GnarlFederated geospatial
Imagery archives that are too large to centralize stay in place while analysts query the whole constellation at once.
Search without
the cluster.
Gnarl is decentralized search infrastructure. Join a network of independent nodes. Keep your data where it lives. Search everywhere.
$ gnarl join$ gnarl search "satellite imagery"$ gnarl node status
Decentralized by design
No central cluster. No single point of failure. Search across independent nodes in a permissionless network.
Your data. Your control.
Data is indexed in place and never leaves the peer that owns it. You decide what is discoverable, by whom, and for how long.
Built for scale
Globally scale to millions of peers and petabytes of data without coordination bottlenecks.
Every device is a peer
Server, desktop, laptop, phone. Nodes run from the cloud to the edge — wherever the data actually is.
Backed by Lucenia
The team behind distributed search at planetary scale.
The future of search isn’t bigger clusters. It’s smarter networks.
— Lucenia
Join the network
Run a node. Contribute. Help build the future of search.
How it works
Three commands from zero to a searchable network.
A Gnarl node is a single process that indexes what it can already read and answers the queries it can already satisfy. There is no cluster to stand up first, because there is no cluster.
- 01
Run a node
One binary. Point it at the data you already have and it announces itself to the network — no cluster to provision, no quorum to keep alive.
$ gnarl node up --data ./corpus - 02
Publish an index
Your node builds and signs a local index, then publishes only the manifest — the summary that tells the network what you can answer, never the documents themselves.
$ gnarl index publish imagery - 03
Query everywhere
A query fans out to the nodes whose manifests can satisfy it, executes where the data lives, and merges into one ranked result set.
$ gnarl search "satellite imagery"
Peers of all kinds
From the cloud to the phone in your pocket.
A peer is not a machine class. A rack of servers, a workstation, a laptop on a plane, and a phone in the field are all first-class members of the same network — each answering for the data it holds. Scale from the cloud to the edge, with no single point of failure anywhere in between.
Cloud & server
Racks, VMs, and containers carrying the heavy corpora — the peers that answer the widest queries.
Linux · Windows Server · KubernetesDesktop
A workstation with a local archive is a peer like any other. No server required to contribute.
macOS · Windows · LinuxLaptop
Answers from local disk on the move, and reconciles its manifest whenever it finds a network again.
macOS · Windows · LinuxMobile
Phones and tablets in the field join as full peers — collecting, indexing, and answering at the edge.
iOS · AndroidOne protocol, one identity model, one query API — whether the peer has 128 cores or a battery.
Cluster vs. network
The cluster was the workaround. The network is the answer.
Distributed search was built on an assumption that stopped being true: that the cheapest way to search data is to first move all of it into one place.
The clustered worldGnarl
TopologyOne cluster, one coordinator, one failure domainMany independent nodes, no coordinator to lose
Data gravityShip every document to the cluster before you can search itData never moves — the query travels to the data
ScalingRebalance shards, re-plan capacity, take the write hitAdd a node; it advertises what it holds and joins the fan-out
TrustEverything in one blast radius, under one operatorSigned manifests, per-node policy, verifiable results
Cost modelPay to duplicate data you already storePay for the nodes you run, on the hardware you own
For developers
An API you already know how to use.
Gnarl speaks plain JSON over HTTP. Point a query at any node in the network and it plans the fan-out for you — the topology is the runtime’s problem, not yours.
- Query any node, get results from every node that can answer
- Results carry the manifest and signature of the node that produced them
- Per-node policy decides what is discoverable, and by whom
- Streaming responses, so first results land before the slow tail
POST /v1/search
{
"query": {
"match": { "text": "satellite imagery" }
},
"scope": {
"network": "public",
"max_nodes": 64,
"deadline_ms": 750
},
"verify": true
}41 nodes answered128 mssignatures verified
Where it fits
Built for data that will not move.
Sovereign and regulated data
Residency requirements stop being a blocker when documents never cross a boundary to be searched.
Edge and disconnected sites
Nodes keep serving locally when the link drops, then reconcile their manifests when they come back.
Cross-organization search
Partners expose what they choose to expose. Everyone searches the union without pooling raw data.
Private retrieval for AI
Give agents grounded context from every corpus in reach without building a second copy of the enterprise.
Multi-cloud consolidation
Stop paying egress to funnel three clouds into one index. Search them where they already sit.
Published by O’Reilly MediaWho builds this
Created by the O’Reilly author of Scaling Search and Retrieval for Contextual AI.
Gnarl comes out of the same work as the book — a decade and a half of building search infrastructure for organizations whose data was never going to fit in one cluster. The St. Bernard on the cover is the same one carrying the barrel up top.
- O’Reilly published author on search and retrieval at scale
- 15+ years building distributed search infrastructure
- Eight patents in search and information retrieval
- Trusted by Fortune 500 companies and government agencies
Put a node on the map.
Install Gnarl on a server, a laptop, or a phone, point it at a corpus, and join the network in a few minutes. Free for your first peer.