Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence β€” Sunday

October 25, 2026

πŸ”₯ What Changed

A quiet Sunday: a Go 1.27 HTTP/2 regression that silently caps "unlimited" stream limits at 250 (cloudflared hit it in production), and the Go and mesh angle on this week's Brotli decoder CVE. For the weekend: Fly.io on swapping a three-ledger filesystem for a block device on a bucket.

Holds: Kafka 4.4.0 (website PR #928 still open, no announcement) and 4.3.2 (PRs #927/#929 open). Go #82374 x/net/idna (no x/net tag after v0.61.0, no point release). grpc-go v1.85 is still unreleased. The Go backports #82189, #82190 and #82297 are still open. No root-cause analyses yet from GitHub for the Oct 21, 22 or 23 incidents. Google Cloud's final europe-west2 report is still to come. The AWS European Sovereign Cloud exercise ran Oct 24, but AWS hasn't published an outcome report yet. Also seen: Go #82496 (P-384 ECDSA verify about 3.6x slower than AWS-LC), Go #82488, chi XFF spoof GHSA (fixed Oct 12).

Go #82497: the Go 1.27 x/net/http2 server wrapper silently turns MaxConcurrentStreams above MaxInt32 into 250 (cloudflared tunnels hit it in production)

Filed Oct 25, 05:09 IST; not yet triaged. Downstream report: cloudflared #1777 (05:03 IST). With go1.27.3 and x/net v0.61.0, an http2.Server{MaxConcurrentStreams: math.MaxUint32} passed to ServeConn advertises 250 in its initial SETTINGS frame. Values of 0, 250 and MaxInt32 behave as before, but MaxInt32+1 and MaxUint32 both fall back to 250. Building with -tags=http2legacy preserves 4294967295. The cause, which we confirmed in source: x/net's Go 1.27 wrapper converts the field with int(...) in HTTP2Config(), and net/http's internal setConfigDefaults replaces any value outside 1..MaxInt32 with the default of 250. No error or log is emitted, and the public field is still a uint32 with no documented upper bound. cloudflared 2026.10.3 sets MaxUint32 for its HTTP/2 tunnel server, so each tunnel connection is capped at 250 streams. The reporter had to ship a custom http2legacy build to production.

Why you care: If you set "unlimited" stream limits on an x/net/http2 server (proxies, tunnels, long-poll or streaming backends) and moved to Go 1.27, your concurrency ceiling may have quietly dropped to 250 per connection. Long-lived streams then queue behind each other with no error. Fix it today with math.MaxInt32, or build with http2legacy, and add a test that reads the SETTINGS frame.
Read golang/go#82497 β†’

Brotli CVE-2026-111497 ("Node2Shell", CVSS 9.2): the Go port andybalholm/brotli shipped the fix in v1.2.7, and Envoy/Istio quietly fixed it in releases you may already run

An out-of-bounds write in Google's Brotli decoder (versions 1.0.3–1.2.0). DecodeBlockTypeAndLength returned a boolean that meant either "malformed" or "need more input", and the fast path ignored it. A crafted stream can corrupt memory in any process that decodes attacker-supplied br data, whether that's a server reading request bodies or a client reading responses. Fixed in google/brotli 1.2.1 / 1.1.1 / 1.0.10 (Oct 22, 00:57 IST; commit 0ed4268), Node.js 22.23.4 / 24.22.1 / 26.12.1, Python Brotli 1.2.1 / brotlicffi 1.2.1.0, and Bun 1.4.4. Envoy's fix was embargoed inside 1.40.0 / 1.39.3 / 1.38.6 / 1.37.8 (Oct 6, the same releases as the OAuth2 CONNECT fix covered here on Oct 13). ISTIO-SECURITY-2026-007 (1.29.9 / 1.30.6 / 1.31.2, covered Oct 18) lists this CVE; you're affected only if Brotli decompression is enabled, for example via an EnvoyFilter. Go: the pure-Go andybalholm/brotli had the same ignored-return pattern. v1.2.7 (commit 33c2a98, Oct 23, 22:33 IST) ports the upstream check. Because Go checks bounds, the likely worst case is a panic (DoS) rather than code execution, though that is the reporter's guess and hasn't been reproduced.

Why you care: Go is memory-safe, but a decoder panic on untrusted input is still a remote crash. There's no GHSA or Go vulndb entry yet, so govulncheck and Dependabot won't flag andybalholm/brotli ≀ v1.2.6. Run go list -m all | grep brotli and bump to v1.2.7. HTTP clients that auto-decode br (e.g. imroc/req pins v1.2.4) are the outbound path. If you applied the Oct 13/18 Envoy and Istio patches, your mesh is already covered.
Read andybalholm/brotli#82 β†’

πŸ“š Worth Your Time

Fly.io SBD: they replaced a modded JuiceFS (three separately-durable ledgers) with a block device on a bucket β€” a lesson in single sources of truth

Published Oct 22. A Sprite's root disk lives in object storage. v1 ran JuiceFS with a rewritten SQLite metadata backend (kept durable by Litestream), holding a single 100 GB ext4 image on a loop device, plus a separate checkpoints.db. That meant chunks, metadata and the checkpoint catalog were each made durable independently. In writeback mode JuiceFS recorded a chunk in metadata before uploading it, so a crash between the two steps left a pointer to a missing chunk. The result: a 404, then an ext4 checksum failure, then a read-only filesystem. Because checkpoints were copy-on-write clones sharing chunks, every checkpoint had the same gap. Their harness drove 400 Sprites through about 5,000 write/checkpoint/sleep/wake/restore cycles, and a nonzero number permanently lost the ability to checkpoint. SBD exposes the bucket to the kernel as an NBD block device, keeps metadata in the bucket alongside the blocks, and runs sbd serve with a local cache, background upload of dirty blocks and flush on demand. A checkpoint is a versioned block snapshot in SBD's own catalog, and restore rewinds the volume in place. Every block read back is verified, and if the daemon dies, the whole Sprite restarts.

Why you care: This is a short, honest case study of the "metadata before data" crash window and of why every extra independently-durable ledger multiplies the ways state can diverge. Those lessons carry straight over to outbox tables, Kafka offset and DB commits, and snapshot catalogs. The reliability claim ("orders of magnitude") is Fly.io's own, with no numbers given.
Read Fly.io's post β†’