Repository navigation
fix(cli): retry deploy POST on transport failure - #327
Conversation
A transient socket reset (ECONNRESET on a stale pooled keep-alive connection, or any edge/network blip) during the finalize-deploy POST failed the entire deploy even though no response ever arrived — so the platform could not have started processing. Retry the fetch with 1s/5s/10s backoff (max 3 retries). HTTP responses of any status still surface verbatim since a server answer means the request may have been accepted. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6ba0aa8695
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…ed a deployment A rejected fetch does not prove the request never arrived — the connection may have dropped after delivery. Before replaying the non-idempotent POST, check for a deployment created inside a 120s window; finding one means the earlier attempt reached the platform, so the retry is suppressed (reported via a terminal error event naming the deployment) instead of bouncing off the server's 409 concurrency guard or registering a duplicate deploy. The check is best-effort: on lookup failure the retry still runs and the server's guard remains the backstop. Addresses Codex review on PR #327. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8d651891c6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… retry A recency-only match could wrongly attribute an unrelated deploy (a manual or CI deploy racing ours) to the lost attempt and abort the retry. The deployments table stores the payload's git commit, so require the candidate row's commit to equal ours — null===null counts since ownership cannot be disproved without metadata. A commit mismatch means the row is not ours and the POST is still replayed. Addresses Codex review on PR #327. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: df5c1c6e42
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
createDeployment inserts the deployment row as a concurrency claim with commit=NULL; updateDeploymentManifest fills it only after R2 inventory. For a deploy carrying a commit, our own lost attempt therefore reads commit:null during its first seconds and the equality check missed it, letting the replay hit the 409 guard. Now also treat a still-in-flight null-commit row as a possible placeholder when the payload has a commit (an unrelated in-flight deploy would reject the replay with 409 anyway). Failed/canceled rows are dead attempts and no longer suppress. Addresses Codex review on PR #327. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e9a6c42859
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Commit equality cannot prove ownership of a deployment row — same-SHA redeploys and commit-less manual deploys collide — so matching a recent row on commit could suppress a POST that never reached the server, silently dropping the deploy the user asked for. Suppress only when a row created inside the window has a non-terminal status: getActiveDeployment treats any such row as an active deploy and the replay would be rejected with 409 regardless of ownership, so suppression is equivalent to retrying and produces the clearer error. Terminal rows never suppress — a completed attempt instead replays as an identical re-deploy, the benign side of the ambiguity. This also drops the payload-commit plumbing, which no longer discriminates. Addresses Codex review on PR #327. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d526b9924d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Two suppression gaps found on review: - The deployment lookup ran before the backoff sleep, so a placeholder inserted while the lost POST's precheck was still running could be missed entirely and the replay hit the 409 the dedupe exists to avoid. The lookup now runs after the sleep, immediately before the replay. - `cutover`/`compensating` rows are not guaranteed to 409: the precheck can reconcile them (completed compensation → failed, marker-verified cutover → live) and accept the new deploy, so suppressing on them could discard the requested deployment. Suppression now uses an explicit status allowlist (pending/uploading/migrating/prerendering/ canceling) — only rows whose replay is guaranteed redundant. Addresses Codex review on PR #327. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 21271b7a6f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Postgres `timestamp` mode:'string' values (e.g. `2026-09-18 12:00:00`) carry no zone marker, so `Date.parse` read them as local time: in UTC+N a fresh row looked N hours old and the dedupe missed it; in UTC−N a stale row could appear to be in the future and wrongly suppress. Normalize offset-less values as UTC, mirroring the platform's `databaseTimestampMs` helper. Tests now emit the real wire format. Addresses Codex review on PR #327. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 25884e55cf
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
A client-side GET-then-decide dedupe can never be authoritative — the
deployment row can transition between the lookup and the replay, so any
suppression rule carries an unavoidable race that can drop the deploy
the user asked for. Remove the /deployments lookup, status allowlist,
and timestamp-window machinery entirely.
The retried POST now always reaches the server's precheck — the only
atomic, project-locked verdict. A 409 conflict ({error, deployment})
is surfaced as a structured error event naming the deployment id
instead of a bare HTTP failure. Residual: a lost attempt completing
within seconds replays as an identical re-deploy — the benign side of
not being able to prove ownership client-side.
Addresses Codex review on PR #327.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
`parsed = JSON.parse(body) as typeof parsed` let the declared type query resolve to the narrowed `null` under stricter checking, collapsing the variable to `never` and failing CI typecheck (TS2339 at the conflict handling site). Parse into `unknown` and narrow explicitly instead. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Summary
POST /:id/deployhad no retry: a single socket reset (ECONNRESET, observed on oxc-project/website CI after ~3 min of successful R2 asset upload) failed the whole deploy withdeploymentId=null.streamDeployRequestnow retriesfetchrejections (TypeError— no HTTP response received) with 1s → 5s → 10s backoff, max 3 retries.409({error, deployment}— 'deploy in progress' / 'deploy cancellation requested') is surfaced as a structured error event naming the deployment id; if nothing is in flight the deploy proceeds normally. Residual accepted: a lost attempt completing within seconds replays as an identical re-deploy — unavoidable without a server-persisted request id.TypeErrorrejections, e.g. AbortError, fail fast).deploy_fetch_retrywarnings, anattemptscount ondeploy_fetch_rejected, anddeploy_conflicton 409s.Test plan
direct-r2-client.test.ts): retry-then-succeed, exhaustion (exact backoff schedule + original error), 409-with-deployment → error event (first attempt and after-retry), 409-without-id → PlatformApiError, AbortError no-retry, HTTP-503 no-retrydirect-r2-upload,stream-deploy-errors)tsc --noEmitclean,vp lintcleanGenerated with Devin