Transactional vs Analytical Databases


25 Nov 2020  Sergio Martin Rubio  3 mins read.

Data can be store in two different ways depending on the business goal. Most of the companies use an OLTP (On-Line Transaction Processing) database to store end user data that requires high performance, however some companies that analyze petabytes of data internally require an alternative database called OLAP (On-Line Analytical Processing).

OLTP

OLTP databases are used to store records that come straight from the end users and they are expected to be high available and to process transactions with low latency since they are critical for running the business.

Business analysts are not suppose to use OLTP databases to avoid impacting the performance when querying a huge number of records.

OLAP

Databases are also used for analytics and here is where OLAP databases are useful because business analysts have different query patterns. Usually an analytic queries a huge number of records and particular columns to generate metrics such as total amount of sales, the average amount spent… rather than returning specific records from the database.

The query requirements usually come from business analysts and they use the data to create reports and make better decisions.

OLAP databases usually belong to a data warehouse team. The data warehouse is responsible for aggregating data from different OLTP systems in the company and stores a read-only copy of the aggregated data. The data is consume in the data warehouse as a stream of events or using a periodic data dump.

A data warehouse is a system used for aggregating data from different OLTP systems, creating reports and analyzing that data. It is one of the key pieces of business intelligence.

The process of extracting the data from OLTP, do some transformation and storing the data in OLAP is called ETL (Extract-Transform-Load).

graph LR;
    id1((Customer))-->|stream|id2[Customer Collector: extract & transform];
    id3((Agent))-->|stream|id4[Agent Collector: extract & transform];
    id5((Provider))-->|stream|id6[Agent Collector: extract & transform];
    subgraph Data Warehouse
    id2-->|load|id7[(OLAP DB)]
    id4-->|load|id7[(OLAP DB)]
    id6-->|load|id7[(OLAP DB)]
    end

OLAP databases usually use a relational model and SQL is the preferred query language, however things are quite different behind the scenes in comparison with a OLTP database.

OLAP databases use a star schema (also called dimensional modeling) or a snowflake schema. A star schema structure consist of a fact table a dimension tables.

A fact table stores individual events with the following information:

  • Measures: numeric data like sales transactions for a particular product.
  • Dimension keys: foreign key references to dimension tables.

An snowflake schema is an evolution of a star schema where dimensions are further broken down into sub-dimensions.

Materialized views is one of the features of OLAP databases and allows you to precompute certain queries so that they become very fast. However, one of the disadvantages of having materialized views is that they are not very flexible.

Some of the most popular OLAP providers are Apache Hive, Spark SQL or even BigQuery.

OLTP vs OLAP

OLTP OLAP
Generated by end users Streamed by other systems
Queried by index Queried in bulk
Consumed by end users Consumed by business analysts
Performance in important Performance is not important

One of the advantages of OLAP over OLTP is that queries can be optimized for analytic access patterns