Linux 7.2 is landing real work – cache-aware scheduling, a two-line IOPS fix – while Daniel Stenberg draws a line on CVE noise. Google’s data-agent announcement is mostly previews dressed as GA.

1. Linux 7.2 Cache-Aware Scheduling Merged, Finally Handles Multiple Last-Level Caches on Modern CPUs — Phoronix · Jun 15
   Cache-Aware Scheduling landed in Linux 7.2, merged under the CONFIG_SCHED_CACHE Kconfig option after more than a year of development by Intel engineers. The feature co-locates tasks that share data within the same last-level cache domain, reducing cache misses and bouncing on CPUs with complex cache topologies – particularly relevant for high core-count AMD EPYC and Intel Xeon 6 parts. Earlier benchmarks showed meaningful gains on AMD Zen 5 workloads including PostgreSQL, Valkey, and network performance, though the author notes fresh 7.2-specific numbers are still coming. The same scheduling pull also includes SD_ASYM_CPUCAPACITY SMT awareness improvements and better data locality for cfs_rq and sched_entity allocation. It’s opt-in at build time, so distros will need to decide whether to ship it enabled before you see it in production kernels.
2. Daniel Stenberg Is Closing curl’s Vulnerability Queue July 1 Through August 3 — LWN.net · Jun 15
   Daniel Stenberg is closing curl’s vulnerability intake queue from July 1 through August 3 – paid support contracts excepted – citing four months of unusually high pressure from security reports. He’s calling it the “curl summer of bliss” and is openly encouraging other open-source maintainers to do the same. The curl 8.22.0 release slips two weeks to September 2, 2026 as a result. The practical note for security teams: if you’re sitting on a curl CVE to disclose this summer, the coordinated disclosure window just got narrower.
3. Linux 7.2 Gets +5% IOPS on EXT4 and XFS by Reordering Two Lines in IOmap — Phoronix · Jun 16
   Bytedance engineer Fengnan Chang landed a two-line fix in Linux 7.2’s IOmap layer that skips a now-pointless memset in iomap_iter() once iteration completes – the memset was burning memory write bandwidth for nothing. The gain is ~5% IOPS on ext4 and xfs, measured on 4K random-read NVMe workloads using io_uring polling. If you run high-IOPS NVMe storage with either filesystem and io_uring, this is a free win once you’re on 7.2. The same pull also adds VFS infrastructure for FS-VERITY support on XFS via a post-EOF Merkle tree.
4. Redis Monitoring 101: Key Issues and Best Practices — groundcover · Jun 13
   A broad Redis monitoring primer covering the most common performance problems and the metrics worth tracking. The practical bits: low hit rates (check via INFO keyspace_hits/keyspace_misses), large values in sorted sets and hashes triggering cleanup bugs, MOVED errors from multi-key pipeline operations across shards, and lost atomicity when chaining Exists + INCR separately – fixable with a Lua script eval. On the cluster side, using a general-purpose hash key instead of spreading data across multiple keys pushes all lookup pressure to a single node, which becomes a bottleneck at scale. Built-in tools (INFO, SLOWLOG, MONITOR) give you the raw data; the article recommends exporting INFO output to Prometheus for ongoing monitoring. Nothing groundbreaking if you’ve operated Redis in production before, but a reasonable reference if you’re just starting to instrument it.
5. What’s new in data agents: Supercharging your AI workflows — Google Cloud Blog · Jun 15
   Google announced a large batch of data agent features across BigQuery, AlloyDB, Spanner, Cloud SQL, Looker, and Lakehouse – most in preview, a handful now generally available. The headline claims include near-100% accuracy for natural language-to-SQL conversion via QueryData, Managed MCP Servers for five databases now GA, and MCP Toolbox for Databases hitting 1.0. Practically, the release covers three layers: conversational analytics for business users querying databases in plain English, purpose-built agents for data engineers, scientists, and DBAs that automate pipeline builds and database monitoring, and developer tooling to connect agents to enterprise data via MCP. The accuracy claim is vendor-tested against unspecified benchmarks, so treat it as aspirational until you run it against your own schema complexity. Teams already on Google Cloud data products have the most to gain here; anyone evaluating MCP-based agent infrastructure should look at the Managed MCP Server GA as the lowest-friction entry point.

---

// In other news

Cache-aware scheduling, a two-liner that buys 5% IOPS, and Stenberg taking a break — not a bad week. Back Wednesday.