First, are the records of rainfall depth, flow at a gauging station, the groundwater level or something else? Are the records continuous, peaks-over-threshold, annual maximum values or something else?
From a statistical point of view, uncertainty (e.g. the confidence interval around the best estimate) will always be higher for short records. The best estimate of a particular value (e.g. the median of annual maxima) has a better chance of being close to the "true" value as more data are collected, though the amount of data required for a "good" estimate depends on the type of data, its characteristics and the acceptable bounds on the uncertainty.
Long records may need to be checked for non-stationarity. If they are non-stationary, then some statistical procedures may be less valid. Conversely, short records may benefit from a climatic adjustment procedure, if there is a much longer record covering the same time period (and more), available and it is valid to compare them.
The length of records strongly affects the results of frequency analysis and calculated Recursive Intervals, RI.
The variation of the recorded data and physio-climatological nature can determine the minimum required length for acceptable estimates. According to Subramanya (2000), a 30-yr recorded data is needed and less that 10-yr data is worthless in frequency analysis. The more the better (in homogeneous and reliable data). The assessment of reliability or test of homogeneity is required for long time series (As mentioned by Gianni).
Avoid estimating RI flood values that are greater than TWICE the record length. Estimates of flood RIs can be done with relatively short records (less than 10-yr). The needed length of record within either ±10% or ±25% errors across different RIs is given in the attached table.
The length of data records in hydrological analysis depend on the application purpose A statistical assessment of annual low flows of an appropriate duration can be based on extreme value analysis to give the yield available. For statistical analysis on temporal variability and trends, long-term homogeneous time series are required. In my experiment, a 30-year hydrological data record is very good enough.
Generally speaking, a minimum sample of size 30 years is because most of the methodology of statistical inference is based on normally distributed data and with this size sample the central limit theorem can be invoked and so normality is assumed, allowing the implementation of some test methods and parameter estimation
if your hydrological data reflect a strong cyclical phenomena (e.g. tides, snow & glacier melt, Sun activity, etc.) or a trend (e.g. due to human activity) your analysis should also take into account the observed data length with respect to mathematical methods/models you use to describe such impacts.
In some cases, if hydrological homogeneous time series data is rather short (example: less than 15 years), there are some mathematical statistical model softwares can help to users finding a fitting curve for hydrological analysis.
There must be an analysis done on this line, since very often such questions are asked in the RG site.
But, for any hydrological work related to climate change, a minimum of 100 years of data is optimum. Shorter data sets may give a trend, but not necessarily from climate change perspective, since climate mostly follows a cyclic pattern.
Please visit my RG site for some papers using >100 years of data.
Length of records simply can help you on definition of long-run behaviors (e.g. Long-run mean etc.) and periodicity of variables. For instance you can find out, say 10 year periodicity in data while you have at least 20-30 years record data and otherwise you will not be able to find out if there is a relation or persistence between present values and past ones.