A Delta version identifies a transaction-log state. Reading that old state still requires the data files it referenced. If those files were removed from the current table and later deleted by VACUUM, the old version can no longer be read just because someone wrote its number in an experiment record. Databricks' VACUUM documentation says vacuuming removes files needed for older time travel. Log retention and data-file retention are separate controls. An old log entry remaining available is not proof its files remain.

I would ask for the training run's exact table ID and version, file paths or snapshot manifest, retention settings at that time, vacuum history, and the error on a VERSION AS OF read. Check whether the trainer actually read a fixed snapshot. If it streamed the changing table for days, one version number recorded at job start may not describe all the examples it consumed. Data may also come from multiple tables, external object stores and transforms whose code or configuration changed. Reproducibility is a chain, not a single Delta version.

For future runs, define the required replay horizon before setting retention. A training pipeline that needs six-month exact replay must retain the referenced files for that period or materialize a governed immutable training snapshot. Store its manifest, content checksums, transform version, tokenizer and sample ordering alongside the model checkpoint. Verify by rerunning a sample read at the oldest promised age, not by checking that the metadata record still exists. Storage cost and deletion obligations are real constraints. An immutable copy cannot casually override a customer's erasure or retention policy, so the replay promise must state the allowed exceptions.

Can we simply lengthen VACUUM retention today and recover the old run? No, extending future retention cannot bring back files already removed. Restore from a valid backup if one exists, or mark the run as not exactly reproducible and compare at a weaker level. Say clearly which level of comparison is still possible.