Another key difference between database systems is how they handle updates to the physical records stored on disk.
Relational databases, such as MySQL, maintain a variety of structures in both the memory and disk, where writes from in-flight transactions and writes from completed transactions are persisted. Once the transaction has been committed, the physical record on disk for a given key is updated to reflect that. On the other hand, many NoSQL databases, such as HBase and Cassandra, are variants of what is called a log-structured merge (LSM) database.
In such an LSM database, updates aren't applied to the record at transaction commit. Instead, updates are applied in memory. Once the memory structure gets full, the contents of the memory are flushed to the disk. This means that updates to a single record can be fragmented and located within separate flush files that are created over time. This means that when there is a read for that record, you need to read in fragments of the record from the different flush files and merge the fragments in reverse time order in order to construct the latest snapshot of the given record. We will discuss the mechanics of how an LSM database works in the Chapter 6, HBase.
