There is a perspective on MapReduce that it is a step backwards in several respects:

• In terms of database access, MapReduce is poor implementation
  • Can only use brute force (e.g. no indexes)
• MapReduce is not novel
• MapReduce is missing features
  • e.g. bulk loader, indexing, updates, transactions
• MapReduce is incompatible with DBMS tools

In response, software like Vertica was made

• Analytical database, designed for OLAP workloads/cluster/big data

Regardless of what system is used, some actions are very slow, like parsing text.

• Loading lines, splitting whitespace, integer parsing are slow operations

To deal with such issues:

• Ned a binary format with schema
• Schema should separate logical and physical views

We arrive at using row or column stores:

• Files are 1 dimensional ⇒ need to project high dimensional data into 1D byte sequence
• For row stores
  • Easier to modify a record (in-place update)
  • Unnecessary data may be read while processing