v26.44.0 release notes

[bosconi]

The v26.44.0 release-note train, one commit per snapshot. Dates in RC snapshots are provisional until the release ships; the final snapshot sets the real dates.

rc.1 adds the ## v26.44.0 section — 6 Improvements, 3 Agent Skills, 4 Bug Fixes — and the v26.44 row in the self-managed operator compatibility table. No Features section: nothing in this snapshot's 144 in-scope PRs rises to one. The Agent Skills section is back after five empty windows — the window (2026-09-23T16:49:13Z, 2026-09-25T02:18:32Z] holds 4 PRs and all 4 are published.

final sets the real dates and adds no new bullets. Nothing shipped on the materialize side between v26.44.0-rc.1 and the v26.44.0 tag: GET /compare/v26.44.0-rc.1...v26.44.0 is ahead_by: 1, and that one commit is the version bump, so every rc.1 classification and all four verified flag defaults still hold at the ship tag. This train has only one RC (no v26.44.0-rc.2 tag exists). Its two changes are:

• Real dates: Cloud 2026-09-30, Self-Managed 2026-10-01, from the v26.44.0 tag's commit date (2026-09-30T10:50:59Z) and +1 day — one day earlier than the provisional pair rc.1 carried, exactly as rc.1 predicted. Both date lines in _index.md and the Release date: in the YAML are updated.
• One cross-snapshot rewrite, not a new bullet: agent-skills#74 finishes the mz- rename that #68 left materialize-docs out of. Rather than publishing a second Agent Skills bullet saying skills were renamed to mz-, materialize-docs is now mz-docs is folded into the existing mz- skill names bullet as another example. The rc.1 snapshot directory is left as the record of what was reviewed at the time.

The final commit preserves the two Features a human editor added to this branch — SCIM group-to-role mapping is generally available and Operational dashboards for Self-Managed — and leaves the Improvements and Bug Fixes sections untouched.

⚠️ Reviewer checks

• The hedged-reads entry has an open testing caveat — still open, and the one entry to cut if it stays open. persist: turn on hedged blob gets by default #39110 flips persist_blob_hedged_get_enabled to true in the code default, which is what carries the capability to Self-Managed — Cloud has served it on through LaunchDarkly since 2026-09-24. But the PR states that real GCS and Azure object stores are not yet tested (CI covers MinIO and Azurite) and that testing under PER-120 "should finish before this reaches self-managed users." Nothing in the final increment resolves this, and Linear was not reachable from that run. If PER-120 has not closed, cut this entry before merging. It is the only conditional entry in the release.
• The balancerd connection-limit entry changes the meaning of something already published. balancerd_max_connections shipped an entry in v26.42.0 rc.2 describing over-limit connections as "rejected with an error". balancerd: count connections from accept and bound the time to resolve one #38854 removes that: the TOO_MANY_CONNECTIONS response and the HTTPS 503 are gone, and a client over the limit now sees libpq's "server closed the connection unexpectedly". Anyone who built alerting on the v26.42.0 wording needs this.
• Two PRs are excluded as already published by the v26.43.0 train — cluster-controller: gate graceful cut-over on lag, not hydration alone #38615 and SQL-716: promote mz_session_role_memberships to mz_catalog for restricted role filtering #39029. Both merged to main inside the rc.1 window and were cherry-picked onto the v26.43 release branch, where that train's rc.2 snapshot published them. Accounted for in omitted.md under "Already published in an earlier release".
• This branch will need a rebase if the v26.43.0 docs PR merges first. _index.md here still jumps from ## v26.44.0 to ## v26.42.0 — the v26.43.0 section is absent because that train's PR (#38928) has not merged. Whichever of the two merges second needs the rebase to get the sections in version order.
• Resolved: rc.1's provisional dates were flagged as very likely one day high. They were — the final snapshot corrects them down by one day, on the tag date.

Inclusion rate is 9.7% (14/144) at rc.1, inside the usual 5–15% band, and 20% (1/5) at final, which reads high only because the denominator is five. The rc.1 denominator is dominated by dependency automation — 91 of the 142 materialize PRs are dependency bumps, 65 of them a single Dependabot fan-out of the materialize-terraform-self-managed modules across five test/terraform directories. Strip those and the materialize side is 10 included out of 49.

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: 7 — review borderline calls, final: 0 — none.

The rc.1 borderline calls are unchanged by seeing the whole release, since final adds no materialize PRs. Two are still worth a reviewer's eye: #39125 (the local kind install guide moving to RustFS) — the documented Self-Managed local-kind install fails today for every published tag, because their minio.yaml references a Docker Hub repository MinIO deleted, and this is the likeliest restoration; and #39034/#39066 (the catalog half of the Console cluster-utilization memory and swap columns) — if that Console PR shipped in this release, it belongs in the notes as a Console improvement. Three further items are gated on flags (enable_upsert_v2, enable_compute_index_peek_offload, enable_unified_cluster).

[maheshwarip]

@val-materialize could you include a screenshot here for the upgrade dashboards? Feel free to use one from our internal context graph installation; I think @sjwiesman had one

[maheshwarip]

approved, let's merge in once @val-materialize adds the release notes for advanced SSO