However, I don't see any benefit in TileDB supporting fast and efficient updates (and duplicates) of time series data. Time series should be immutable, and only in rare occasions require updating.
Of course there are trade offs, we could have forced a longer waiting period until we were confident the data was finalized. We also could have ignored the updates. In the end we used a system of staging tables for loading the last 24 hours of data before merging out into a more finalized table. This kept the load of updating records in the database down, and still allowed us to achieve our goals.
At the time, several years ago, I was not aware of TileDB else we would have considered it instead of a more traditional database vendor.