compute next to your data

Scientific datasets are measured in terabytes and petabytes. Inside a cloud region, compute reads that data at enormous throughput. Over the internet — to a laptop, a notebook server, or an on-prem cluster — the same bytes squeeze through a link that is orders of magnitude slower. Every byte that leaves the cloud is billed as egress.

• Slow — hours of downloading before the analysis can even start.

• Expensive — egress charges on every byte that leaves the cloud.

• Duplicated — local copies that go stale and multiply across a team.

Data-proximate compute flips this around: do the heavy, I/O-intensive work in the cloud, right next to the object store, and send back only the result — a table, a map tile, a timeseries.

Earthmover makes it easy to deploy data-proximate compute services connected to your Arraylake data. We handle deployment, orchestration, scaling, auth, and security. You focus on the science.

• Fast — heavy reads run at in-region throughput, with no download standing between you and the answer.

• Cost-effective — only the result leaves the cloud, so egress shrinks to the size of the answer.

• Single source of truth — everyone queries the same governed dataset, with no local copies to go stale.

Every service runs on its own cluster, and you choose two things: a node size, from XS to 3XL, and a replica range, how many nodes the cluster may run. Scale up by picking a bigger node; scale out as traffic rises, up to 24 nodes, and back in as it falls; and scale to zero when a service sits idle, so an occasional workload costs nothing between uses.

Usage is metered in credits. Node cost is proportional to compute capacity, every service draws on the same pool, and you pay only for the nodes that are running. On Professional plans, keep a minimum of nodes warm for always-on services that can’t wait for a cold start.

Managed Compute services runs on Zax, the query engine we're building for multidimensional chunked arrays. Written in Rust, Zax plans a query the way a database does, but in multidimensional space: it works out which chunks the query touches, schedules I/O against the data's layout in object storage, and sizes the work to the resources available. Zax makes managed compute services reliable, fast, cost-effective.

Managed Compute is optional. Your data stays in open formats in your own bucket — arrays in Icechunk, tables in Apache Iceberg — so the tools you already run read and write it directly, with Arraylake vending the credentials. Open a repository with Xarray or Zarr, stream it into a PyTorch training loop, or point DuckDB, Spark, Polars, or Snowflake at the Iceberg REST catalog.