26.4.4rc5

Features

• Add VFolder subpath support in model service session creation (#9149)
• Validate deployment revision creation includes all globally required resource slots (#11580)
• Add RouteSubStatus enum and DB columns (sub_status, updated_at) to track fine-grained provisioning stages; change traffic_status default to inactive so new routes start with traffic disabled until healthy. (#11602)
• Add three-stage PROVISIONING sub-status pipeline (PENDING→STARTING→WARMING_UP→RUNNING) for model service routes with ReplicaID typed identifier. (#11613)
• Add UUID id primary key to domains and scaling_groups tables and demote name to a UNIQUE constraint, plus introduce DomainID and ResourceGroupID identifier types. (#11623)
• Expose subpath on the v2 REST/GraphQL session MountItemInput so EnqueueSession callers can mount a vfolder subpath. Follow-up to #11608, which only exposed subpath on the legacy CreationConfigV*.mount_options wire schema. (#11628)
• Expose the parent deployment directly on the ModelRevision GraphQL Node so any revision (from a mutation result, read query, or connection edge) can be traversed back to its deployment in a single round trip. Adds deploymentId: ID! (raw ID for client-side joins) and a DataLoader-backed deployment: ModelDeployment resolver. (#11631)
• Add --version option to the root backend.ai CLI to display the versions of loaded backend.ai-* packages in the current Python environment. (#11641)

Improvements

• Apply ImageID NewType to ImageRow.id, KernelRow.image_id, and ImageAliasRow.image_id for stronger static typing. (#11588)
• Prune dead model serving code: the unused create_model_service and search_auto_scaling_rules action chains plus the stale ModelServiceCreator, RouteConnectionInfo, and RequesterCtx data containers. (#11590)
• Rename the resource group field on deployment metadata inputs to align with the resource group naming used elsewhere in the deployment API: ModelDeploymentMetadataInput.resourceGroup → resourceGroupName (v2 GraphQL/DTO) and DeploymentMetadataInput.resource_group → resource_group_name (v1 REST DTO; POST /deployments body). (#11600)
• Simplify the container live-stat Prometheus pipeline by replacing implicit metric-classification with explicit per-query results (#11604)
• Extend route coordinator to support sub_status and traffic_status transitions across all lifecycle handler axes. (#11606)
• Unify per-vfolder mount-option DTOs into a single MountOption type and formally declare subpath / mount_destination on the session-creation wire schema (CreationConfigV*.mount_options). The previously separate SDK ExtraMountOption and ExtraMountModel types are removed; both session creation and inference service creation now share the same MountOption. (#11608)
• Split route health Valkey record into ReplicaProbeTarget (probe config) and ReplicaHealthStatus (TTL-based result), removing initial_delay from Valkey and switching to DB-based timeout in check_warming_up_health. (#11632)

Fixes

• Stop the appproxy worker /setup?token=X endpoint from emitting a cacheable redirect: it now responds with 302 Found and Cache-Control: no-store so browsers cannot cache the Set-Cookie-bearing response and lose the ability to reissue the permit cookie after expiry. (#11571)
• Fix TooManyKernelsFound error on multi-node inference sessions by splitting kernel groups into main + sub roles (#11575)
• Fix prometheus safe metric wrappers to catch all exceptions (not just ValueError), preventing mmap IndexError from propagating into business logic. (#11577)
• Restore session creation REST API support for the legacy mounts=['vfname/subpath'] form, including the v1 CLI -v vfname/sub:/dest shape that routes the destination through mount_map. (#11582)
• Fix resource_group being specified per-revision by moving it to deployment metadata so the network assignment is fixed at deployment creation time. (#11583)
• Make AddRevisionInput fields optional except deployment_id. (#11585)
• Expose resource_group on ModelDeploymentMetadata so the resource group selected at deployment creation is visible in detail responses (GraphQL and REST v2). (#11598)
• Expose status, traffic_status, health_status on ModelReplica in the v2 deployment GQL/REST schema. (#11605)
• Make the TUI installer write the halfstack docker compose file as docker-compose.halfstack.current.yml to match the convention used by other dev scripts (start-dev.sh, refresh-graphql-gateway.sh, etc.), fixing silent failures of those scripts on TUI-installed environments. (#11611)
• Remove deprecated scope_type='global' RBAC rows from fixtures and existing databases so role detail queries no longer fail with RBACTypeConversionError. (#11636)
• Fix local scie builds by forwarding SCIENCE_AUTH_GITHUB_COM_BEARER (and SCIENCE_AUTH_API_GITHUB_COM_BEARER) to pants subprocesses so that science binary could download other dependencies (scie-jump, ptex, and PBS artifacts) from GitHub without hitting the per-IP anonymous rate limit. (#11640)

Miscellaneous

• Increase pants lint batch_size from the default 128 to 1024 to reduce the number of spawned ruff processes and their startup overhead. (#11626)

Full Changelog

Check out the full changelog until this release (26.4.4rc5).

Full Commit Logs

Check out the full commit logs between release (26.4.4rc4) and (26.4.4rc5).