags" for miren debug test load → hey, because the old command is gone from the tree. Users will copy flags straight across, so someone who knows the old command should confirm it.

Smaller wording fixes, neither blocking

  • service claim: the entry says a verifier "can tell an app's services apart" with no caveat. workload-identity.md says that only holds when nothing else can run code as the service. I left an inline note.
  • TTL default: the "How It Works" list in pr-environments.md still says "(default 24 hours)". The deploy flag defaults to 7d (cli/commands/deploy.go:111), and the "Previews last a week" entry links to that same page. This PR didn't cause it, but it's a one-line fix while you're here.

The other entries I could check against this branch hold up:

  • The new anchors all resolve.
  • The sealing-key path and the refusal behaviour match server-config.md.
  • The miren app run truncation entry matches runOutputTruncatedError / reportTruncatedRun in cli/commands/app_run.go.
  • The via_cloud note matches the cloud-routed docs.

I'm not recommending approval. The etcd authorization question, the completeness check and the hey claim all need someone with the git history to settle them.

Inline comments

  • docs/docs/changelog.md:16 — server-config.md says the coordinator enables etcd certificate-identity authorization "before importing coordinator credentials", which is this PR's #1349 migration. It also lists two operator steps that this note leaves out:

  • Before upgrading the coordinator: check that every runner certificate's CommonName is runner-<runner_id>, and re-provision mismatched runners (miren runner remove + join). Otherwise their etcd networking may stop once auth is enabled.

  • Downgrading: an in-place binary downgrade leaves authorization on. Going back means restoring the pre-upgrade etcd snapshot through the managed rollback.

If authorization first ships in v0.18 (I couldn't find it in any v0.17.x notes), the runner pre-check belongs in Upgrade Notes, either in this entry or as its own bullet. Distributed clusters are the ones it can break.

  • docs/docs/changelog.md:24 — workload-identity.md says service only keeps an app's services apart when nothing else can run code as one of them. That means the app runs as app-readonly or none, and no app on the cluster runs under a cluster role other than cluster-readonly. As written, this entry reads like an unconditional guarantee, and someone could set up per-service GCP roles on an app-admin app that don't actually isolate anything. Consider adding a short clause, e.g. "…tell an app's services apart, as long as no role can deploy or exec into the app (see the limits)".

Verdict: ready