Repository navigation
Conversation
…ts jobs The configurator trusted projects.list to say where an account can run query jobs. That list can leave out a project the account does run jobs in (for one user it was empty), so the picker demanded a separate query project and had none to offer, blocking the connection. A project the list leaves out is now checked with a free dry-run job; if BigQuery accepts it, the project runs its own queries. BigQueryNode.listed becomes runsQueries to match.
|
Review: 3 findings. Posted all three as inline comments. |
Preview:
|
A dry-run probe that failed for a reason other than a refusal (a 5xx, a rate limit) failed the whole lookup or listing of a project BigQuery doesn't list, which worked before the probe. It now answers no, as before, and isn't cached, so the next lookup asks again.
There was a problem hiding this comment.
Superseded by the review of
9f7d1b97.
This review, of 94eba74a
✅ Approved — The earlier WARNING is fixed: a failed dry-run probe now answers no and isn't cached. No new issues were found.
The decision is advisory: this review is a comment and never approves, blocks or requests changes on the pull request.
✅ All checks passed
🔒 Security: ✅ LGTM
The dry-run probe uses the user's own credentials and reads no data. It runs only for project IDs that have already been parsed.
📖 AGENTS.md: ✅ LGTM
No issues found.
📋 Code Quality: ✅ LGTM
#runsQueries now maps any probe failure other than a refusal to false, outside the cached loader. A failure is therefore not cached and gets retried, and listing and lookup of unlisted projects work again. The new workerd test covers a 503 from the probe.
⚡ Performance: ✅ LGTM
The probe runs at most once per project per picker, and concurrent callers share one request. Listed projects are never probed. An unlisted project's first listing costs one extra round trip.
📝 Documentation: ✅ LGTM
No issues found.
Reviewed commit: 94eba74a
The previous commit evicted a failed dry run so a later lookup would ask again, but the node and listing caches kept the fallback, so the same lookup never did. Any failed dry run now answers no and is kept for the picker, like a refusal; reopening the picker asks again.
|
Review: 2 findings. Posted both as inline comments. |
The fake BigQuery now accepts a job in the probed project only as a POST with dryRun: true, so a probe that would run a real, billed job fails the test.
There was a problem hiding this comment.
Superseded by the review of
85570957.
This review, of 9f7d1b97
✅ Approved — No findings. A failed dry run now answers no and is cached once per picker, so listing and lookup of unlisted projects keep working, and a failed probe never grants query access.
The decision is advisory: this review is a comment and never approves, blocks or requests changes on the pull request.
✅ All checks passed
🔒 Security: ✅ LGTM
The dry-run probe (SELECT 1) uses the user's own credentials, reads no data and runs only for project IDs that have already been parsed. A failed probe keeps the "Run queries in" requirement, so an error never grants access.
📖 AGENTS.md: ✅ LGTM
No issues found.
📋 Code Quality: ✅ LGTM
#runsQueries turns every dry-run failure into false inside the cached loader, so each project is asked once per picker. #heldProjects still rejects on failure and is retried. The rename is applied consistently, and the new workerd tests cover the unlisted-project, 503 and single-probe cases.
⚡ Performance: ✅ LGTM
Each unlisted project is probed at most once per picker. Listed projects are never probed.
📝 Documentation: ✅ LGTM
No issues found.
Reviewed commit: 9f7d1b97
|
Review: 1 findings. Posted one inline finding. |
A stalled jobs.insert held up the picker's lookup or listing of a project BigQuery doesn't list for the client's 70-second timeout. BigQueryApi.dryRun takes a timeoutMs, and the picker's probe answers no once 10 seconds pass, as for any other failed dry run.
There was a problem hiding this comment.
Superseded by the review of
78403d2e.
This review, of 85570957
✅ Approved — No findings. A dry run that times out or fails now answers no, and that answer is cached once per picker. Other callers keep their existing timeout.
The decision is advisory: this review is a comment and never approves, blocks or requests changes on the pull request.
✅ All checks passed
🔒 Security: ✅ LGTM
The probe is a SELECT 1 dry run made with the user's own credentials, and only for project IDs that have already been parsed. A failed or timed-out probe keeps the "Run queries in" requirement, so an error never grants query access.
📖 AGENTS.md: ✅ LGTM
No issues found.
📋 Code Quality: ✅ LGTM
dryRun now takes an optional timeoutMs, passed through to callRest, which falls back to the existing request timeout when it isn't given. The picker's probe uses 10 seconds. A timeout is caught in #runsQueries like any other failure, and the cached false means each unlisted project is asked once per picker. The workerd tests check that the probe is a POST with dryRun: true, that unlisted projects are probed once, and that listed projects are never probed.
⚡ Performance: ✅ LGTM
The probe is limited to 10 seconds and runs at most once per unlisted project per picker. Listed projects are never probed.
📝 Documentation: ✅ LGTM
No issues found.
Reviewed commit: 85570957
|
Review: 1 findings. Posted one inline finding. |
…cluded timeoutMs bounded each attempt only, so a 401 that waited on a token refresh (up to 20 s) and then replayed could hold up the picker past its 10 s. The probe now races one deadline whose signal also ends the request, so a replay after the deadline is never sent. BigQueryApi.dryRun takes that signal instead of a timeoutMs.
There was a problem hiding this comment.
✅ Approved — No findings. The dry run's single deadline, the signal passed through to the request, and the cached "no" answer on failure are all correct.
The decision is advisory: this review is a comment and never approves, blocks or requests changes on the pull request.
✅ All checks passed
🔒 Security: ✅ LGTM
The probe is a SELECT 1 dry run made with the user's own credentials, and only for project IDs that have already been parsed. A failure or timeout answers "no", so an error never grants query access.
📖 AGENTS.md: ✅ LGTM
No issues found.
📋 Code Quality: ✅ LGTM
dryRun passes its signal to callRest, so the deadline stops a request in flight and keeps a retry from being sent once it has passed. In #runsQueries the race always settles to a boolean, and a probe that rejects late is still handled. In #node, listDatasets now runs only for unlisted projects, so a project this account can't see still comes back as null. The listed → runsQueries rename is applied everywhere.
⚡ Performance: ✅ LGTM
The probe has a 10-second limit that includes any token refresh, and it runs at most once per unlisted project per picker.
📝 Documentation: ✅ LGTM
No issues found.
Reviewed commit: 78403d2e
|
Review: 1 findings. Posted one inline finding on missing regression coverage for the picker deadline during token refresh. |
A user who can run queries in a project couldn't connect it. The BigQuery configurator decides where an account can run query jobs from
projects.list, which came back empty for them. So the picker called the project one they have "no role in", required a separate "Run queries in" project, and had none to offer.Now a project the list leaves out gets a free dry-run job (
SELECT 1), once per project per picker. If BigQuery accepts it, the project runs its own queries and "Run queries in" stays hidden. Listed projects behave as before and are never probed.BigQueryNode.listedis renamedrunsQueries, since it no longer means "listed".Tested: the gatekeeper-google configurator tests (workerd and unit), and
pnpm lint.