Released Oct 15 (20:11 IST); advisories published Oct 16, 13:30 IST. GHSA-p8wx-ph99-xgcr (High): the Kubernetes Ingress NGINX provider attached the client-certificate TLS option from nginx.ingress.kubernetes.io/auth-tls-secret per Ingress, not per host. If another Ingress declared a path on the same host without the same auth-tls settings, the TLS options conflicted and Traefik fell back to the default TLS options, which don't ask for a client certificate (unless core.strictTLSOptions is enabled). Anyone could then reach backends that were supposed to require mTLS. Affects v3.7.0 to v3.7.14 (the annotation arrived in 3.7); the fix applies auth-tls to the whole host. GHSA-ww7c-q3mv-xx22 (High): with experimental.fastProxy, after a backend answers 101 Switching Protocols the connection pool and the tunnel copier both read the same bufio.Reader with no synchronization; repeatedly opening and aborting upgraded connections corrupts it and panics the whole process. v3.2 to v3.6 are affected and unmaintained, so they get no patch. A Low advisory (GHSA-28hp) fixes a cache hit that skipped the cross-namespace check on CA secrets. Also missed earlier: oauth2-proxy v7.15.5 (Oct 1) fixed two Critical 9.1 bypasses: --trusted-ip with --reverse-proxy trusted a client-IP header without checking the connecting peer (and took the leftmost X-Forwarded-For), and skip-auth routes could be matched by paths that downstream servers resolve differently (%2e%2e, //, ..;).
π₯ What Changed
Quiet Monday after a quiet weekend. Traefik's ingress-nginx provider could silently drop mTLS for a host (and oauth2-proxy had two Critical bypasses on Oct 1); NVIDIA released a permissive 0.63B document parser; Materialize shows how to build a linearizable log on Postgres READ COMMITTED.
Holds (unchanged): Kafka 4.4.0 is listed as released on 15 Oct in the Apache release catalog, but there is still no announcement (kafka-site PR #925 open). grpc-go v1.85 still unreleased (latest v1.84.0). Go #81957 CL 847406 / alt CL 844405 unmerged, backports #82189/#82190 open; #82177 CL 848985, #82026 CL 846325, #81759 CL 840845 and #82272 CLs under review; #81864 tip-only. No new Go point release. New: AWS European Sovereign Cloud cuts its global backbone for a few hours on Sat Oct 24 β waiting for results; Go 1.26 backport of the AES-GCM short-tag out-of-bounds write (#82297, fixed in 1.27) requested.
Released on Hugging Face Oct 17. ~0.63B parameters: a C-RADIOv4 vision encoder plus a 10-layer mBART decoder. One prompt returns markdown text, semantic layout classes, bounding boxes and reading order; tables can be converted to LaTeX, HTML, markdown, JSON or CSV. Weights and code are under OpenMDW-1.1 and the card says it is ready for commercial use; the tokenizer is MIT. NVIDIA's ParseBench table (2,078 pages, one H100 per model): triple-pass 75.20 at 10.62 pages/s, against LightOnOCR-3-4B at 75.10 and 2.76 pages/s; single-pass 72.74 at 15.42 pages/s. Double pass gets most of the gain (+2.18 over single); triple adds 0.28 at ~9% lower throughput. Caveats: all numbers are vendor-run, the triple-pass merge keeps differing full-page/crop alternatives and the reported scores include them, and it targets GPUs with a vLLM plugin or trust_remote_code.
π Worth Your Time
Oct 13, Petros Angelatos. Materialize stores a per-shard log of state diffs in one Postgres table and extends it with compare-and-set. Under SERIALIZABLE, SSI kept hundreds of thousands of SIRead locks; once a transient slowdown hit, every insert got slower and latency went from milliseconds to 30 s, a metastable failure (Bronson et al.). In a repro it held ~4,300 CaS/s, then collapsed. The redesign rests on one invariant: a shard's sequence numbers are always a contiguous interval. Then "predecessor present and the primary key rejects the successor" means "predecessor is the head", so CaS is one statement: a SELECT β¦ FOR KEY SHARE on the expected row plus an INSERT of the next one. The shared lock only keeps truncation from deleting the predecessor mid-statement. A -1 marker row that is never truncated stops stale first writers from recreating a shard. The design is proven linearizable in Lean 4 by refinement, and the mixed-version rollout stayed SERIALIZABLE until a new flag was flipped. Result: ~13,000 CaS/s sustained, and latency now falls when load falls.