Daily Engineering Intelligence

Technology Newspaper

← Back to Home

Engineering Intelligence — Sunday

November 8, 2026

🔥 What Changed

A quiet Sunday with three items. Keycloak has a new CVE that the 26.8.2 batch from Friday does not fix: CVE-2026-120984 (CVSS 9.1) lets a user with manage-clients run server-side code through an imported inline JavaScript authorization policy, and Red Hat lists no fix or mitigation yet, so the practical step today is auditing who holds manage-clients. For Worth Your Time, GitHub published Part 3 of its new Git storage architecture (packs in blob storage, refs in a quorum-replicated compare-and-swap log, disposable cache replicas), and PlanetScale's Shlomi Noach explains how Neki Postgres detects a dead primary and fences it without leases.

Also seen: Go #82917 (opened 11-07 23:38 IST) — _panic.sp resurrects a dead stack argument, ending in fatal error: found pointer to free object; reproduces on go1.26.5 and tip via an unusual Goexit/panic path, sibling of #82394 (fixed for 1.28), no maintainer response yet. KAFKA-21439 (filed 11-08) — after an unclean leader election a share group never delivers new records at already-acknowledged offsets; unconfirmed. Go blog "A new Go API for JSON" (11-04) — encoding/json/v2 stable in Go 1.27. Fly.io postmortem on a kernel checksum-offload bug (CVE-2026-99999) that let VMs read neighbours' traffic; fleet patched in 10 hours.

Holds: Go 1.27.4/1.26.11 (pre-announced Tue 11-10; go.dev/dl still 1.27.3/1.26.10); gopls release for #82821 + #82891 (proxy still v0.24.0); CL 858982 (#82853/#82861) still NEW; Go #82717 MPTCP default; CVE-2026-98432 OpenShift CoreOS still unfixed, no Ubuntu USN; x/oauth2 CL 854305 still NEW; Kafka 3.9.3 now scheduled (code freeze 11-20); Keycloak fix for CVE-2026-120984; moby/swarmkit GHSA pages still 404; NATS 2.15.3/2.14.10 GA; Postgres 19 RC2 ~11-12 / GA ~11-19.

Keycloak CVE-2026-120984: manage-clients can run server-side code via imported inline JS policy — not fixed by 26.8.2

When a client is created or imported with authorization services enabled, Keycloak does not clear an internal setup session attribute, so a user holding manage-clients can get past the block on uploaded JavaScript policies; the server then executes the policy during authorization. CVSS 9.1; Red Hat rates it Important (needs high privileges) but says it breaks realm boundaries and can compromise co-hosted realms. Red Hat Build of Keycloak and RH-SSO 7 are listed Affected, no fix, no mitigation. The fix is not in the 26.8.2 release notes, and there is no upstream Keycloak GHSA or issue yet; no reported exploitation. NVD published 2026-11-07 19:47 IST; GHSA-2539-2x8g-x597 followed at 21:00 IST.

Why you care: Auth/IAM is a rising board category, and this sits right after Friday's 26.8.2 batch, which people may assume closed things out. If you run multi-realm Keycloak, the real control today is who holds manage-clients (including delegated realm admins and service accounts). Hold for the fixed release.
Read Red Hat CVE page → Also: nvd.nist.gov → Also: github.com →

📚 Worth Your Time

GitHub: Inside GitHub's new Git storage architecture (Project Rubicon, Part 3)

Only reference updates need ordering, so pack data lives in Azure Blob Storage while references live in Ref-CAS, a replicated log where a majority quorum commits each per-reference compare-and-swap. Requests are served by disposable "virtual replicas" whose disks are just caches; an unchanged repo costs one version check per read, and compaction runs on separate nodes under leases. It explains rejected designs too: a single-blob log topped out in the mid-hundreds of writes per second, and splitting refs across blobs made reads slow. Self-reported internal testing: 5,000 concurrent agents on one repo, 7,200 pushes/s, "tens of thousands" of Ref-CAS ops/s. Private preview now, broader rollout early 2027.

Why you care: A clean primary-source case study in separating the small strongly-ordered part of a system (refs) from the bulk immutable part (packs), with quorum CAS and stateless cache replicas. Closes the Part 3 hold from the 10-22 brief.
Read GitHub article →

PlanetScale: Designing a reliable unplanned failover (Neki Postgres)

Shlomi Noach's follow-up to the planned-switchover post (reported 10-30). Instead of leases, Neki declares the primary dead only when reachable replicas agree replication broke and none contradict it; detection uses TCP keepalives plus wal_receiver_timeout, which also catches a frozen primary. Fencing is a reversible // prefix on the old primary's host in each replica's primary_conninfo, which stops reconnects and freezes WAL positions so a returning primary cannot commit under ANY 1 sync. It promotes the replica with the most WAL and tears down the old primary rather than rejoining it, and it states its limits honestly, including the commit-loss case under ANY 1.

Why you care: Failure detection, consensus-by-observation, and fencing explained concretely on real Postgres knobs; a good companion to Marc Brooker and Jepsen material on lost writes.
Read PlanetScale article →