I'm developing a time-series database. My main use case is not monitoring but time-series data mining and I want to discuss some ideas behind it.<p>This TSDB should act as raw-data storage that can generate various useful representations of the original data (SAX, PAA, HWT, FFT, MA/MM, etc).<p>For example, you can ask it to generate SAX representation from original data and use it to find motifs or discords or to answer 1-NN queries approximately. Maybe you'll want to build index using Lucene or jMotif. But maybe SAX is not what you want and you need to index data in frequency domain or something else...<p>There is plenty of different indexing schemes and models for time-series data and I can't implement just one or two of them. I'm planning to implement different preprocessing steps for this models instead. I think that this approach (store raw data, generate various representations) would be much more useful. What do you think about this?