Traditionally, databases are designed whereby a row is a record. The sub-elements of the row are fields such as date, part number, customer number, customer address, etc.
When a record is updated or created, the row is locked. A row is the smallest unit that can be locked which enables high transaction rates.
Imagine a data warehouse. Analysis is done by selecting from a huge table that could go back many years. You will select fields with certain parameters in that table - for example, certain dates, customers, zip codes, part numbers and prices. These selections are columnar in nature. For example you will have a column of dates, a colunn of zip codes, a colunn of part numbers, etc. These columns, unlike rows, are highly compressable and will yield great efficiencies when compressed.
Wikibon members have indicated compression rates ranging from 4:1 for transaction data and 40:1 for archive data. The trend in data warehousing is clear as innovators such as Netezza and vertica and now Oracle and IBM are rapidly introducing columnar compression into the market.