Product Performance Architecture Ecosystem Use Cases Community Early Access
Rust × DuckDB · Distributed Data Fabric

Distributed ETL & compute at the speed of Rust.

Arolium fuses DuckDB's vectorized engine with a Rust-native distributed fabric — declarative pipelines, elastic SQL compute, and exactly-once delivery in one open platform. One binary from laptop to thousand-core cluster.

Get Early Access View on GitHub
99 / 99TPC-DS · value-exact
5.2×vs Spark
< 3 mstask dispatch p50
~40 MBsingle static binary
arolium — fabric shell

Powering data platforms at

NordwindVektra LabsDatalaneHelioscaleQuanta MetricsOsmixFerrostream
Core Capabilities

Three pillars. One fabric.

ETL and analytics have lived in separate systems for too long. Arolium collapses the pipeline and the query engine into a single distributed runtime.

Elastic SQL Compute

DuckDB's vectorized kernels on every worker, stitched into an MPP fabric.

  • Distributed shuffle engineAlias-aware repartitioning drives distributed joins and aggregations — including self-joins and grouping sets.
  • Full SQL surfaceWindow functions, CTEs, ROLLUP / CUBE, correlated subqueries. TPC-DS 99/99, value-exact.
  • Scale without rewritesThe same query runs on one core or a thousand. The planner decides; your SQL doesn't change.

Declarative ETL Pipelines

From CDC to lakehouse in one DAG — checkpointed, versioned, exactly-once.

  • Pipelines as codeSQL + YAML DAGs that live in git, review like code, and deploy in one command.
  • 40+ connectorsKafka, Postgres & MySQL CDC, S3, Iceberg, Parquet, HTTP — sources and sinks alike.
  • Incremental by defaultCheckpointed state and idempotent sinks give exactly-once delivery without ceremony.

Rust-Native Reliability

No GC pauses. No JVM tax. Predictable latency under real pressure.

  • Memory-safe end to endFrom wire protocol to storage engine — one language, zero segfault class, flat p99s.
  • Raft-replicated control planeMetadata survives node loss with no single point of failure and no external ZooKeeper.
  • One static binary~40 MB, cold-starts in milliseconds. Deploy with a copy, not a cluster of sidecars.
Benchmarks

Fast is the feature.

Vectorized execution where the data lives, a shuffle fabric that respects your network, and a scheduler that thinks in microseconds.

  • Whole-plan vectorizationOperators execute on DuckDB's columnar kernels — SIMD all the way down, spilling gracefully past RAM.
  • Shuffle that stays out of the wayArrow Flight transport with zero-copy hand-off; partition pruning keeps bytes off the wire.
  • Honest benchmarksFull TPC-DS suite, value-exact results, identical hardware envelopes. Reproduce every number from the repo.
TPC-DS · total wall-clock lower is better
Arolium418 s
418 s
Spark 3.52 167 s
2 167 s
4-node cluster · 2 vCPU / 4 GiB per worker · identical hardware & data · all 99 queries, results value-exact against the reference. Methodology and raw runs published in the benchmark repo.
5.2×faster end-to-end
8.4×less memory per worker
0GC pauses, ever
Architecture

One engine. Two topologies.

A Raft-replicated control plane schedules work onto stateless workers. Run compute next to NVMe for the hottest paths, or float it over object storage and scale to zero.

Mode 01 · Co-located

Compute beside NVMe

Workers own local storage for the hottest tables. Zero network hops on the critical path — the lowest latency an MPP topology can give you, for customer-facing analytics.

Mode 02 · Disaggregated

Stateless over object storage

Workers hold nothing but cache. Point the fabric at S3, scale from zero to hundreds of workers for a backfill, and back to zero when it's done. Pay for compute you actually use.

Ecosystem

At home in your data stack.

Arolium sits at the center — ingesting from the systems you run, serving the tools your teams already love.

Ingest from

Kafka · Redpanda PostgreSQL CDC MySQL CDC S3 · GCS · Azure Iceberg · Parquet HTTP · Webhooks
Arolium AROLIUM ETL + COMPUTE

Serve to

Iceberg tables Superset · Metabase Grafana Arrow Flight SQL AI agents · MCP Webhooks

40+ CONNECTORS · POSTGRES WIRE-COMPATIBLE · ARROW-NATIVE EVERYWHERE

Use Cases

Built for the workloads that matter.

01

Customer-facing analytics

Sub-second dashboards under thousands of concurrent sessions. Vectorized scans plus a shuffle fabric that keeps tail latency flat when your traffic isn't.

sub-secondhigh concurrencyflat p99
02

Streaming ETL & CDC

Replicate OLTP databases into the lakehouse continuously. Checkpointed state and idempotent sinks make exactly-once the default, not a research project.

exactly-onceCDCsecond-level freshness
03

Lakehouse compute

Query Iceberg and Parquet in place — no ingestion detour, no second copy. Partition pruning and metadata caching keep object-storage scans honest.

Icebergzero-copyin-place SQL
04

AI data engineering

Feature pipelines, training-set builds, and vector-ready transforms — exposed to agents over MCP so your AI stack can query the fabric directly.

feature pipelinesMCPagent-ready
Community

Built in the open.

Apache-2.0 licensed, developed in public, shaped by the engineers who run it. 180+ contributors and counting.

GitHub Source, issues, roadmap, and RFCs. Star the repo to follow releases. github.com/arolium → Discord Setup help, performance tuning, and answers from maintainers. discord.gg/arolium → X / Twitter Release notes, benchmark drops, and engineering threads. @aroliumdata → Dev mailing list Design discussions and release votes, in the open. dev@arolium.dev →
Coming Soon

Something fast is coming.

Arolium is in active development, being forged in the open. Watch the repo or join the community to be first in line when the initial release lands.

$ arolium install Coming soon
Watch on GitHub Join the community