Entropy is a partition-tolerant distributed compute fabric designed for unreliable infrastructure.

It provides durable job scheduling, lease-based assignment, deterministic recovery, and eventual consistency under:

• Power loss
• Worker crashes
• Network partitions
• Duplicate messages
• Out-of-order delivery
• Control-plane restarts

Entropy is designed for environments where infrastructure cannot be trusted.

Entropy consists of two roles:

• Durable SQLite-backed state
• Job queue & state machine
• Lease-based assignment
• Worker registration & capability tracking
• Reconciliation engine
• Retry & recovery transitions
• Structured audit events
• Metrics aggregation

• Poll-based job acquisition
• Assignment ACK handshake
• Execution with timeout enforcement
• Artifact generation & checksum validation
• Heartbeat with monotonic sequence numbers
• Local replay buffer for reconnect
• Deterministic completion replay

Entropy provides the following execution semantics:

A job lease (assignment_id) ensures only one worker can legally complete a job.

Concurrent claims are prevented by serializable DB transactions.

• assignment_id validation
• Monotonic completion_seq
• Terminal-state no-op protection

Duplicate completion messages do not corrupt state.

If a worker:

• Crashes
• Loses power
• Stops heartbeating

The control plane transitions:

RUNNING → LOST → RETRYING → PENDING → (reassigned)

Attempt counts increment safely and deterministically.

• Durable SQLite state
• Immediate reconciliation sweep
• Stale leases reclaimed
• No orphaned RUNNING state

Workers may:

• Replay completions
• Replay state on reconnect
• Resume after network partitions

Control plane reconciliation ensures final deterministic state.

Entropy implements:

• At-least-once execution model
• Idempotent completion enforcement
• Lease-based execution safety
• Priority-aware scheduling
• Capability-based assignment
• Scheduled (delayed) job eligibility
• Retry concurrency controls
• High-priority preemption gates
• Queue aging fairness model

Entropy explicitly handles:

Entropy provides:

• GET /metrics
• Structured audit events (/events)
• Worker liveness tracking
• Queue depth by status
• Retry & lost counters
• Completion/failure rate windows
• Lightweight UI dashboard

• Admin vs worker token separation
• Token rotation support
• Optional signed worker tokens
• Optional TLS
• Optional mTLS
• Worker command allowlists
• Working directory allowlists
• Optional non-root enforcement

Entropy is particularly suited for:

• GPU compute in unstable power environments
• Local multi-host clusters
• Home-lab distributed systems
• Academic distributed scheduling experiments
• Infrastructure failure simulations

See:

• ARCHITECTURE.md
• DESIGN_GUARANTEES.md
• SRS.md
• SRS_TEST_SCENARIOS.md

Entropy intentionally avoids:

• Kubernetes
• Distributed consensus systems
• External coordination layers
• Cloud dependencies

It focuses on:

Deterministic state transitions
Lease-based safety
Durable recovery
Minimal moving parts
Explicit failure modeling

Entropy implements the full SRS specification including:

• Lease-based assignment
• Retry lifecycle
• Stale worker recovery
• Concurrency-safe claim logic
• Stress-tested recovery model
• Duplicate completion protection
• Dual-host recovery validation

Entropy is not a demo.

It is a small, production-grade distributed systems engine.