The data set is written once a month in bulk, and read many times by different users. The last month of data (~2TB) is the hotspot. A year's worth is ~25TB.

The attributes of one record are as follows:

  * 1-4 character string
  * float
  * float
  * integer
  * integer
  * integer
  * integer
  * integer
  * 1 character

For each 1-4 character string, there are many records -- sometimes several per second. As an example, in the span of a month, one of these strings can be associated with 18 million records. There are about 10,000 unique 1-4 character strings, but not all as active as the previous example. The data is queried by two attributes: 1-4 character string and timestamp.

Potential solutions I've come up with (feel free to debate any of these):

  * Put everything (or just the hotspot) in a MyISAM compressed database.
  * Put everything (or just the hotspot) in an InnoDB database with a proper clustered index.
  * Put everything (or just the hotspot) into CouchDB with proper views.
  * Put everything (or just the hotspot) into MongoDB with proper indexes.
  * Put everything (or just the hotspot) into Redis ZSETs with timestamp as SCORE and distribute across nodes.
  * Load all of the data into a long-running Hadoop job.

Feel free to ask any questions too.