Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence — Monday

November 16, 2026

🔥 What Changed

A quiet weekend, so two items. A fresh Go runtime report (#83101): since Go 1.26, sending SIGQUIT to a -race or -asan binary to get a goroutine dump can lose the dump. Instead the process hangs on darwin or segfaults on linux. A runtime CL is already up. For Worth Your Time, a new net/http proposal (#83106) asks for a minimum-data-rate RateHandler to stop slow clients from tying up server resources. It's a good, compact review of what Go's server timeouts do and don't cover.

Also seen: labstack/echo v4.16.5/v5.4.6 fix low-severity GHSA-h6x5-phvw-c55c, where a client could strip X-Forwarded-Proto via the Connection header in the proxy middleware. Go #83103 crypto/tls QUICConn.Close hangs when a QUICResumeSession event isn't drained. Go #83119 html/template panic after a subtemplate escape failure. Go #83118 time.ZoneBounds leap-year TZ-footer bug (1.27 backport candidate). Cyrus SASL ≤2.1.28 SRP authentication-bypass advisory. InfoQ: Uber Secure AI Framework for agents; Cloudflare Private Origins.

Holds: Go #83036 phiopt fix (#83076 open); Go #83068 fix (#83078 open); KAFKA-21491 (open, no fix version); Go 1.27.5/1.26.12 (CVE-2026-97047 #82948, CVE-2026-97044 #82374 backports); Keycloak CVE-2026-120984 fix; OTel-Go v1.47.0 adoption; Kafka 3.9.3 (code freeze 11-20); Postgres 19 GA ~11-19.

Go #83101: since Go 1.26, SIGQUIT on a -race or -asan binary can lose the goroutine dump and hang (darwin) or segfault (linux)

The reporter sends SIGQUIT 100 ms after start to a child process built with -race that keeps starting bursts of goroutines, and counts the outcomes. Instead of the usual all-goroutine dump, the child sometimes dies with fatal error: unsafe.Slice: len out of range or hangs (darwin/arm64), or segfaults (linux/amd64). The report lists go1.27.4 darwin/arm64 and go1.27.2 linux/amd64 as affected and reproduces it at tip (2026-11-14). Go 1.26.0 is affected and Go 1.25.0 is not. On darwin, the signal was seen to land while the race runtime was in a kernel call made as a goroutine started or exited. Gopherbot has linked CL 863961, "runtime: don't hexdump a frame that is not on the stack". The issue is labelled compiler/runtime with no milestone yet. The reporter says AI assistants helped with the investigation and the patch.

Why you care: Go is your top board weight. kill -QUIT (or Ctrl-\) is the standard way to see where a stuck Go service or test is blocked, and -race builds are exactly what runs in CI when a test hangs on a deadlock. On Go 1.26+, a missing or garbled dump from a race build may be this runtime bug rather than your code. Until the CL lands, rerun without -race, or use GOTRACEBACK=all with a test timeout (go test -timeout already prints all goroutines).
Read Go issue #83101 → Also: go.dev →

📚 Worth Your Time

Proposal #83106: net/http request and response data-rate enforcement (a minimum-progress RateHandler against slow clients)

Today http.Server.ReadTimeout/WriteTimeout apply to every handler, which is awkward when some endpoints, such as uploads, legitimately run long. ReadHeaderTimeout covers only the headers, and since Go 1.20 per-handler deadlines through http.ResponseController have to be wired into each handler by hand. None of them makes sure a client keeps making progress: a client sending 1 byte under a 30 s read timeout still pins a connection or stream. The proposal adds a stdlib RateHandler with Rate (bytes/s) and Grace (a ramp-up period before enforcement), possibly split into header and body rates, to sit alongside http.TimeoutHandler. Prior art is cited: Kestrel MinDataRate (default 240 B/s with a 5 s grace), Apache mod_reqtimeout/mod_qos, freenginx client_body_min_rate/send_min_rate, and Caddy idle timeouts with a rate. The proposer found nothing equivalent in Gin, Chi, Echo or Iris. A first reply (seankhliao) pushes back: you can do this by wrapping the net.Conn from the listener, and there's little demand. In other words, it's likely to stay outside std.

Why you care: Go services are often exposed directly because the TLS/HTTP stack is mature, and slow-client resource exhaustion is the classic gap between "has timeouts" and "is protected". It ties into your context.Context/deadline-ownership theme. The thread is a fast checklist for any public Go server: ReadHeaderTimeout, per-route ResponseController deadlines, MaxBytesReader, and a floor-rate conn wrapper if you need one.
Read proposal #83106 →