Tasks can execute on dedicated compute instances with microVM support

nicktrn

·Mar 29, 2026·#3114feat(supervisor): compute workload manager

A new compute workload manager routes runs to a dedicated gateway, enabling template pre-warming, fast checkpoint restorations, and delayed snapshots.

The platform can now route task execution through a dedicated compute gateway, unlocking microVM-based workloads. Previously, execution was limited strictly to Docker and Kubernetes environments.

Tasks are now pre-warmed during the deployment phase, creating templates ahead of time for faster boot sequences. When tasks suspend, a new delayed snapshot system waits to see if the waitpoint resolves quickly before committing to a snapshot. This saves unnecessary computational overhead on short pauses. The compute gateway handles instant checkpointing and restoration of running instances, emitting nanosecond-precision OpenTelemetry spans to track exactly where time is spent.

These capabilities are wired directly through the supervisor daemon and the web deployment pipeline, backed by a new shared internal compute package.

View Original GitHub Description

Adds the ComputeWorkloadManager for routing task execution through the compute gateway, including full checkpoint/restore support, OTel trace integration, and template pre-warming.

Changes

Compute workload manager (apps/supervisor/src/workloadManager/compute.ts)

Routes instance create, snapshot, delete, and restore through the compute gateway API
Wide event logging on create with full timing and context
Configurable gateway timeout, auth token, image digest stripping

Compute snapshot service (apps/supervisor/src/services/computeSnapshotService.ts)

Timer wheel for delayed snapshot dispatch (avoids wasted work on short-lived waitpoints)
Configurable dispatch concurrency limit (COMPUTE_SNAPSHOT_DISPATCH_LIMIT)
Snapshot-complete callback handler with suspend completion reporting
Trace context management and OTel span emission for snapshot operations

OTel trace service (apps/supervisor/src/services/otlpTraceService.ts)

Fire-and-forget OTLP span emission for compute operations (provision, restore, snapshot)
BigInt nanosecond conversion preserving sub-ms precision for span ordering

Template creation (apps/webapp/app/v3/services/computeTemplateCreation.server.ts)

Three-mode rollout: required (MICROVM projects), shadow (feature flag / percentage), skip
Integrated into deploy finalize flow

Shared compute package (internal-packages/compute/)

Gateway client with namespace-based API (instances, templates, snapshots)
Zod schemas for all gateway request/response types

Database

COMPUTE variant added to TaskRunCheckpointType enum
WorkloadType enum and column on WorkerInstanceGroup
hasComputeAccess feature flag

Env / config

Compute gateway URL, auth token, timeout
Snapshot enable flag, delay, dispatch limit
Dedicated OTLP endpoint for compute spans (COMPUTE_TRACE_OTLP_ENDPOINT)