Keldra

| DISTRIBUTED OBJECT STORAGE - RUST - APACHE-2.0

Keldra.

Distributed object storage for application state.

Stable paths and opaque bytes, with the coordination primitives applications usually have to assemble around a blob store: streaming writes, compare-and-swap, immutable namespaces, bounded atomic programs, Zanzibar authorization, change notification, and materialized search indices.

View on GitHub

Read the guide

CURRENT RELEASE

0.11.3

Keldra system map

FIG. 02

objects

surface

indices

surface

Keldra

app model

gateways

surface

One authorized storage core serves objects, indices, and standard protocol gateways from any active node.

Any node

can accept public requests

2+1

fixed erasure-coded payload durability

materialized index engines

3 APIs

native gRPC, S3, and Git on one endpoint

SECTION

Storage primitives that compose.

The object API keeps each operation explicit, so applications can select the ordering and durability contract they actually need.

Objects

Streaming writes and stable paths

Put, get, head, delete, batch reads, bulk writes, prefix listing, optional retained versions, content deduplication, and change watches.

Coordination

CAS and write-once data

PutIfAbsent, exact-version compare-and-swap, conditional delete, and immutable puts make publication intent visible at the API boundary.

State transitions

Bounded atomic programs

Deterministic JSON programs can protect a small multi-path invariant without routing ordinary objects or large media through a transaction engine.

Scale out without changing the API.

Start with one process, then add capacity and availability while clients continue addressing the same tenant, bucket, and path model.

DISTRIBUTED CORE

Any-node ingress

Every active node can accept a request and route work to the correct authority.

Weighted placement

Capacity-weighted rendezvous hashing places data across heterogeneous nodes.

Separated authority

Raft carries compact membership and publication decisions, not object bodies or index files.

Online growth

Large objects use complete replicas below erasure width, then move online to the fixed erasure profile as nodes join.

One authoritative object model.

Every successful write creates immutable content and advances one exact-path head. Derived structures remain reconstructible.

AUTHORITY MODEL

Request

→

Immutable content

→

Exact-path head

Stable numeric IDs

BLAKE3 identity

Ordered journals

Disposable caches

SECTION

Search is a materialized view, not a second database.

Bucket-local definitions consume ordered object journals, publish complete immutable generations, and return freshness evidence with every query.

Path

Prefix discovery and stable pagination over object paths.

Metadata

Predicates over path, content type, length, and retained head fields.

Typed JSON

Declared predicates, ordering, facets, full text, and numeric aggregates.

Full text

Fielded text and phrase search over selected JSON pointers.

Vector

Cosine, dot-product, or Euclidean similarity with fixed dimensions.

Hybrid

Weighted full-text and vector retrieval in one definition.

Git source

Commit and tree-path lookup over Git manifests and pack locations.

Tensor

Model and tensor-name lookup over typed tensor manifests.

Ordered journals → bounded immutable segments → complete generation → freshness-bearing results

NO PARTIAL GENERATIONS

SECTION

Protocol-native at the edge.

Keldra exposes native storage capabilities and familiar ecosystem protocols on the same authorized public listener.

Native API

gRPC and Rust client

Generated versioned contracts cover objects, indices, authorization, administration, accounting, and PersonalDB, with authenticated Rust helpers for common workflows.

Object ecosystem

S3 SigV4 gateway

Path-style S3 requests work with the AWS CLI and official SDKs while Keldra application credentials and Zanzibar policy remain authoritative.

Source control

Git smart HTTP

Authenticated push, pull, and clone use ordinary Keldra authorization; bucket owners can explicitly enable anonymous reads for public repositories.

SECTION

Private by default, observable in production.

Authentication, authorization, placement, and operational evidence are designed as first-class parts of the storage system.

Zanzibar authorization

Tenant and bucket roles, customer schemas, relationship tuples, and authorization-filtered index results.

TLS

Authenticated peers

Short-lived client JWTs protect public APIs while mandatory peer mTLS isolates cluster traffic.

PersonalDB

Witnessed predecessor-linked SQLite changeset logs, explicit projections, catch-up streams, and signed descriptors.

Usage accounting

Authorized bucket or path-prefix aggregates report objects, logical bytes, traffic, and source freshness.

From one node to a cluster.

The published multi-platform container runs on Linux AMD64 and ARM64. The repository includes Docker Compose and single- and three-node qualification workflows.

FIVE-MINUTE PATH

docker compose -f crates/keldra/docker-compose.yml up -d

keldra create-bucket objects

keldra put example objects hello.txt ./hello.txt

keldra get example objects hello.txt

Public, executable evidence.

Release gates exercise the same public interfaces clients use, including real S3 SDK traffic, Git clients, index queries, rolling restarts, and online cluster growth.

RELEASE EVIDENCE

Single-node qualification

Three-node cluster qualification

All eight index engines

S3 and Git real-client workflows

Linux AMD64 + ARM64 image

OPEN SOURCE

Read the code, contracts, and qualification evidence.

Keldra is Apache-2.0 licensed. The repository is the source of truth for released capabilities, setup, architecture RFCs, and current operational limits.

Explore Keldra

Courtney Robinson

zcourts.com

Building the infrastructure for the agentic future.

EXPLORE

RESOURCES