Two awkward shapes, and the cloud in the middle
Machine learning and data teams move training sets to an annotation vendor and get labels back, ship datasets from a lab to the cloud where the GPUs are, and pull checkpoints of tens of gigabytes per epoch back down for evaluation. The data is either millions of small samples, where per-file overhead dominates, or a few sharded archives of hundreds of gigabytes, where a single stalled stream wastes a night.
Consumer sharing tools cap sizes or meter the bytes, sync tools choke on millions of files, and a browser upload over TCP to another continent never reaches the line rate you pay for.
Small files batched, big shards in parallel
Boita moves data with an accelerated transfer engine over UDP with its own rate control, holding your line rate whatever the distance. A few huge shards run as parallel sessions sized to your plan. A tree of millions of small files is grouped into transfers by the desktop app or the agent, in batches of up to 500 files or a ten-second window, so the per-file cost of an ordinary upload does not apply. Transfers checkpoint and resume after any interruption; completed files are never re-sent.
The server records a checksum for every file, and the delivery receipt carries per-file checksums with a CSV verification report, so an evaluation run uses the checkpoint that was actually written.
Fits training scripts and object storage
The boita command line runs in a training script or a scheduler job on the Linux agent, and the REST API with a Python SDK, scoped API keys and signed webhooks (Pro and above) does the same from your own code. Import from an S3-compatible bucket pulls a dataset into the workspace on our servers, not through a laptop. On Enterprise, auto-delivery pushes everything landing in a workspace folder to your own S3-compatible bucket within minutes, sub-folders kept and retried on failure. GCS works today through its S3-compatible endpoint; Azure Blob and native GCS connectors are coming.
An annotation vendor sends labels back through an upload portal with a form for batch identifiers and file-type rules, into its own dated inbox folder, with an email to both sides when it lands.
A worked example
A 1 TB dataset leaves a lab on Pro, over a 1 Gbps line, in about five hours, plan-limited at 450 Mbps; a 40 GB checkpoint on the same plan takes about twelve minutes. The same 1 TB on Studio over a 1 Gbps line is line-limited and takes about two and a quarter hours. What the far end sees depends on the distance and loss on its side; the calculator on the home page models it.
- Millions of small files batched into transfers; huge shards in parallel sessions
- boita CLI, REST API and Python SDK for scripts (Pro and above)
- Import from an S3-compatible bucket; auto-delivery to your bucket (Enterprise)
- Upload portals with forms for annotation vendors
- Checksum on every file; receipts with a CSV verification report
Controls on proprietary data
Sender-side encrypted delivery with a passphrase keeps a dataset unreadable to anyone, including Boita, until the recipient enters it. Share links carry an expiry, an optional password and a download limit, recipient verification by email code names who downloaded, and a revoked link stops resumed downloads within minutes. Folder permissions per member or group (Studio and above) give a vendor’s contact one folder and nothing else. Two-factor authentication and an audit log are on every plan; SSO, IP allow-lists and audit export to a SIEM are on Enterprise.