The read path is not necessarily as efficient as the write path. To get the required columns for row, the read might need to get data from the memtable and various SSTables that might have different parts of the data. The read is effectively a merge across all these data structures:

There are ancillary data structures, such as bloom filters, which help in figuring out whether an SSTable contains a given row, thereby avoiding reading (expensive disk I/O ) for each SSTable. Compaction also helps read performance since the number of SSTables to be compared is reduced. The SSTable format includes an offset lookup, which helps in figuring out the offset within the SSTable file for a given key.