Sounds like you’re treating the vineyard like a versioned dataset. I usually break the problem into a few layers. First, raw data—timestamped readings, sensor logs, weather feeds—kept in a normalized table so each column has one meaning. Next, a feature table: you derive things like moving averages, growth indices, or lagged values that capture temporal patterns. Finally, a model table where you store the coefficients or neural net weights that map features to the next phase. I keep the schemas flat for quick lookups, but add a small metadata layer so I know what each field means and when it was last updated. That way, when the model needs a fresh input, I just pull the latest row from the feature table, feed it to the prediction routine, and get the next expected state. It keeps the logic separate from the data, so debugging is just a matter of tracing a single column instead of hunting through mixed logs.