Why: the 2026.x → 27.1 crossing includes layout step 2→3, which isn't rolling-safe. A plain rolling update stalls: the new pod waits for the old ones to stop, and they never stop because the new one never gets ready (spike S5, inbuxa/inbuxa-k8s-lab#5). The hook has to stop every server pod first, but only when a step needs it. S5 said the hook should ask the server rather than hard-code that.
inbuxa --upgrade-plan boots like --upgrade-store, writes nothing, and prints one line of JSON:
Can't tell (pre-0.16 store, or an error); the message says why
The exit code is there because the image has sh but no jq.
Tested
store_version integration test (RocksDB): a current store, stores marked 1 and 2 (the steps and their rolling-safe flags), a newer store, and that the plan leaves the mark untouched. 1 test passes.
Unit test for steps_from.
The debug binary on a new RocksDB store prints newStore: true, exit 0. After --upgrade-store marks it: store 3, no steps, exit 0.
fmt ok. The lock gains one line (serde_json for the inbuxa crate).
Note: this branch adds lines next to the K-S5 comments that #454 renames, so whichever merges second may conflict. I'll rebase it.
For the chart's upgrade hook (K-C9, tracker inbuxa/inbuxa-charts#14).
**Why:** the 2026.x → 27.1 crossing includes layout step 2→3, which isn't rolling-safe. A plain rolling update stalls: the new pod waits for the old ones to stop, and they never stop because the new one never gets ready (spike S5, inbuxa/inbuxa-k8s-lab#5). The hook has to stop every server pod first, but only when a step needs it. S5 said the hook should ask the server rather than hard-code that.
**`inbuxa --upgrade-plan`** boots like `--upgrade-store`, writes nothing, and prints one line of JSON:
```json
{"newStore":false,"steps":[{"from":1,"to":2,"rollingSafe":true,"what":"…"},{"from":2,"to":3,"rollingSafe":false,"what":"…"}],"stopAllNodes":true,"store":1,"target":3}
```
| Exit | Meaning |
|---|---|
| 0 | Nothing to do (current, or a new store) |
| 3 | Steps to run, all rolling-safe |
| 4 | A step needs every other node stopped first |
| 2 | A newer release changed the store |
| 1 | Can't tell (pre-0.16 store, or an error); the message says why |
The exit code is there because the image has `sh` but no `jq`.
**Tested**
- `store_version` integration test (RocksDB): a current store, stores marked 1 and 2 (the steps and their rolling-safe flags), a newer store, and that the plan leaves the mark untouched. 1 test passes.
- Unit test for `steps_from`.
- The debug binary on a new RocksDB store prints `newStore: true`, exit 0. After `--upgrade-store` marks it: store 3, no steps, exit 0.
- fmt ok. The lock gains one line (`serde_json` for the `inbuxa` crate).
**Note:** this branch adds lines next to the K-S5 comments that #454 renames, so whichever merges second may conflict. I'll rebase it.
jcoffey-dev
added this to the 27.1 milestone 2026-11-18 20:11:06 +00:00
For the Helm chart's upgrade hook (K-C9). The 2026.x to 27.1 crossing has
a layout step that isn't rolling-safe: a rolling update stalls, because
the new pod waits for the old ones to stop. The hook has to stop every
server pod first, but only when a step needs it, which the server knows
and the chart shouldn't hard-code (S5).
inbuxa --upgrade-plan boots as --upgrade-store does, writes nothing, and
prints one line of JSON: the store's layout version, this build's, and
each step with whether it's rolling-safe. Its exit code is for a shell
to branch on: 0 nothing to do, 3 rolling-safe steps only, 4 a step needs
every other node stopped, 2 a newer release changed the store, 1 can't
tell.
The store_version integration test covers a current store, layout 1 and
2, and a newer store, and that the plan leaves the mark alone.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
For the chart's upgrade hook (K-C9, tracker inbuxa/inbuxa-charts#14).
Why: the 2026.x → 27.1 crossing includes layout step 2→3, which isn't rolling-safe. A plain rolling update stalls: the new pod waits for the old ones to stop, and they never stop because the new one never gets ready (spike S5, inbuxa/inbuxa-k8s-lab#5). The hook has to stop every server pod first, but only when a step needs it. S5 said the hook should ask the server rather than hard-code that.
inbuxa --upgrade-planboots like--upgrade-store, writes nothing, and prints one line of JSON:The exit code is there because the image has
shbut nojq.Tested
store_versionintegration test (RocksDB): a current store, stores marked 1 and 2 (the steps and their rolling-safe flags), a newer store, and that the plan leaves the mark untouched. 1 test passes.steps_from.newStore: true, exit 0. After--upgrade-storemarks it: store 3, no steps, exit 0.serde_jsonfor theinbuxacrate).Note: this branch adds lines next to the K-S5 comments that #454 renames, so whichever merges second may conflict. I'll rebase it.
7e61761de0to9f3aa90127