Should is a dangerous word. It presupposes a measure of goodness without explicitly stating it.
If you have a noiseless system then sometimes a 1:1 ratio is acceptable. If you are trying to prove using pass-fail tests with 95% CI that your maximum error rate is under 2% - then you might want at least 400 samples giving you a ratio in the hundreds.
If you have a slow-learning Kalman filter and the right (wrong sorts of noise) then you might want thousands of measurements for a few parameters.
Bottom line: your mileage is going to vary depending on the nature of the information you get from each sample and what you are trying to do with it. My personal rule of thumb is that I prefer to get 30 samples per parameter.