Drain guest responses before scale-to-zero - #369
Conversation
831acb0 to
16df82d
Compare
4a3f7d2 to
616817c
Compare
Sayan-
left a comment
There was a problem hiding this comment.
Static review of the response drain at 616817c. Comment only, no verdict. Items are ranked by severity. The last two entries are scope notes rather than defects. No fixes proposed here; this is intended as input for follow-up work.
| # | Severity | Area | Concern | Where |
|---|---|---|---|---|
| 1 | Medium | Protocol / resource | A peer RST (or TCP_USER_TIMEOUT expiry) on the duplicated, half-closed socket leaves the kernel socket in TCP_CLOSE with TIOCOUTQ still nonzero. waitForClosedResponse requires queued == 0 && acked, so it never releases early and retains the hold, the dup fd and the goroutine for the full 5 minute timeout. The drain is then recorded as timeout although the socket was definitively dead. |
connection.go:511-534, socket_linux.go:16-18 and 43-59 |
| 2 | Medium | Correctness | drainConn.Write sets one absolute deadline per Write call. A WriteTo-backed body (bytes.Reader wrapped in io.NopCloser) reaches the connection as a single Write of the whole body, so a slow but progressing transfer that runs longer than 5 minutes is reset. DownloadDirZip is the affected endpoint. The per 1 MiB refresh exists only in ReadFrom, so the PR description's "rolling" deadline does not hold on this path. |
connection.go:68-93 vs 95-131, fs.go:859 |
| 3 | Medium-low | Observability | The deferred writeError check runs when the middleware returns, before net/http finishRequest, which is where unflushed responses (roughly anything under 2 KiB) first touch the socket. Write failures there surface as connection_closed at debug level and are never counted as write_error. Conversely, a write failure inside the handler is counted twice: connection_closed via abortNow and write_error via the deferred check. |
middleware.go:225-230 and 156-158; go server.go:2109 and 2116 |
| 4 | Low-medium | Correctness | Every non-EOF error from TCPConn.ReadFrom is treated as a write error, including source-side read errors (pipe CloseWithError, sendfile EIO). The abortive linger-0 close purges already-queued response bytes and sends RST where the pre-PR behavior was a graceful FIN with a truncated body. Affects DownloadDirZstd, TakeScreenshot, ReadFile. SSE responses are unaffected because the generated visit method takes the Flusher path. |
connection.go:117-126 |
| 5 | Low | Resource | There is no cap on concurrent close-delimited monitors. Each stalled close holds one fd and one goroutine for up to the timeout. When F_DUPFD_CLOEXEC fails with EMFILE, Close takes the abortive path and RSTs the response tail. This is the aggregate consequence of item 1. |
connection.go:330-349 |
| 6 | Low | Efficiency | drainConn.ReadFrom wraps a second io.LimitedReader around the one io.CopyN already created. sendFile and spliceFrom unwrap a single layer, so http.ServeContent and http.FileServer now fall back to a 32 KiB userspace copy. Only /extensions/* is affected; *os.File bodies still reach sendfile. |
connection.go:116; go net/sendfile.go:27-39 |
Scope notes (intentional or pre-existing, listed for completeness):
- Hijack completes every pending drain immediately, including a previous request's carried-forward drain. Once a WebSocket handler returns, nothing covers an unacked close frame or FIN. This matches the PR description and connection_test.go:24-83. Residual exposure is small: coder/websocket
Closewaits for the peer's close frame before returning, and the cooldown follows release. Same as pre-PR behavior. - Responses that net/http produces before entering the handler (
OPTIONS *, 400/431/501 parse errors, 417 for unsupportedExpect) never acquire a hold. Pre-existing and inherent to middleware placement. chi routes 404/405 through the chain, so those do get holds.
Evidence for item 1 (Linux v6.12): tcp_reset calls tcp_done_with_error, which runs tcp_write_queue_purge and tcp_done; neither touches snd_una or write_seq (tcp_input.c:4509-4550, tcp.c:3254-3268 and 4847-4871). tcp_ioctl SIOCOUTQ zeroes only SYN_SENT/SYN_RECV and otherwise returns write_seq - snd_una (tcp.c:636-644). tcp_write_err takes the same path (tcp_timer.c:75-79), and the TCP_USER_TIMEOUT value equals the poll timeout, so it cannot shorten detection. The keep-alive idle path is fine: an RST arriving before Close makes shutdown return ENOTCONN, which isTerminalConnectionError accepts.
Evidence for item 2 (Go 1.25): io.NopCloser returns nopCloserWriterTo when the reader implements WriterTo (io.go:682-687); bytes.Reader.WriteTo issues one Write for the remaining slice (bytes/reader.go:137-153); both bufio layers pass a large slice straight through once the buffer is empty (bufio.go:679-686); FD.Write loops over the whole slice under the absolute deadline (internal/poll/fd_unix.go:360-401). middleware_test.go:332-365 exercises this shape and asserts that the handler errors.
Checked and found not to be issues: long-lived SSE streams under the 1 MiB chunk deadline (the Flusher path refreshes the deadline per write), and hold leaks from a stalled partial next request (StateActive fires only after a complete header block is read).
91886a1 to
ee50103
Compare
ee50103 to
a1d8298
Compare
Sayan-
left a comment
There was a problem hiding this comment.
thanks for iterating! overall looks good. wanted to confirm if you want to fix these before merging
- p2, the new response-drain metrics do not retain their intended dimensions through production collection. The browser-VM telemetry pipeline drops the
outcomelabel before aggregation, andguest_terminationis incremented immediately before shutdown closes the metrics listener. SigNoz therefore cannot distinguish drain outcomes or reliably observe guest termination, although guest and VMM logs retain separate evidence.
problem
An HTTP handler can return while response bytes remain unacknowledged in the guest TCP send queue. Re-enabling scale-to-zero at that point can suspend the guest before the peer receives the response tail.
change
This implements the fix entirely inside the guest image. It requires no caller or control-plane protocol changes.
TIOCOUTQafter net/http finalizes explicit response framingio.ReaderFrom/sendfile transfers by advancing them once per 1 MiB chunkio.LimitedReaderbounds sohttp.ServeContentandhttp.FileServerretain sendfileshutdown(SHUT_WR), closing the Go connection promptly, and retaining the hold until TCP acknowledges the closeSO_LINGER=0TCP_USER_TIMEOUTsockets as terminal even when Linux retains a nonzero historical output-queue countclose(2), including when close reportsEINTRTCP_USER_TIMEOUT; terminal duplication failures close the original Go connection before releasing its holdhttp.ErrAbortHandler, preventing net/http from synthesizing a successful response for an operation that never ranwrite_erroroutcomeThe release invariant is: response data and required close framing were acknowledged outside the guest, or the connection was abortively terminated. A timeout, inspection failure, or cleanup failure alone never releases the hold.
tests
go build ./...go vet ./...go test -race ./lib/scaletozero ./lib/metrics ./cmd/api/apigo test -race $(go list ./... | grep -v '/e2e$')EINTRis terminal and the descriptor is never inspected or closed againfinishRequestand handler-visible write failures each emit exactly onewrite_erroroutcomeTCP_USER_TIMEOUTfailures reach bounded guest termination without releasing the holdNote
High Risk
Changes core HTTP serving and scale-to-zero lifecycle for all external traffic, with Linux-specific socket recovery that can abort connections or terminate the entire guest on persistent failures.
Overview
Fixes premature scale-to-zero by keeping the guest awake until each external HTTP response is safely finished on the wire, not merely when the handler returns.
Scale-to-zero middleware now acquires a per-request hold at
Disable, releases it only after both the handler finishes and the connection reaches a terminal drain outcome, then callsEnable. It tracks TCP viascaletozero.Serveon the API, DevTools, and ChromeDriver listeners (metrics stay on plainListenAndServe). Untracked connections or failedDisableabort withhttp.ErrAbortHandlerinstead of returning 500.Response draining polls Linux
TIOCOUTQafter net/http goes idle, with adaptive polling, rolling write deadlines per 1 MiB chunk (includingReadFrom/sendfile), keep-alive hold handoff (one pending drain per connection), and hijacked/WebSocket paths that skip tail tracking. Close-delimited responses duplicate the socket,SHUT_WR, and monitorTCP_INFOuntil close is acknowledged; stalled paths use abortive close, capped (256) close monitors, bounded recovery, and guest termination (SIGTERM/SIGKILL to PID 1) if the hold cannot be released safely.Observability: new Prometheus metrics for drain outcomes, active/fail-closed holds, close monitors, and monitor rejections via
ResponseDrainCollector.Reviewed by Cursor Bugbot for commit a76f270. Bugbot is set up for automated code reviews on this repo. Configure here.