Groundwater level time series are of great value for a variety of groundwater studies, particularly for those dealing with the impacts of anthropogenic and climate change. Quality control of groundwater level observations is an essential step prior to any further application, e.g., trend analysis. Often the quality control of data is limited to the removal of outliers or elimination of entire time series from a dataset, while such approaches drastically reduce the spatial coverage of initially huge datasets. Frequently studies tend to present already qualitycontrolleddata, but neglect to demonstrate how the data were selected, judged, and modified. We present a data rescue approach developed for correcting the Latvian national groundwater level database, containing 1.68 million groundwater level observations since 1959, including 0.69 million manual measurements. A web-based R-Shiny interface was developed and used for visual identification and manual correction of erroneous measurements in groundwater level time series. All data manipulations were performed programmatically. Reproducibilityand traceability were ensured by deploying separate data tables for raw observations, data repair actions and the final dataset. As a result of applied actions, 34.3% of all automatic measurements were either deleted or corrected, while only 6.5% of manual measurements were edited. Commonly found errors ingroundwater level time series were grouped into: errors in measurement and data recording; technical problems at the observation site; local anthropogenic impact and other unclassified problems. The improvement from the rescue approach was assessed by comparing the Akaike information criterion derived from fitted ARMA and ARIMA models to both original and repaired time series. The results showed that models fitted using repaired time series were better than those fitted on the original time series for the same time series sections. The presented rescue approach and results can be of great value for all studies using groundwater level time series as an input.