Distributed Intelligence

Consistent-Hash Sharding

Deterministic shard placement with minimal redistribution when the cluster topology changes.

Deterministic

Zero-Coordination Placement

Given an entity ID and shard count, every node in the cluster independently computes the same shard assignment. No coordination service, no consensus round, no single point of failure for placement decisions.

Minimal Disruption

O(1/n) Redistribution

When a node joins or leaves, only 1/n of keys need to migrate — the theoretical minimum. Minimal memory footprint with optimal redistribution bounds.

Performance

Constant-Time Lookups

Shard assignment is near-instantaneous with minimal resource usage — no hash ring or virtual node tables required.

Gossip-Based Membership Protocol

Scalable failure detection and membership management without a centralized coordinator.

Detection

Sub-Second Failure Detection

Scalable gossip-based membership detects node failures through randomized probe-and-gossip rounds. Multi-stage health checks with configurable timeouts detect node failures reliably before declaring the node suspect.

Propagation

Epidemic Dissemination

Membership changes propagate via piggybacked gossip messages — no dedicated protocol messages needed. Information reaches all N nodes in O(log N) protocol rounds, keeping bandwidth overhead constant regardless of cluster size.

Convergence

Consistent Cluster View

Every node maintains a local membership list that converges to the same state. Suspect nodes go through a configurable grace period before removal, preventing flapping in networks with transient partitions. Automatic conflict resolution ensures consistent cluster views.

CRDTs for Conflict-Free Replication

Replicate state across nodes without coordination. Merge any two replicas and get the correct result — mathematically guaranteed.

GSet

Grow-Only Sets

Track entity membership, active tenant lists, and registered entity types using grow-only sets. Elements can be added from any node; monotonically growing membership. Simple, fast, and guaranteed to converge.

LWWRegister

Last-Writer-Wins Registers

Configuration values, threshold settings, and per-entity metadata use LWW registers with logical timestamps. Concurrent writes resolve deterministically — the write with the higher timestamp wins, with deterministic conflict resolution.

ORSet

Observed-Remove Sets

When elements need both add and remove semantics — like dynamic shard assignments or active query subscriptions — ORSets track add/remove pairs with unique tags. Concurrent add and remove of the same element resolves deterministically.

Query Coordination & Entity Migration

Distributed queries feel like local queries. Topology changes happen without downtime.

Queries

Scatter-Gather Coordination

A query arrives at any node. The coordinator scatters sub-queries to all relevant shards in parallel, gathers partial results, and merges them into a unified response. KNN queries merge by distance; anomaly queries merge by score. The client sees a single result.

Migration

Live Entity Migration

When the cluster topology changes, entities are rebalanced across the new shard configuration. Migration happens in the background — reads continue from the source shard until the entity is fully transferred, then the routing table updates atomically.

Federation

Federated Template Exchange

Clusters can share behavioral templates discovered via density-based clustering. A template describes a behavioral pattern without exposing raw data. Remote clusters match incoming entities against shared templates, enabling cross-cluster intelligence without data centralization.

One Substrate, Microcontroller to Datacenter

VectorScaleDB is a single adaptive substrate, not a set of fixed product tiers. The same engine scales across a continuum of hardware — auto-tuning compression, indexing, and resource allocation to whatever it runs on — and operates in three network modes.

Edge

Microcontroller to Mobile

The same binary runs on microcontrollers, wearables, phones, edge gateways, and single-board computers — with battery-aware throttling and the same full compression and query capabilities. The substrate scales itself down to the device rather than forcing a stripped-down build.

Server

Workstation to Single Server

On workstations and servers the substrate lights up accelerator support (GPU and beyond), cross-domain analysis, multi-tenancy, and full federation. A dynamic hierarchy overlay enables hub topology for efficient cross-node coordination.

Datacenter

Clusters & Data Centers

At cluster scale, multiple instances coordinate with sharding, replication, and scatter-gather queries over petabyte-scale tiered storage and full adaptive intelligence — acting as a network backbone for smaller deployments.

Mode

Fully Federated

Nodes join the wider network, exchanging behavioral templates and intelligence so every node makes every other node smarter — with no single entity holding all the keys.

Mode

Lightly Coupled / Island

Nodes operate semi-independently, syncing selectively when connectivity allows and continuing fully autonomously when it does not.

Mode

Air-Gapped

Fully isolated operation with attested, offline data bundles for environments that never touch the public network.

See It in Action

Check cluster health, membership, and shard distribution through a single API call.

# Check cluster membership, shard distribution, and node health
curl "https://api.vectorscaledb.com/v1/cluster/status" \
  -H "Authorization: Bearer $TOKEN"

{
  "cluster_id": "prod-us-east-1",
  "node_count": 5,
  "shard_count": 32,
  "status": "healthy",
  "members": [
    {
      "node_id": "node-01",
      "address": "10.0.1.10:8080",
      "state": "alive",
      "shards": [0, 1, 6, 7, 12, 13, 25],
      "entities": 284019,
      "uptime_seconds": 1847293
    },
    {
      "node_id": "node-02",
      "address": "10.0.1.11:8080",
      "state": "alive",
      "shards": [2, 3, 8, 9, 14, 15, 26],
      "entities": 291847,
      "uptime_seconds": 1847290
    },
    {
      "node_id": "node-03",
      "address": "10.0.1.12:8080",
      "state": "alive",
      "shards": [4, 5, 10, 11, 16, 17, 27],
      "entities": 278432
    }
  ],
  "replication": {
    "crdt_types": ["GSet", "LWWRegister", "ORSet"],
    "convergence_status": "converged"
  },
  "federation": {
    "connected_clusters": 2,
    "shared_templates": 47,
    "last_sync": "2026-03-09T12:00:01Z"
  }
}