v26.46.0 release notes

[bosconi]

Release notes for the v26.46.0 train. Dates in RC snapshots are provisional until the release ships. One commit is added per snapshot; the final snapshot sets the real dates.

• Content: 1 Feature (Metric Sinks), 1 Kafka sink message attribution section (split out of Improvements by an editor on this branch), 9 Improvements, 7 Bug Fixes — all from rc.1.
• final adds no entries. v26.46.0-rc.1...v26.46.0 is ahead_by: 1 and that commit is the version bump, so nothing was cherry-picked onto the release branch after rc.1. The five agent-skills PRs in the window are four automated mz-docs regenerations plus one prose clarification, all omitted.
• Dates: real Cloud 2026-10-15, Self-Managed 2026-10-16 — identical to the provisional pair rc.1 estimated, so no date line changed in this run. The v26.46.0 tag was cut 2026-10-14 (Wednesday); the first Thursday on or after it is 2026-10-15, Self-Managed the day after.
• Operator compatibility: row present for v26.46, release date 2026-10-16.
• Branch updated from main in this run (it was 89 commits behind). The merge conflict was the top-of-list insertion point in both files; resolved keeping v26.46.0 above main's consolidated v26.45.1. Section order is now v26.46.0, v26.45.1, v26.44.1.

⚠️ Open questions for the release owner

• Metric sinks (#39653) — still open at the final tag, and this is the last snapshot before ship. enable_metric_sink is true in src/sql/src/session/vars/definitions.rs, re-verified at v26.46.0, so availability is established — but doc/user still has no page for CREATE METRIC SINK, DROP METRIC SINK, or SHOW METRIC SINKS. Either a docs page lands with the release, or the Metric Sinks section is held for the release that ships its documentation; publishing the heading as-is sends readers to a 404. (Also stale, harmlessly: doc/developer/generated/sql/session/vars/definitions.md still describes the default as false.)
• Bounded persist part decoding (#39587, omitted): persist_part_decode_batch_rows is unset at the tag. If the batch size is being turned on for this release, this is a strong Improvement — the PR measures an OOM-killed hydration becoming 303 s on TPC-H SF100 lineitem.
• SQL Server change-table cleanup (#39463, omitted): behind CDC_CLEANUP_CHANGE_TABLE, which defaults off. Publish if the flag is being enabled.
• Persist per-shard metric cardinality (#39357, #39591, #39601, omitted): adds persist_shard_metrics and friends, but production keeps both, so no scrape shrinks in this release. Should self-managed operators be told the knob exists before their scrape is the one that breaks at ~390k series?
• Nullable-UNIQUE constraint fix and its revert (#39598, #39660, both omitted): net zero shipped behavior, but the planner defect is still open and the near-miss was an environmentd boot panic after upgrade, not a cosmetic one.
• Per-subsystem chunk spill gates (#39659, omitted): shipped as the sole substantive commit of v26.45.1 and reaches main in this window. If it warrants a note, that note belongs on ## v26.45.1 (#39661), not here.
• Kafka sink message attribution (#39566, #39590): confirm the private-preview wording. Settled on this branch — the entry is its own section using the {{< private-preview />}} shortcode and linking CREATE SINK: Headers.

Snapshot drafts (with PR links) in mz-skills:

• rc.1: release_notes.md / omitted.md
• final: release_notes.md / omitted.md

Borderline PRs omitted: rc.1: 10 — omitted.md#borderline, final: 1 — omitted.md#borderline

🤖 Generated with Claude Code