Skip to main content

​
Compatibility and qualification policy

InferCrane follows semantic versioning after 1.0. The complete API, DeploymentSpec, database, integration-contract, deprecation, mixed-version, and support-window rules are in Upgrade and compatibility. PostgreSQL migrations are forward-only and must be backed up before rollout. Mixed-version operation is unsupported unless a release explicitly says otherwise.

​
Compatibility dimensions

Every release candidate must record the tested versions of Go, PostgreSQL, vLLM Router, vLLM, Python, container runtime, GPU driver, CUDA, model, and infrastructure provider. Absence from that matrix means unqualified, not incompatible. infercrane integrations --output json is the executable capability inventory. RunPod elastic/serverless, AWS EC2 elastic, GCP Compute elastic, and Kubernetes elastic are distinct provider profiles; OpenRouter is a governed external target profile. Local conformance and real-provider qualification are separate fields. AWS ASG/EKS/SageMaker/Bedrock, GCP MIG/GKE/Vertex, and CoreWeave CKS have independent registered boundaries but remain non-executable and deferred. The optional Dynamo DGD adapter is locally API-qualified but not real-GPU qualified; llm-d and unregistered external adapters remain unqualified.

​
Check one exact serving combination

There is no universal “supported model” flag. Qualification is the intersection of adapter, runtime, compute mode, runtime version, immutable model artifact, accelerator, and real evidence:
  • integrations proves which provider/runtime/mode combinations are registered and locally or externally qualified.
  • models inspect exposes reviewed immutable configuration and license evidence, not GPU fit or performance.
  • plan checks the requested serving intent without allocating capacity, but cannot manufacture real provider, model-load, or benchmark evidence.
If no release evidence names the complete combination, treat it as unqualified and run the documented real-infrastructure gate. Do not combine several partial “implemented” rows into a production-support claim.

​
Reconcile a partial declarative apply

apply and the control API persist desired state plus a durable operation before provider mutation. Automation must preserve both the exact desired document and its idempotency key:
If the API response is lost or the watcher disconnects, do not submit a new intent. Query the operation. Retrying is safe only with the byte-equivalent desired state and original idempotency key; a conflicting intent must fail rather than reuse the key. The worker re-observes deterministic provider identity and adopts a matching resource before create. For an update, the active revision remains routed while the immutable candidate converges. A failed candidate is rejected and cleaned independently; it does not authorize in-place mutation of active capacity. For a create/update whose provider outcome is ambiguous, compare persisted resource ID, intent digest, ownership tags, revision, and ordinal with direct provider inventory. Exact match may resume; proven absence may recreate from the original intent; mismatch or incomplete inventory stops all create/delete work for manual ownership recovery. CLI, generated SDK, Terraform, and raw API callers share this same operation contract. An experimental client surface does not weaken server-side idempotency or qualify a provider. See Provisioning recovery and the provider outage runbook. The OpenAI-compatible surface is contract-tested for model listing and chat completions. New API fields should pass through unless InferCrane must interpret them. Removing or changing an accepted field requires a deprecation period after 1.0. Endpoint admission applies uniformly across qualified protocols. Buffered requests to managed capacity may use a bounded retry budget; streaming and external paid routes are never replayed. Durable async execution supports Chat Completions, Responses, Embeddings, Completions and bounded chat batch only when the selected runtime declares the corresponding protocol capability. Async does not make an unsupported runtime protocol compatible.

​
Release qualification gates

  1. Unit and PostgreSQL integration tests pass with the race detector and go vet.
  2. The Docker stack smoke test passes, including planning, diagnostics, routing, metrics, and load.
  3. Upgrade and backup/restore drills pass against a copy of production-like data.
  4. Worker loss, router failure, PostgreSQL failover, cancellation, and shutdown are exercised.
  5. A sustained real-vLLM GPU benchmark records throughput, p50/p95 latency, errors, and versions.
Local fake workers validate control flow only. They cannot satisfy gates 3–5 or support performance claims. Repository commands: