Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence β€” Tuesday

November 17, 2026

πŸ”₯ What Changed

A quiet Tuesday, so three items. The big one: Go 1.27.4 and 1.26.11 refuse to link Rust static libraries through cgo (#83126). The cgo-linker security fix in last week's point releases rejects the LLVM bitcode sections that Rust's std ships. The Go security team calls it a clear regression, a fix CL is up, and backports for 1.27 and 1.26 are open. Also: MSK Express brokers now deliver topics straight to S3 with no Kafka Connect. For Worth Your Time, Meesho on running 33 Strimzi Kafka clusters on a dedicated Kubernetes cluster.

Also seen: Go #83130 ServeMux {v} wildcard doesn't match a segment that is exactly %2F (it does match a%2F). Go #83133 go/types CheckExpr mutates an existing FuncLit scope (data race; NeedsFix). A batch of cmd/go module-resolution bugs (#83137–#83141, e.g. go get mod@none infinite loop). Kubeshark MCP-server advisories GHSA-8vf2-hxg5-crqh (high: argument injection leading to cluster takeover) and GHSA-gj7r-p794-hhq5. No new releases since Sunday evening from Kubernetes, NATS, etcd, grpc-go, Kafka, franz-go, OTel-Go, SpiceDB, Keycloak, ZITADEL, Istio, Envoy, Cilium, Temporal or Vault.

Holds: Go #83126 fix (CL 864082; backports #83136/#83135) β†’ likely in 1.27.5/1.26.12 with the CVE-2026-97047 #82948 and CVE-2026-97044 #82374 backports; Go #83036 phiopt fix (#83076 open); Go #83068 fix (#83078 open); KAFKA-21491 (open); Keycloak CVE-2026-120984 fix; Kafka 3.9.3 (code freeze 11-20); Postgres 19 GA ~11-19.

Go 1.27.4 / 1.26.11 regression: cgo rejects Rust static libraries ("disallowed LTO section"); fix CL up, backports opened

Linking any Rust staticlib that pulls in Rust std via #cgo LDFLAGS: -L… -ladder now fails on go1.27.4 and go1.26.11 with disallowed LTO section __LLVM,__bitcode in #cgo LDFLAGS (darwin) or .llvmbc (linux). It worked on go1.27.3 and go1.26.10. The cause is CL 862151, the fix for security issue #82572 ("arbitrary code execution via crafted .syso files and cgo linker input files", one of the cgo/linker CVEs shipped in the 11-10 point releases). Rust std objects carry an embedded bitcode section, so they trip the new check. Teleport maintainers report the same breakage. Ian Lance Taylor noted there's no escape hatch like CGO_CFLAGS_ALLOW. Roland Shoemaker (Go security) called it an oversight and a clear regression with no reasonable workaround: .llvmbc, .llvmcmd and __LLVM,__bitcode should be safe. He has posted CL 864082 ("allow embedded LLVM bitcode sections and honor CGO_LDFLAGS_ALLOW"), and gopherbot opened backports #83136 (1.27) and #83135 (1.26) at 04:07 IST today.

Why you care: Go is your top board weight, and last week's 1.27.4/1.26.11 was a must-upgrade security release. If any service or CI image links Rust (or other LLVM-built) static libraries through cgo, that upgrade breaks the build, and you can't simply pin back without giving up the 12 CVE fixes. Watch for 1.27.5/1.26.12. The fix also makes CGO_LDFLAGS_ALLOW work as the documented bypass for this check. Pure-Go and gcc-only cgo builds aren't affected by this report.
Read Go issue #83126 β†’ Also: CL 864082 β†’

Amazon MSK Express brokers deliver Kafka topics straight to S3, with no Kafka Connect and no broker throughput used

You create a channel on a topic of an MSK Express cluster (console, MSK API or aws kafka create-channel). MSK then batches records into S3 objects in the producer's original format (JSON, string or raw bytes, no schema registry), with your choice of compression (NONE/GZIP/ZSTD), storage class (Standard, Intelligent-Tiering, Glacier IR) and an output key template (for example date prefixes plus partition and offset in the object name). JSON/string records that fail validation go to a DLQ bucket and delivery continues. Billing is per byte delivered, with no worker fees and no charge for the broker egress or DLQ records. AWS says the channel reads the topic without consuming broker throughput, so it doesn't compete with real-time consumers, and claims up to 60% lower cost than self-managed pipelines. The same topic can also feed streaming tables (Iceberg on S3 Tables). AWS's migration advice: run the channel next to your existing S3 sink connector until the delivered data overlaps, then retire the connector.

Why you care: Kafka is a board-weighted Messaging skill, and "S3 sink connector on a Connect cluster" is one of the most common bits of Kafka plumbing to run. This turns archive, replay-beyond-retention and compliance copies into a managed per-topic setting. It's worth knowing as a design option, and as an interview trade-off: managed delivery vs Connect, and object layout for replay. It's Express-only and AWS-specific, so check supported Regions and configurations in the MSK Developer Guide.
Read the AWS post β†’

πŸ“š Worth Your Time

Kafka at Meesho: 33 open-source Kafka clusters on a dedicated Kubernetes cluster with Strimzi (474 β†’ 594 brokers on sale day, no network change)

Part 1 of Meesho's "Queueing at Meesho: Infrastructure" series. Kafka gets its own Kubernetes cluster that hosts nothing else: brokers, controllers, client proxies, Cruise Control and metrics. Each Kafka cluster gets tainted node pools, with one broker per node so a node failure or provider maintenance costs exactly one broker. One Strimzi cluster operator deployment (three replicas, elected leader) runs all 33 clusters, reconciling drift every two minutes and rolling brokers one at a time. The user operator manages thousands of KafkaUsers, and KafkaRebalance drives Cruise Control after upscales. The networking half starts from the fact that "a Kafka client never talks to the cluster", only to individual brokers: per-broker advertised listeners and Services, one DNS name per broker, routing by port through the ingress-nginx tcp-services map, ports pre-wired for brokers that don't exist yet, a proxy tier per business unit, and a headless path that keeps most traffic off the per-byte load balancer. Peak is over two million client connections.

Why you care: This hits your Kafka, Kubernetes and Distributed Systems board weights at once, from a Bengaluru-based company at real scale. The per-broker addressing and "scaling is just a broker count" design is the part most Kafka-on-K8s write-ups skip, and it's a strong system-design story to be able to explain.
Read the Meesho post β†’