Draw the data path before launching
Write down the source dataset, temporary training cache, durable checkpoints and final artifacts. Give each a storage location and retention period. This takes a few minutes and often reveals that the same dataset will be copied several times.
Estimate both bytes and time. If a worker spends its first hour downloading data, that hour belongs in the job budget even if the network transfer itself is free. Repeating the download on every restart can dominate a short experiment.
Local and persistent storage fail differently
Local scratch space is useful for caches that can be rebuilt. Durable state needs a storage service with the lifetime and recovery behavior the project requires. The word persistent still needs a definition: persistent across a process restart, across a stopped instance, or after deletion?
Lambda warns that local data is destroyed when an instance terminates and documents export to filesystems or external object storage. Its filesystems must be attached at launch and reside in the same region as the instance. Plan those choices before allocating expensive compute.
References: Lambda data import and export · Lambda filesystems
Make retention an explicit line item
Consider an illustrative project that keeps five 100 GB checkpoints for two months. That is 1,000 GB-months before accounting for replicas, snapshots or billing-unit differences. Apply the provider rate to that quantity, then add expected transfer costs.
Keep the latest recoverable checkpoint and any artifacts required for reproducibility. Retaining every intermediate result forever is a policy choice with a cost, not a free default.
| Data | Owner decision |
|---|---|
| Raw dataset | Authoritative location and access controls |
| Scratch cache | Whether rebuilding is cheaper than retaining |
| Checkpoints | Restore requirement and deletion date |
| Final model | Version, license and long-term destination |
Close the project with a resource inventory
After the job, list instances, volumes, snapshots and object prefixes by run identifier. Export the final artifact and test that it opens. Only then remove disposable resources and record what remains.
Assign an owner to the retained resources and a review date. A billing alert can reveal a forgotten volume, but it cannot decide whether deleting that volume will destroy the only usable checkpoint.
Questions readers ask
Is free egress enough to choose a provider?
No. Loading time, storage price, regional restrictions and repeated copies may matter more for your workflow.
Should I keep checkpoints on the worker?
A local copy can help resume a process, but keep the recovery copy somewhere that survives loss of that worker.
Sources and editorial notes
Primary documentation reviewed October 9, 2026. Provider descriptions reflect published documentation. Recommendations are our analysis; worked figures are illustrative unless explicitly stated otherwise. No independent provider benchmark or uptime ranking is claimed.
Read the editorial policy. NeuraMesh is a prototype, not live verified GPU inventory.