How do SSD drives change things for main memory databases?

Most main memory databases have a log which contains data modifications, as well as a copy of the data for persistence.  Write performance may be increased compared to spinning disk because the log can be written to faster.  In addition checkpoints which write data to persistent storage can complete faster.  Finally, since main memory databases normally load data into RAM at startup from the persistent store, start time can be improved since SSD have faster read throughput.  Even if data is loaded lazily, improved read performance will lead to faster access when data must be loaded from disk.

My experience is a bit out of date, but I will try to shed some light on this question anyway. Solid state drives are a mixture of technologies between old fashioned disk drives and integrated circuit memory.  They are accessed in a similar fashion to the interface between central processors and disk drives.  That means the data transfer and data selection commands are transmitted serially - over a single electrical circuit. Main memory in computers may use the same sort of integrated circuits as a solid state storage device but probably not exactly the same.  This is because main memory is usually volatile while SSD memory should maintain its contents when power is removed.  Main memory is accessed by the CPU over parallel address and data buses.  The width (number of electrical circuits) in these buses varies from one computer architecture to another, but it is safe to say access to main memory is a lot faster than access to an external disk drive.  Over the history of computers the general difference in speed for access to a single byte of main memory compared to a single byte of data on disk is 4 orders of magnitude faster.  Stated differently, it takes about 10,000 times as long to get a byte off a disk as to get one from main memory.  This ratio has remained relatively constant because the speeds of processor parts and buses has increased about the same amount as the speed of disk drives over the years. Solid state storage devices offer two obvious advantages and one disadvantage over spinning disks.  Access speeds will be somewhat better.  You don't have to wait for heads to move and there is no "Rotational latency" waiting for the information you want to arrive at a read/write head.  Secondly, SSDs have no moving parts that can crash like disk heads do under really poor conditions.  The negative side is that SSDs will be a lot more expensive than disks of the same capacity.  It is relatively easy to trade disk access speed for capacity, but with SSDs you need to buy more chips to get more memory. Like all other parts of computer system engineering there are trade-offs to be made for any application.  If you are facing a harsh environment like a factory floor or airplane cockpit then the SSD is probably a good idea.  For normal data processing applications the rotating disk will provide a lot more capacity at a lower price.

VoltDB is a relational data store, with SQL as its DDL and query language, and is ACID compliant.  It falls in the "NewSQL" category, vs the NoSQL category, in which the DynamoDB resides. As for SSD usage, VoltDB handled all data in memory, making for fast transactions for both data ingestion, and query.  Hundreds of thousands of transactions a second on a small commodity linux cluster (where a transaction can be many SQL statements)  Since data is in memory random reads are fast - it doesn't touch the disk (spindle or ssd) when reading.  As for durability, VoltDB's durability strategy is Snapshots coupled with write-ahead logs (sync or async).  We work great on spindle drives and slightly better on SSDs. John Piekos (Note, I work for VoltDB)