* Implement count-min-sketch (supposedly more space-efficient than count-sketch) * Time-adaptive sketch: decay accuracy & storage for older values (using inflation) * Integrate with atomic multilog for offline/online approximate queries. - Figure out what the basic interface should look like - Consider: - cross-stream correlation queries - use-cases where the same stream is partitioned across multiple servers - storage-accuracy tradeoff, [blowfish](https://www.usenix.org/system/files/conference/nsdi16/nsdi16-paper-khandelwal.pdf)