v1.7.4

A fix release, and the two worker fixes in it matter to anyone running the memory veto. A worker under a cgroup limit counted every page of file cache its own I/O had touched as memory in use, so a busy worker reported almost nothing free and declined every action that asked for more; free memory is now the limit less the working set, with dirty and writeback pages still counted as used, and the figures are read on both sides so cache that moves mid-read makes the worker report less free rather than more. A worker that declined on free memory while idle could then sit paused waiting for a figure it would never report, and an idle worker that admits work is no longer held on that check. On the store side, an Azure read now stops at the byte count the digest declares and a retry resumes from what was already sent, which ends the oversized reads that failed size verification on large blobs. Worth taking if you enable live_memory_veto or read a CAS backed by Azure Blob Storage. No new features, and nothing to change in configuration.

What's Changed

🐛 Bug Fixes

• Leave reclaimable page cache out of the free memory a worker reports by @corcillo in (#2873) - (db375f0)
• Let paused idle workers accept new work by @corcillo in (#2876) - (7d3d20c)
• Store: bound an Azure read to the bytes the digest declares, and resume a retry by @amankrx in (#2872) - (83c1f81)

🧪 Testing & CI

• (nix) key the Bazel Dev lane to the hermetic toolchain identity so it can hit the warm cache by @b7r6 in (#2874) - (5b75558)
• Build the OCI image on merge-to-main, not on every PR by @b7r6 in (#2879) - (bb3ead5)
• Skip the sanitizer lane on docs-only changes by @b7r6 in (#2878) - (59d8ede)
• Disable the Bazel disk cache when the NativeLink cloud cache is enabled by @palfrey in (#2864) - (b9a0b1e)

New Contributors

• @b7r6 made their first contribution in #2874

Full Changelog: v1.7.3...v1.7.4