Repository navigation
v0.7.0
·
168 commits
to main
since this release
Immutable
release. Only release title and notes can be modified.
Gate flags: breaking, migration
One Tenkai on a bare host. Tenkai installs a single-node cluster as a
privileged host release and delivers releases into it from one control
plane, with the server never running as root (ADR 0041). The release
candidate v0.7.0-rc.1 passed the bare-host drill on its published binaries.
Full operator notes: docs/releases/0.7.0.md.
What's Changed
- chore(release): prepare v0.7.0-rc.1 (#980)
- feat(ops): bare host to cluster to in-cluster release with one Tenkai (#979)
- feat(executor)!: uninstall a host release through an approved plan (#973)
- fix(host): keep CAP_SETUID in the host applier's permitted set (#975)
- feat(executor)!: host executor runs install and health steps through the root-side applier (#970)
- fix(executor): an unavailable executor target leaves the deployment as it was (#972)
- fix(tenkaictl): print the schema type count on init, not every name (#964)
- docs: make the README a five-minute front door (#961)
- feat(release): ship tenkai-host-applier with the Linux release hosts (#965)
- test(connectivity): name the error when a staged bundle import errors (#966)
- feat(host): tenkai-host-applier binary, systemd unit, and polkit rule (#948)
- feat(host): durable applied-state record and generation critical section (#941)
- feat(host): host apply request format and root-side verification (#937)
- fix(providers)!: scope tenant-mode feedback targets under tenkai.feedback_export.v2 (#959)
- test(postgres): give every live test unique tenant and environment names (#957)
- fix(catalog): settle a just-released file lock before refusing as held (#955)
- fix(providers): capture feedback observations in tenant-mode catalog partitions (#954)
- fix(auth): report a committed grant-rule change as done, and fail closed when unsure (#952)
- fix(providers)!: move terminal outcomes to tenkai.terminal_outcome.v2 with tenant-scoped targets (#944)
- fix(catalog)!: sign recall notices and topology drills over the bytes sent (#951)
- test(worker-pool): poll the live fixture instead of sleeping a fixed 80 ms (#950)
- fix(providers): mark a feedback observation delivered only when admitted (#946)
- fix(auth): read stored grant rules one row at a time (#936)
- feat(apply)!: resolve the software executor per release (#938)
- fix(catalog): index a prefix of PostgreSQL property values so large plans store (#942)
- fix(criticality)!: sign derived criticality over the bytes the producer sent (#933)
- fix(postgres): live lease tests do not wait on the runner's disk at commit (#939)
- feat(storage): take a verified pre-migration copy of the SQLite store (#927)
- fix(executor)!: sign runner attestations over the bytes sent and name accepted schemas (#929)
- feat(ops): install Tenkai hosts with one verified command (#928)
- fix(release): stop the Linux host build failing on SIGPIPE from ldd (#930)
- fix(trust): check signer windows at an artifact's signed time (#925)
- feat(manifest): signed deploy.executor field (#922)
- fix(auth): stored grant rules only add to what the grant file grants (#924)
- ci(release): accept signed vX.Y.Z-rc.N tags and publish them as pre-releases (#923)
- fix(release): build Linux hosts at a glibc 2.28 floor and refuse newer (#921)
- docs(capabilities): list server self-upgrade and release binaries as shipped (#918)
- docs(release): complete the 0.6.0 upgrade notes (#919)
- docs(adr): per-release software executor and privileged host releases (#908)
- ci(release): signed-tag verification, SBOM, family release manifest, notes from CHANGELOG (#897)
Upgrade
- Signed statements (#929, #933, #951): runner attestations, derived criticality, recall notices and topology drills move to
.v2schemas signed over the bytes sent..v1is still verified and never produced. Upgrade every server andtenkaictlthat imports derived criticality, recall notices or drill evidence before switching producers to.v2; releases before 0.7.0 refuse.v2as malformed. Runner agents and servers upgrade in either order. Re-sign.v1derived criticality beforetenkaictl bundle build. - Outcome export (#944): register
tenkai.terminal_outcome.v2@2.0.0fromapi/tenkai-terminal-outcome-v2.schema.json(keep.v1; tenant-mode hubs also add target kindtenkai.tenant_deployment), then upgrade withTENKAI_OUTCOME_PROVIDER_REGISTRATIONnaming.v2. A host whose registration still attests.v1@1.0.0refuses to start. Remove the.v1registration once no older replica runs and nothing queued before the upgrade is undelivered. - Tenant-mode feedback export (#959): add target kind
tenant_feedback_observation(keepfeedback_observation), then upgrade withTENKAI_FEEDBACK_PROVIDER_REGISTRATIONusingtarget_kind=tenant_feedback_observation. Rows queued before the upgrade keep their old target. Community hosts are unchanged. - Per-release executor (#922, #938): declare
[deploy] executoronly after every server and runtime that reads the catalog runs 0.7.0. Moving a product to another executor is a remove followed by an install. On a server that setsTENKAI_SOFTWARE_EXECUTOR, management API apply now uses it for undeclared releases, like the reconciler. - Host releases (#937, #941, #948, #965, #970, #973, #975): hosts without
executor = "host"releases need no change. To use them, install the applier withinstall.sh --role server --with-host-applier, add the root-owned trust roots and host approval policy in/etc/tenkai, and setTENKAI_HOST_SPOOL_DIR. Host steps are never auto-approved. A 0.6.0 server refuses plans with host orremovesteps. - In-cluster releases (#972):
env cluster-config setaccepts a path whose file does not exist yet; a release with a missing cluster config is refused before it changes anything and converges once the file appears. - Signer windows (#925, ADR 0042): stored artifacts are checked at their signed time, so
not_afteron a rotated key no longer breaks what it already signed. Live admissions still check at now. Older binaries in a rolling deploy still check at their own clock (they refuse more, never accept more). - Stored grant rules (#924, #936, #952): a stored rule only adds to what the grant file grants. An unreadable or invalid row is withheld and reported (
stored_grant_rules_withheldon/healthz). A write whose outcome is unknown is auditedgrant_rule.*.failed(wasrefused) and makes that replica not ready until it reads again; alerts on*.refusedshould also match*.failed. Do not write a later rule format until every replica runs 0.7.0. - SQLite (#927): the first start that migrates an older schema writes
<db>.pre-v<found>.<ms>.<id>.dbbeside the store; it needs free space about the store's size, or the start is refused with the store untouched. The schema stays 13, so a 0.6.0 store opens without migration. - Hub PostgreSQL (#942, migration): the first use of each tenant schema rebuilds
catalog_object_properties_lookupon a 256-character value prefix. Writes to that tenant's catalog properties wait for it. Rolling deploys and rollback stay correct. - Linux hosts (#921): glibc 2.28+ on x86-64 (0.6.0 needed 2.39). Musl is not supported.
- Rollback to 0.6.0: the SQLite schema is unchanged, so the store opens without a restore. Restore the 0.6.0 values of
TENKAI_OUTCOME_PROVIDER_REGISTRATIONand, on tenant-mode hubs,TENKAI_FEEDBACK_PROVIDER_REGISTRATIONfirst; 0.6.0 refuses to start with the 0.7.0 values. 0.6.0 refuses.v2signed statements, manifests that declare[deploy] executor, and plans with host orremovesteps. - Rust library API (breaking):
ExecutionOptions.software_executorsreplacessoftware_executor;DeploySection.executor,plan::Step.hostandplan::Action::Removeare new; grant-rule store types,CatalogGraphobservation writes and runner attestation signing changed. See the pull requests for exact signatures.