Decentralized Architecture

P2P Transport Layer

Peer-to-peer communication without centralized infrastructure. Nodes discover, connect, and exchange data directly.

Discovery

Peer Discovery

Nodes discover peers through mDNS on local networks and Kademlia DHT across the internet. No bootstrap server required for LAN deployments. For WAN, a minimal set of bootstrap peers seeds the DHT — after initial contact, the network is self-organizing.

Transport

Multi-Transport Support

P2P negotiates the best available transport: TCP for data center deployments, QUIC for mobile and edge nodes, WebSocket for browser-based dashboards. Encryption is mandatory — every connection uses Noise protocol for authenticated key exchange.

Abstraction

Unified Transport Layer

The query engine and ingestion pipeline do not know whether they are talking to a peer over HTTP or P2P. The transport abstraction layer presents a uniform interface — switch from centralized to decentralized by changing configuration, not code.

IPFS Segment Archival

Compressed segments are archived to IPFS with content-addressed CIDs. Immutable, verifiable, and retrievable from any IPFS node worldwide.

Content Addressing

CID-Based Identity

Every archived segment receives a Content Identifier (CID) derived from its cryptographic hash. The CID is the data — if two nodes archive the same segment, they produce the same CID. Deduplication is automatic and verifiable.

Integrity

Verifiable Retrieval

When a segment is retrieved by CID, the recipient recomputes the hash and verifies it matches. Any corruption — in transit, at rest, or through tampering — is detected immediately. No trust required in the storage provider.

Durability

Distributed Persistence

Pinned segments are replicated across IPFS nodes. Even if the originating VectorScaleDB instance goes offline, archived segments remain available from any IPFS gateway. Cold storage with built-in geographic redundancy.

Differential Privacy on Shared Segments

Share behavioral intelligence without exposing raw data. Laplacian noise is applied before any segment leaves the node.

Mechanism

Laplacian Noise Injection

Before a segment is shared via IPFS or federated exchange, mathematically calibrated noise is added to centroid vectors and metadata. The noise magnitude is tuned to a configurable privacy parameter — lower values mean stronger privacy, higher noise.

Guarantees

Mathematical Privacy Bounds

Differential privacy provides a formal guarantee: an observer cannot determine whether any individual entity's data was included in the shared segment. This holds even against adversaries with unlimited computational power — it is an information-theoretic bound.

Utility

Preserved Pattern Fidelity

At typical privacy settings, shared segments retain enough behavioral signal for cross-cluster pattern matching and template exchange. The noise obscures individual contributions while preserving aggregate behavioral structure.

Edge-Native Deployment

Run VectorScaleDB on edge hardware with full compression, indexing, and query capabilities. Sync to the cloud when connectivity allows.

Footprint

Minimal Resource Requirements

The Rust engine compiles to a single binary with no runtime dependencies. Edge deployments run on ARM or x86 hardware with as little as 512MB RAM. Vector-aware compression means local storage requirements are 200–16,500x smaller than raw data.

Autonomy

Offline-Capable Operation

Edge nodes ingest, compress, index, and query data independently. No cloud connectivity required for core operations. When connectivity is available, segments sync to central clusters or IPFS — when it is not, the edge node operates fully autonomously.

Sync

Selective Segment Sync

Edge nodes do not sync everything. Configurable sync policies determine which segments upload based on anomaly score, regime type, entity priority, or time window. An edge node on a vehicle might only sync anomalous segments and daily summaries.

Trust Scoring & Partition Recovery

Behavioral trust scoring for decentralized nodes, with data-preserving partition recovery when network splits heal.

Trust

Behavioral Trust Scoring

Each node accumulates a trust score based on uptime, data integrity, and behavioral consistency. Low-trust nodes are excluded from sensitive operations. Trust events from quarantined nodes are discarded automatically.

Recovery

Data-Preserving Partition Recovery

When a network partition heals, conflicting data is resolved using integrity verification with higher aggregate trust winning. Conflicting segments are flagged for manual review. No automatic data deletion during recovery — ever.

Revocation

Network-Wide Revocation

Revoked licenses propagate via membership gossip to all nodes. Every node checks the revocation list before any federated operation. Revocations are irrevocable and permanent. Compact set representations minimize propagation overhead.

Edge-Cloud Data Handshake

Edge nodes and cloud clusters negotiate what data matters most. Local fast-access for critical entities, cloud orchestration for the full picture.

Local Cache

Relevance-based local cache

Edge nodes retain the segments most relevant to their local workload — recent entities, high-anomaly segments, and frequently queried patterns stay local for sub-millisecond access.

Backhaul

Cloud orchestration backhaul

Edge nodes periodically sync finalized segments and behavioral templates to the cloud cluster. The cloud aggregates cross-edge patterns, cascade predictions, and fleet-wide anomaly baselines.

Negotiation

Bidirectional priority negotiation

Cloud and edge exchange priority manifests — the cloud tells the edge which entity types and time windows to retain locally, while the edge reports which patterns are accessed most frequently. Both sides adapt their retention policies based on actual usage.

See It in Action

Inspect cluster topology including P2P peers, IPFS archival status, and transport configuration.

# Query cluster topology with P2P node details
curl "https://api.vectorscaledb.com/v1/cluster/status" \
  -H "Authorization: Bearer $TOKEN"

{
  "cluster_id": "edge-fleet-west",
  "topology": "hybrid",
  "node_count": 8,
  "members": [
    {
      "node_id": "cloud-us-west-01",
      "transport": "tcp",
      "address": "10.0.1.10:8080",
      "state": "alive",
      "role": "coordinator",
      "entities": 1847293,
      "shards": [0, 1, 2, 3, 4, 5, 6, 7]
    },
    {
      "node_id": "edge-vehicle-042",
      "transport": "P2P",
      "peer_id": "QmYwAPJzv5CZsnEf18gnKJRv5bKF7...",
      "state": "alive",
      "role": "edge",
      "entities": 4219,
      "last_sync": "2026-03-09T11:58:22Z",
      "sync_policy": "anomalous_only"
    },
    {
      "node_id": "edge-drone-17",
      "transport": "P2P",
      "peer_id": "QmT5NvUtoM5nX3FVGa7Lq4kGTlN...",
      "state": "alive",
      "role": "edge",
      "entities": 891,
      "last_sync": "2026-03-09T12:01:07Z",
      "sync_policy": "daily_summary"
    }
  ],
  "ipfs": {
    "archived_segments": 284019,
    "total_cids": 284019,
    "pinned_nodes": 3,
    "last_archive": "2026-03-09T12:00:00Z"
  },
  "privacy": {
    "differential_privacy_enabled": true,
    "shared_segments_noised": 47291
  }
}