Retention ledger design for large browser uploads using presigned multipart parts

The article argues that large training artifacts uploaded from a browser should use presigned multipart uploads, with the application database tracking a three-state record of uploading, retained, and deletion_due. Object bytes stay off the Node.js control plane, and every multipart session must be completed or aborted explicitly. Retention is treated as a property of the application record rather than something inferred from bucket state, since incomplete parts have no automatic cleanup rule. The author separates a short deadline for abandoned sessions from the longer retention deadline for completed artifacts, using bucket lifecycle rules only as a backstop with a one-day minimum. The piece suggests tus or UploadThing when interruption recovery matters most, and AWS S3 when versioning or Object Lock is required.