Command Palette

Search for a command to run...

Reactors — GPU Computing Resource Management

Beta notice: Reactors is currently in beta. Some features may change.

1. Overview

1.1 What is Reactors?

Foundry Reactors is an integrated GPU computing resource management platform for AI/ML workloads. Through a web interface you can manage the entire lifecycle — from cluster creation and job execution to model serving.

Core value:

  • Ready to use immediately: Create GPU clusters without building separate infrastructure
  • Multi-cloud support: Integrate diverse environments including AWS, GCP, Azure, and Kubernetes
  • Cost optimisation: Reduce costs with automatic Spot instance usage and Autostop
  • Enterprise-grade security: Role-based access control, audit logs, and built-in security validation

1.2 Key Features

FeatureDescription
Cluster managementCreate / start / stop / terminate GPU clusters, configure Autostop
Managed JobsLong-running jobs with automatic recovery and Spot instance support
SkyServeAI model serving with autoscaling and load balancing
Distributed trainingSupport for PyTorch DDP, Ray, DeepSpeed, and Horovod
Infrastructure managementIntegration with Kubernetes, SSH Node Pool, and Cloud Policy
Volume managementPersistent / ephemeral storage and cloud storage mounts

2. System Architecture

2.1 Overall Composition

Foundry Reactors consists of a user interface, a job queue, workers, and connections to computing backends.

2.2 Job Processing Flow

  1. Request received: User submits a cluster creation request via the web interface
  2. Queue registration: Registered in the job queue according to priority
  3. Distributed lock acquired: Lock obtained to prevent concurrent job conflicts
  4. Job executed: Reactor Worker executes commands through the computing API
  5. State synchronised: Results are saved to the database and reflected in the UI in real time

2.3 Queue Architecture

QueueConcurrencyPurposeKey Jobs
Control2Cluster controllaunch, stop, start, down, autostop
Jobs5Job managementjob.submit, job.cancel, cluster.exec
IO10Input / outputlogs.collect, rsync.sync, output.collect

3. Connection Types

Foundry Reactors provides three connection modes to support a variety of environments.

3.1 Embedded

A ready-to-use option suited for development and test environments

ItemDetail
InfrastructureIncluded in the Foundry package
ConfigurationAutomatic; no separate setup required
Use casesDevelopment environments, PoCs, small teams
AdvantageFast start, simple management

3.2 Managed

An auto-provisioned option for production environments

ItemDetail
InfrastructureAutomatically deployed to your Kubernetes cluster
ConfigurationUpload kubeconfig; provisioning is automatic
Use casesProduction environments, large teams, multi-cloud
AdvantageScalability, high availability, automatic updates

Key characteristics:

  • Automatic namespace creation and isolation
  • Service Account automatically issued (renewed every 30 days)
  • Computing controller automatically deployed

3.3 External

A direct-connection option for leveraging existing infrastructure

ItemDetail
InfrastructureA computing server you already operate
ConfigurationManually enter an endpoint URL and token
Use casesReusing existing infrastructure, special network environments
AdvantageFlexible configuration, leverage existing investment

4. Cluster Management

4.1 Cluster Lifecycle

StateDescriptionCost incurred
INITCluster being createdYes
UPRunningYes
STOPPEDStopped (storage retained)Partial
TERMINATEDFully terminatedNo

4.2 Cluster Operations

OperationDescriptionTime required
LaunchCreate and provision a new cluster5–15 min
StartRestart a stopped cluster1–3 min
StopStop cluster (storage retained)1–2 min
DownFully terminate cluster1–2 min

4.3 Autostop

Automatically stops idle clusters to reduce costs:

PresetIdle timeRecommended use
Development30 minDevelopment / testing
Production2 hoursProduction workloads
CustomUser-definedSpecial requirements
DisabledNoneAlways-on operation (not recommended)

4.4 Spot Instances

Leverage Spot / Preemptible instances from cloud providers to reduce costs.

Recovery strategyDescription
FAILOVERSwitch to On-Demand on interruption
ALWAYS_SPOTAlways maintain Spot instances
EAGER_NEXT_REGIONRetry Spot in another region

5. Job Management

5.1 Managed Jobs

A job type with automatic recovery for long-running workloads.

Key characteristics:

  • Automatic checkpointing and recovery
  • Automatic restart on Spot instance interruption
  • Real-time progress monitoring
  • Detailed log streaming
StatusDescription
PENDINGWaiting to run
STARTINGAllocating resources
RUNNINGCurrently executing
SUCCEEDEDSuccessfully completed
FAILEDFailed (check logs)
CANCELLEDCancelled by user
RECOVERINGRecovering after failure

5.2 Cluster Jobs

One-off jobs run directly on a cluster.

Use cases:

  • Interactive development and debugging
  • Running short scripts
  • Environment setup and package installation

5.3 Job Pool (Batch Jobs)

A batch system for efficiently processing large numbers of similar jobs.

Characteristics:

  • Job group management
  • Automatic scaling
  • Automatic retry for failed jobs

6. Service Deployment (SkyServe)

A managed platform for deploying AI models as services.

6.1 Key Features

FeatureDescription
AutoscalingAutomatic replica adjustment based on load
Load balancingAutomatic request distribution
Rolling updatesZero-downtime service updates
Health checksAutomatic status monitoring and recovery

6.2 Service Status

StatusDescription
INITIALIZINGService initialising
READYReady to handle traffic
FAILEDService error
SHUTTING_DOWNShutting down

7. Distributed Training

Supports multi-node distributed training for large-scale AI model training.

7.1 Supported Frameworks

FrameworkDescriptionRecommended use
PyTorch DDPNative distributed data parallelismGeneral distributed training
Ray TrainAuto-scaling distributed trainingLarge-scale experiments
HorovodHigh-performance distributed trainingExisting Horovod users
DeepSpeedMicrosoft optimisation libraryLarge language models

7.2 GPU Allocation Strategies

StrategyDescriptionRecommended when
PackMaximise GPU utilisation per nodeCost optimisation
SpreadDistribute across nodesFault tolerance
BalancedBalanced placementGeneral use

7.3 Example Supported Cloud Instances

CloudInstanceGPURecommended use
AWSp4d.24xlargeA100 x 8Large-scale training
AWSg5.12xlargeA10G x 4Cost-effective
GCPa2-ultragpu-8gA100-80GB x 8Maximum memory
AzureStandard_ND96asr_v4A100 x 8InfiniBand support

8. Infrastructure Management

8.1 Infrastructure Types

TypeDescriptionUse cases
KubernetesConnect a Kubernetes clusterCloud / on-premises k8s
SSH Node PoolSSH-based node connectionOn-premises servers, VMs
Cloud PolicyCloud provider policyDirect connection to AWS, GCP, Azure

8.2 Secret Management

Securely manage sensitive credentials.

Secret typePurpose
KubeconfigAccess a Kubernetes cluster
SSH Private KeySSH access to nodes
Cloud CredentialAccess cloud APIs

Security characteristics:

  • Encrypted with AES-256-GCM
  • Decrypted values are never exposed in API responses
  • Version control and audit logs

8.3 Apply Workflow

Safely apply infrastructure changes.

  1. Validate: Check configuration validity
  2. Apply: Apply all changes atomically
  3. Verify: Confirm the result; rollback is supported

9. Security

9.1 YAML Security Validation

Security checks are performed on user-submitted YAML configurations.

Blocked items:

  • Attempts to access the local file system
  • Access to cloud metadata services (169.254.169.254)
  • Reverse shell patterns
  • References to sensitive environment variables

Validated areas:

  • workdir: local paths are forbidden
  • file_mounts: only allowed cloud storage
  • docker: local build contexts are forbidden
  • Commands: dangerous patterns are detected

9.2 Access Control

Permission levelAllowed operations
OwnerAll operations, including deleting the Connection
AdminManage clusters / jobs, change settings
AccessView, check logs

9.3 Audit Logs

Detailed audit logs are recorded for all major operations.

Logged events:

  • Cluster creation / start / stop / termination
  • Job submission / cancellation
  • Service deployment / update
  • Infrastructure changes
  • Secret creation / modification

9.4 Distributed Locks

A distributed lock mechanism prevents concurrent job conflicts:

  • Redis-based distributed lock: Fast lock acquisition and release
  • TTL management: Appropriate lock timeout per job type
  • Automatic extension: Lock extension for long-running jobs
  • Database fallback: Fallback locking when the cache is unavailable

10. Monitoring

10.1 Dashboard

An integrated dashboard for real-time status monitoring:

  • GPU utilisation overview
  • Cluster status summary
  • Cost overview
  • Recent job history

10.2 Queue Monitoring

Monitor job queue status and performance:

  • Number of queued jobs
  • Throughput trends
  • Failure rate analysis
  • Worker status

10.3 Notifications

Notifications are provided for key events.

EventIn-appEmail
Cluster startedYesNo
Cluster stoppedYesNo
Cluster failedYesYes
Job completedYesYes
Job failedYesYes
Service deployedYesYes
Service failedYesYes

11. Limits

ItemLimit
Connections per organisationMax 10
Volumes per ConnectionMax 100
Infrastructures per ConnectionMax 50
Secrets per organisationMax 100
YAML file sizeMax 1 MB
Cluster name lengthMax 100 characters
Maximum Autostop duration24 hours

12. Glossary

TermDescription
ConnectionConfiguration linking Foundry to a computing backend
ClusterA group of GPU instances
Managed JobA long-running job with automatic recovery
SkyServeThe model serving service
Spot instanceA discounted cloud instance (subject to interruption)
AutostopAutomatic cluster shutdown after an idle period
InfrastructureThe backend infrastructure on which clusters run
VolumePersistent or ephemeral storage
Job QueueThe job queue management system