Upgrade the gitea recipe app image 28.0.0-rootless → 28.1.0-rootless (upstream v28.1.0, released 2026-10-06 — the first maintenance release of the 28 series).
Context / reconciliation. Upstream's default branch master is at 97eb7bf = release 3.6.8+28.0.0-rootless, i.e. the 1.27.3 → 28.0.0 major bump is already merged and published upstream. (The local abra checkout often sits on a stale local master branch at 1.27.3, which is why abra recipe upgrade still reports from 1.27.3.) Per the "re-plan from the tip of origin/main" rule, this PR is the 28.0.0 → 28.1.0 delta against the real current upstream main (97eb7bf). The 1.27 → 28 notes are retained below for reference.
What changed
service
image
old → new
app
gitea/gitea
28.0.0-rootless → 28.1.0-rootless
dbpostgres overlay 15.19 — HELD. The only available bumps are majors (16.x/17.x/18.x); this recipe pins plain postgres (not pgautoupgrade) and ships no pg_upgrade tooling (pg_backup.sh is logical pg_dump/psql only), so a major bump breaks every in-place upgrade. No within-major 15.x bump exists (15.19 is the latest 15.x; Docker Hub 15.20 → 404).
dbmariadb overlay 10.11.19 — current (Docker Hub 10.11.20 → 404) — no bump.
No app.ini.tmpl / compose-overlay changes required.
No breaking changes in 28.1.0. The release notes contain only Enhancements + Bug Fixes + security fixes.
mariadb overlay behaviour note: 28.1.0 switches MySQL/MariaDB to READ COMMITTED transaction isolation (#39506) and MSSQL to READ_COMMITTED_SNAPSHOT. Standard container default; no config change needed.
egress note:#39560 reclassifies private/link-local/CGNAT/documentation ranges as restricted rather than reserved, so they can be dialed in lax mode when explicitly allowed. This recipe sets no egress policy (default lax) ⇒ no action.
For reference — the 28.0.0 BREAKING items now in the base (no action for this recipe):
git-egress rework (#39426): ALLOWED_DOMAINS/BLOCKED_DOMAINS/ALLOW_LOCALNETWORKS deprecated in favour of ALLOWED_HOST_LIST/BLOCKED_HOST_LIST + EGRESS_MODE; the external deny preset is removed. This recipe sets none of them.
9 CVEs fixed by this upgrade (3 high, 6 medium) — the deterministic GitHub-advisory scan and the vendor blog agree exactly (all fixed via #39501/#39507):
CVE
severity
issue
CVE-2026-86684
high
push-mirror API checked owner's (not caller's) local-path permission
CVE-2026-96594
high
stored XSS — repo media API rendered small non-LFS HTML on the Gitea origin
CVE-2026-97208
high
push-mirror API ignored [mirror] DISABLE_NEW_PUSH
CVE-2026-101023
medium
OAuth2 refresh-token grant didn't check token kind (lifetime extension)
1 further advisory the scanner could not auto-judge (CVE-2026-62925, medium) is affected <= 1.27.2 and was already fixed in 1.27.3 (--adjudicate evidence: advisory text "Fixed in Gitea 1.27.3 (#39026, #39029)"), i.e. it predates this window — not counted. Scan sources: 25 (20 registry URLs + advisory APIs); no failed sources.
Version bump (operator's final step — the PR does NOT touch the label)
28.0.0 → 28.1.0 is a minor upstream release (no BREAKING) ⇒ release as a minor:
abra recipe release gitea -y
Verification
Direct --chaos deploy on cc-ci (dev-gitea.ci.commoninternet.net, mariadb overlay, chaos version f2b5089, 2026-10-16): converged 1/1 app (gitea/gitea:28.1.0-rootless) + 1/1 db (mariadb:10.11.19); app log Gitea version: 28.1.0 built with go1.27.1; DB collation preprocess ran + fresh schema initialised (117 tables); /api/healthz200. Dev deploy torn down (0 leaked stacks/volumes/env). warm-gitea untouched.
!testme on this PR (existing tests, full suite) — verified in the PR comments/runs on the PR head.
Upgrade the `gitea` recipe app image **`28.0.0-rootless` → `28.1.0-rootless`** (upstream **v28.1.0**, released 2026-10-06 — the first maintenance release of the 28 series).
> **Context / reconciliation.** Upstream's default branch `master` is at `97eb7bf` = release `3.6.8+28.0.0-rootless`, i.e. the **1.27.3 → 28.0.0** major bump is **already merged and published upstream**. (The local abra checkout often sits on a stale local `master` branch at 1.27.3, which is why `abra recipe upgrade` still reports *from* 1.27.3.) Per the "re-plan from the tip of origin/main" rule, this PR is the **28.0.0 → 28.1.0** delta against the real current upstream main (`97eb7bf`). The 1.27 → 28 notes are retained below for reference.
## What changed
| service | image | old → new |
|---------|-------|-----------|
| app | gitea/gitea | `28.0.0-rootless` → `28.1.0-rootless` |
- `db` **postgres overlay `15.19` — HELD.** The only available bumps are majors (16.x/17.x/18.x); this recipe pins plain `postgres` (not pgautoupgrade) and ships no pg_upgrade tooling (`pg_backup.sh` is logical pg_dump/psql only), so a major bump breaks every in-place upgrade. No within-major 15.x bump exists (15.19 is the latest 15.x; Docker Hub 15.20 → 404).
- `db` **mariadb overlay `10.11.19` — current** (Docker Hub 10.11.20 → 404) — no bump.
- No `app.ini.tmpl` / compose-overlay changes required.
**Upstream release notes:** app gitea/gitea 28.0.0→28.1.0: https://github.com/go-gitea/gitea/releases/tag/v28.1.0
**Vendor security write-up:** https://blog.gitea.com/release-of-28.1.0/
## Operator Action Required
- **No breaking changes in 28.1.0.** The release notes contain only Enhancements + Bug Fixes + security fixes.
- **mariadb overlay behaviour note:** 28.1.0 switches MySQL/MariaDB to `READ COMMITTED` transaction isolation ([#39506](https://github.com/go-gitea/gitea/pull/39506)) and MSSQL to `READ_COMMITTED_SNAPSHOT`. Standard container default; no config change needed.
- **egress note:** [#39560](https://github.com/go-gitea/gitea/pull/39560) reclassifies private/link-local/CGNAT/documentation ranges as *restricted* rather than reserved, so they can be dialed in `lax` mode when explicitly allowed. This recipe sets no egress policy (default `lax`) ⇒ no action.
- **For reference — the 28.0.0 BREAKING items now in the base** (no action for this recipe):
- git-egress rework ([#39426](https://github.com/go-gitea/gitea/pull/39426)): `ALLOWED_DOMAINS`/`BLOCKED_DOMAINS`/`ALLOW_LOCALNETWORKS` deprecated in favour of `ALLOWED_HOST_LIST`/`BLOCKED_HOST_LIST` + `EGRESS_MODE`; the `external` deny preset is removed. This recipe sets none of them.
- actions `RUN_RETENTION_DAYS` ([#38855](https://github.com/go-gitea/gitea/pull/38855)): new `[actions] RUN_RETENTION_DAYS` (0 = keep forever). Defaults apply.
## Security — advisory scan (deterministic) + vendor notes, window 28.0.0 → 28.1.0
**9 CVEs fixed by this upgrade** (3 high, 6 medium) — the deterministic GitHub-advisory scan and the vendor blog agree exactly (all fixed via [#39501](https://github.com/go-gitea/gitea/pull/39501)/[#39507](https://github.com/go-gitea/gitea/pull/39507)):
| CVE | severity | issue |
|---|---|---|
| CVE-2026-86684 | high | push-mirror API checked owner's (not caller's) local-path permission |
| CVE-2026-96594 | high | stored XSS — repo media API rendered small non-LFS HTML on the Gitea origin |
| CVE-2026-97208 | high | push-mirror API ignored `[mirror] DISABLE_NEW_PUSH` |
| CVE-2026-101023 | medium | OAuth2 refresh-token grant didn't check token kind (lifetime extension) |
| CVE-2026-104633 | medium | migration page-size `0` ⇒ downloader loop / memory exhaustion |
| CVE-2026-105267 | medium | tag-delete route also deleted releases + attachments |
| CVE-2026-105268 | medium | issue attachment API allowed renaming/deleting comment attachments |
| CVE-2026-89182 | medium | push-to-create ignored `[repository] FORCE_PRIVATE` |
| CVE-2026-97626 | medium | profile RSS/Atom leaked private/limited users' activity feed |
1 further advisory the scanner could not auto-judge (CVE-2026-62925, medium) is affected `<= 1.27.2` and was **already fixed in 1.27.3** (`--adjudicate` evidence: advisory text "Fixed in Gitea 1.27.3 (#39026, #39029)"), i.e. it predates this window — not counted. Scan sources: 25 (20 registry URLs + advisory APIs); no failed sources.
## Version bump (operator's final step — the PR does NOT touch the label)
28.0.0 → 28.1.0 is a **minor** upstream release (no BREAKING) ⇒ release as a minor:
abra recipe release gitea -y
## Verification
- **Direct `--chaos` deploy on cc-ci** (`dev-gitea.ci.commoninternet.net`, mariadb overlay, chaos version `f2b5089`, 2026-10-16): converged `1/1` app (`gitea/gitea:28.1.0-rootless`) + `1/1` db (`mariadb:10.11.19`); app log `Gitea version: 28.1.0 built with go1.27.1`; DB collation preprocess ran + fresh schema initialised (117 tables); `/api/healthz` **200**. Dev deploy torn down (0 leaked stacks/volumes/env). `warm-gitea` untouched.
- `!testme` on this PR (existing tests, full suite) — verified in the PR comments/runs on the PR head.
- **NOT merged** — for operator review.
cc @trav @notplants
Tested green on the cc-ci recipe CI server (full suite, cold, against this PR head). NOT merged — for operator review.
cc @trav @notplants
autonomic-bot
requested review from trav 2026-10-02 04:01:42 +00:00
autonomic-bot
requested review from notplants 2026-10-02 04:01:42 +00:00
Three of the four custom tests pass (test_admin_api_user_org_token_lifecycle — admin API CRUD, test_git_push — full SCM push round-trip, test_gitea_root_returns_200). The one failure is test_lfs_roundtrip, and it fails in the test's own restart polling loop, not in the LFS
round-trip or the recipe.
What the failing test asserts vs. what actually failed
# after `docker service update --force <stack>_app` (restart):deadline=time.time()+120whiletime.time()<deadline:status2,_=_api(live_app,"/version",user=user,password=password)# urllib, timeout=20ifstatus2==200:breaktime.sleep(5)assertstatus2==200,"gitea did not come back up after restart"
The traceback is a TimeoutError: The read operation timed out on that _api() call
(urllib.request.urlopen(..., timeout=20)) — i.e. the single /api/v1/version request stalled
~20 s and raised, aborting the whole test. Crucially:
the test is running _api()from inside the cc-ci runner/harness process on the cc-ci host
(it shells out to docker service update directly), so this poll is a local HTTP call to the
app's public /api/v1/version — it is not the app's own healthcheck;
the app was healthy and serving — all four other tests in the same custom stage pass, the
install/upgrade stages already asserted the same app healthz/version endpoint, and the App
registry in this very run reports install/upgrade/backup/restore = pass;
so the restart did not break the app; the test's fixed 20 s socket timeout simply expired on one
poll during the app's post-restart window.
Why this reads as genuinely stale
This exact test is green on the current 1-safe-line base: the 2026-08-31 run (app 1.27.2 →
1.27.3-rootless) reports !testme GREEN on the first run, custom tier included.
gitea 28.0.0 is a major version jump that substantially refactored git/model layers and boot
wiring; a cold-restart boot + first-request window is exactly the kind of timing the fixed 20 s
timeout (and the comment's own # Wait for gitea to come back up TODO) is brittle to.
Per tests/STYLE.md §5, "derive waits from the recipe's declared readiness, not a guess" — this
loop hard-codes timeout=20 per attempt instead of using the recipe's declared HTTP_TIMEOUT
(600 s) / READY_PROBE. That is the test's bug, not the recipe's.
Nothing in the recipe needed to change for it: the dire dev deploy of this same PR head
converged, and admin + repo API + /api/v1/version all returned 200 on 28.0.0.
What I did NOT do (DEFAULT mode)
I did not modify the test or the harness. !testme cannot go fully green on this PR without a
test change, and test changes are gated behind --with-tests. So this PR is recorded as SUCCESS-PENDING-TESTS in DEFAULT mode.
Operator options
Re-run /recipe-upgrade gitea --with-tests — that will open a verified cc-ci test PR to
replace the brittle fixed-timeout restart poll with the recipe's declared readiness
(HTTP_TIMEOUT / READY_PROBE, per tests/STYLE.md §5). This PR's !testme then goes green
once the test PR is merged.
Treat the run as a flake and simply re-post !testme on this PR (within the budget) — if the
restart window is kind, the same test passes; but the fixed-timeout brittleness remains.
The recipe upgrade itself (single line: gitea/gitea:1.27.3-rootless → 28.0.0-rootless) is
verified live on the cc-ci dev swarm and by the install/upgrade/backup/restore/lint tiers. NOT merged.
## `!testme` is RED — but only on a stale LFS custom test (recipe upgrade itself verified GREEN on every lifecycle tier)
The upgrade itself looks correct. The cc-ci run (`#31`, on `91e42c8a`) is:
```
install pass
upgrade pass ← in-place 1.27.3 → 28.0.0 redeploy, ci-marker repo survived
backup pass
restore pass
lint pass
custom FAIL ← tests/gitea/custom/test_lfs_roundtrip.py (only)
```
Three of the four custom tests pass (`test_admin_api_user_org_token_lifecycle` — admin API CRUD,
`test_git_push` — full SCM push round-trip, `test_gitea_root_returns_200`). The one failure is
`test_lfs_roundtrip`, and it fails **in the test's own restart polling loop, not in the LFS
round-trip or the recipe**.
### What the failing test asserts vs. what actually failed
```python
# after `docker service update --force <stack>_app` (restart):
deadline = time.time() + 120
while time.time() < deadline:
status2, _ = _api(live_app, "/version", user=user, password=password) # urllib, timeout=20
if status2 == 200:
break
time.sleep(5)
assert status2 == 200, "gitea did not come back up after restart"
```
The traceback is a **`TimeoutError: The read operation timed out`** on that `_api()` call
(`urllib.request.urlopen(..., timeout=20)`) — i.e. the single `/api/v1/version` request stalled
~20 s and raised, aborting the whole test. Crucially:
- the test is running `_api()` **from inside the cc-ci runner/harness process on the cc-ci host**
(it shells out to `docker service update` directly), so this poll is a *local* HTTP call to the
app's public `/api/v1/version` — it is **not** the app's own healthcheck;
- the app **was healthy and serving** — all four other tests in the same `custom` stage pass, the
install/upgrade stages already asserted the same app healthz/version endpoint, and the App
registry in this very run reports `install/upgrade/backup/restore = pass`;
- so the restart did not break the app; the test's fixed 20 s socket timeout simply expired on one
poll during the app's post-restart window.
### Why this reads as genuinely stale
- This exact test is green on the current 1-safe-line base: the **2026-08-31** run (app 1.27.2 →
1.27.3-rootless) reports `!testme GREEN on the first run`, custom tier included.
- gitea 28.0.0 is a major version jump that substantially refactored git/model layers and boot
wiring; a cold-restart boot + first-request window is exactly the kind of timing the fixed 20 s
timeout (and the comment's own `# Wait for gitea to come back up` TODO) is brittle to.
- Per `tests/STYLE.md` §5, "derive waits from the recipe's declared readiness, not a guess" — this
loop hard-codes `timeout=20` per attempt instead of using the recipe's declared `HTTP_TIMEOUT`
(600 s) / `READY_PROBE`. That is the test's bug, not the recipe's.
- Nothing in the recipe needed to change for it: the dire dev deploy of this same PR head
converged, and admin + repo API + `/api/v1/version` all returned 200 on 28.0.0.
### What I did NOT do (DEFAULT mode)
I did **not** modify the test or the harness. `!testme` cannot go fully green on this PR without a
test change, and test changes are gated behind `--with-tests`. So this PR is recorded as
**`SUCCESS-PENDING-TESTS`** in DEFAULT mode.
### Operator options
1. **Re-run `/recipe-upgrade gitea --with-tests`** — that will open a verified cc-ci test PR to
replace the brittle fixed-timeout restart poll with the recipe's declared readiness
(`HTTP_TIMEOUT` / `READY_PROBE`, per `tests/STYLE.md` §5). This PR's `!testme` then goes green
once the test PR is merged.
2. Treat the run as a **flake** and simply re-post `!testme` on this PR (within the budget) — if the
restart window is kind, the same test passes; but the fixed-timeout brittleness remains.
The recipe upgrade itself (single line: `gitea/gitea:1.27.3-rootless` → `28.0.0-rootless`) is
verified live on the cc-ci dev swarm and by the install/upgrade/backup/restore/lint tiers. NOT merged.
Diagnosis (from run #31): the failure was in the test's own post-restart wait — _api() caught only HTTPError, so one urlopen(timeout=20) socket timeout during the 28.0.0 migration-window restart raised out of the poll loop instead of being retried; the poll also hard-coded a 120s deadline / 20s request timeout while recipe_meta declares HTTP_TIMEOUT = 600 (STYLE.md §5). The app itself was fine — install/upgrade/backup/restore + every other custom test passed on run #31, and 28.0.0 served /api/healthz + authed /api/v1/version on a separate dev deploy.
Fix (assertions untouched): _api() is now never-raising (mirrors lifecycle.http_fetch → transient errors return (0, {}) and the bounded poll retries); poll deadline derives from recipe_meta.HTTP_TIMEOUT (600).
Verification: with PR #44 checked out on the CI host, the full cold suite against this PR's head (91e42c8) passed 5/5 tiers with test_lfs_roundtrip green (incl. the restart + wait) — evidence in the #44 thread.
Dependent pair: !testme on this PR goes green only after #44 is merged (it runs the deployed cc-ci tests). Operator: merge #44, then re-!testme here. Merge of #10 remains yours; recommended release after upstream merge: abra recipe release gitea -x.
Stale-test fix PR is up and verified: https://git.autonomic.zone/recipe-maintainers/cc-ci/pulls/44
Diagnosis (from run #31): the failure was in the test's own post-restart wait — `_api()` caught only `HTTPError`, so one `urlopen(timeout=20)` socket timeout during the 28.0.0 migration-window restart raised out of the poll loop instead of being retried; the poll also hard-coded a 120s deadline / 20s request timeout while `recipe_meta` declares `HTTP_TIMEOUT = 600` (STYLE.md §5). The app itself was fine — install/upgrade/backup/restore + every other custom test passed on run #31, and 28.0.0 served `/api/healthz` + authed `/api/v1/version` on a separate dev deploy.
Fix (assertions untouched): `_api()` is now never-raising (mirrors `lifecycle.http_fetch` → transient errors return `(0, {})` and the bounded poll retries); poll deadline derives from `recipe_meta.HTTP_TIMEOUT` (600).
Verification: with PR #44 checked out on the CI host, the full cold suite against this PR's head (`91e42c8`) passed 5/5 tiers with `test_lfs_roundtrip` green (incl. the restart + wait) — evidence in the #44 thread.
**Dependent pair: `!testme` on this PR goes green only after #44 is merged** (it runs the deployed cc-ci tests). Operator: merge #44, then re-`!testme` here. Merge of #10 remains yours; recommended release after upstream merge: `abra recipe release gitea -x`.
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.
Upgrade the
gitearecipe app image28.0.0-rootless→28.1.0-rootless(upstream v28.1.0, released 2026-10-06 — the first maintenance release of the 28 series).What changed
28.0.0-rootless→28.1.0-rootlessdbpostgres overlay15.19— HELD. The only available bumps are majors (16.x/17.x/18.x); this recipe pins plainpostgres(not pgautoupgrade) and ships no pg_upgrade tooling (pg_backup.shis logical pg_dump/psql only), so a major bump breaks every in-place upgrade. No within-major 15.x bump exists (15.19 is the latest 15.x; Docker Hub 15.20 → 404).dbmariadb overlay10.11.19— current (Docker Hub 10.11.20 → 404) — no bump.app.ini.tmpl/ compose-overlay changes required.Upstream release notes: app gitea/gitea 28.0.0→28.1.0: https://github.com/go-gitea/gitea/releases/tag/v28.1.0
Vendor security write-up: https://blog.gitea.com/release-of-28.1.0/
Operator Action Required
READ COMMITTEDtransaction isolation (#39506) and MSSQL toREAD_COMMITTED_SNAPSHOT. Standard container default; no config change needed.laxmode when explicitly allowed. This recipe sets no egress policy (defaultlax) ⇒ no action.ALLOWED_DOMAINS/BLOCKED_DOMAINS/ALLOW_LOCALNETWORKSdeprecated in favour ofALLOWED_HOST_LIST/BLOCKED_HOST_LIST+EGRESS_MODE; theexternaldeny preset is removed. This recipe sets none of them.RUN_RETENTION_DAYS(#38855): new[actions] RUN_RETENTION_DAYS(0 = keep forever). Defaults apply.Security — advisory scan (deterministic) + vendor notes, window 28.0.0 → 28.1.0
9 CVEs fixed by this upgrade (3 high, 6 medium) — the deterministic GitHub-advisory scan and the vendor blog agree exactly (all fixed via #39501/#39507):
[mirror] DISABLE_NEW_PUSH0⇒ downloader loop / memory exhaustion[repository] FORCE_PRIVATE1 further advisory the scanner could not auto-judge (CVE-2026-62925, medium) is affected
<= 1.27.2and was already fixed in 1.27.3 (--adjudicateevidence: advisory text "Fixed in Gitea 1.27.3 (#39026, #39029)"), i.e. it predates this window — not counted. Scan sources: 25 (20 registry URLs + advisory APIs); no failed sources.Version bump (operator's final step — the PR does NOT touch the label)
28.0.0 → 28.1.0 is a minor upstream release (no BREAKING) ⇒ release as a minor:
Verification
--chaosdeploy on cc-ci (dev-gitea.ci.commoninternet.net, mariadb overlay, chaos versionf2b5089, 2026-10-16): converged1/1app (gitea/gitea:28.1.0-rootless) +1/1db (mariadb:10.11.19); app logGitea version: 28.1.0 built with go1.27.1; DB collation preprocess ran + fresh schema initialised (117 tables);/api/healthz200. Dev deploy torn down (0 leaked stacks/volumes/env).warm-giteauntouched.!testmeon this PR (existing tests, full suite) — verified in the PR comments/runs on the PR head.cc @trav @notplants
Tested green on the cc-ci recipe CI server (full suite, cold, against this PR head). NOT merged — for operator review.
cc @trav @notplants
!testme
🌻 cc-ci —
gitea@91e42c8a❌ failurefull logs · dashboard
!testmeis RED — but only on a stale LFS custom test (recipe upgrade itself verified GREEN on every lifecycle tier)The upgrade itself looks correct. The cc-ci run (
#31, on91e42c8a) is:Three of the four custom tests pass (
test_admin_api_user_org_token_lifecycle— admin API CRUD,test_git_push— full SCM push round-trip,test_gitea_root_returns_200). The one failure istest_lfs_roundtrip, and it fails in the test's own restart polling loop, not in the LFSround-trip or the recipe.
What the failing test asserts vs. what actually failed
The traceback is a
TimeoutError: The read operation timed outon that_api()call(
urllib.request.urlopen(..., timeout=20)) — i.e. the single/api/v1/versionrequest stalled~20 s and raised, aborting the whole test. Crucially:
_api()from inside the cc-ci runner/harness process on the cc-ci host(it shells out to
docker service updatedirectly), so this poll is a local HTTP call to theapp's public
/api/v1/version— it is not the app's own healthcheck;customstage pass, theinstall/upgrade stages already asserted the same app healthz/version endpoint, and the App
registry in this very run reports
install/upgrade/backup/restore = pass;poll during the app's post-restart window.
Why this reads as genuinely stale
1.27.3-rootless) reports
!testme GREEN on the first run, custom tier included.wiring; a cold-restart boot + first-request window is exactly the kind of timing the fixed 20 s
timeout (and the comment's own
# Wait for gitea to come back upTODO) is brittle to.tests/STYLE.md§5, "derive waits from the recipe's declared readiness, not a guess" — thisloop hard-codes
timeout=20per attempt instead of using the recipe's declaredHTTP_TIMEOUT(600 s) /
READY_PROBE. That is the test's bug, not the recipe's.converged, and admin + repo API +
/api/v1/versionall returned 200 on 28.0.0.What I did NOT do (DEFAULT mode)
I did not modify the test or the harness.
!testmecannot go fully green on this PR without atest change, and test changes are gated behind
--with-tests. So this PR is recorded asSUCCESS-PENDING-TESTSin DEFAULT mode.Operator options
/recipe-upgrade gitea --with-tests— that will open a verified cc-ci test PR toreplace the brittle fixed-timeout restart poll with the recipe's declared readiness
(
HTTP_TIMEOUT/READY_PROBE, pertests/STYLE.md§5). This PR's!testmethen goes greenonce the test PR is merged.
!testmeon this PR (within the budget) — if therestart window is kind, the same test passes; but the fixed-timeout brittleness remains.
The recipe upgrade itself (single line:
gitea/gitea:1.27.3-rootless→28.0.0-rootless) isverified live on the cc-ci dev swarm and by the install/upgrade/backup/restore/lint tiers. NOT merged.
Stale-test fix PR is up and verified: recipe-maintainers/cc-ci#44
Diagnosis (from run #31): the failure was in the test's own post-restart wait —
_api()caught onlyHTTPError, so oneurlopen(timeout=20)socket timeout during the 28.0.0 migration-window restart raised out of the poll loop instead of being retried; the poll also hard-coded a 120s deadline / 20s request timeout whilerecipe_metadeclaresHTTP_TIMEOUT = 600(STYLE.md §5). The app itself was fine — install/upgrade/backup/restore + every other custom test passed on run #31, and 28.0.0 served/api/healthz+ authed/api/v1/versionon a separate dev deploy.Fix (assertions untouched):
_api()is now never-raising (mirrorslifecycle.http_fetch→ transient errors return(0, {})and the bounded poll retries); poll deadline derives fromrecipe_meta.HTTP_TIMEOUT(600).Verification: with PR #44 checked out on the CI host, the full cold suite against this PR's head (
91e42c8) passed 5/5 tiers withtest_lfs_roundtripgreen (incl. the restart + wait) — evidence in the #44 thread.Dependent pair:
!testmeon this PR goes green only after #44 is merged (it runs the deployed cc-ci tests). Operator: merge #44, then re-!testmehere. Merge of #10 remains yours; recommended release after upstream merge:abra recipe release gitea -x.!testme
🌻 cc-ci —
gitea@91e42c8a✅ passedfull logs · dashboard
!testme
🌻 cc-ci —
gitea@91e42c8a✅ passedfull logs · dashboard
chore: upgrade to 28.0.0-rootlessto chore: upgrade to 28.1.0-rootless!testme
🌻 cc-ci —
gitea@f2b5089c✅ passedfull logs · dashboard
!testme
🌻 cc-ci —
gitea@f2b5089c✅ passedfull logs · dashboard