OLAP was a departure from OLTP, online transaction processing, the mainstay of business transactional networks. And of course, both are contrasted to batch processing which occurs later or after the fact, though in practice, much OLAP was specific triggered reports or ad hoc analysis, at least from the world as I saw it.
Where OLTP and OLAP differ is in that OLTP deals with a transaction at a time, typically following a single entity or account through a single transaction or operation. The data loads are small though the transaction counts are high, and there's virtually no summation of information. Database optimisation is built around these needs.
OLAP, in theory, was based around realtime or near-realitime access to transactional data, but the analysis looked across entities (typically accounts, products, divisions, regions, time, etc.), and relied heavily on summarised data (counts, sums, means, medians, min, max, standard deviation, occasionally other metrics). This was typically rationalised by either defining views, or (more often in my experience) by precomputing what were thought to be sufficient and useful summary statistics. (In practice, neither ambition ever seemed to be reliably met, requiring returning to source or disaggregated data.)
From a statistical stadpoint, one of the chief failures of OLAP is that it attempts to substitute summarisation for sampling. In practice, a small sample of a very large dataset is quite frequently orders of magnitude more easily compiled and analysed, with very little loss of information. Even aggregate computations can often be closely estimated by this method.
The article explicitly makes most of these points, FWIW.