When choosing an analytical database, of course, you should test on your own data and queries that your users will ask. But to create a short list of DBMSs for testing, everyone wants to see real data from other people's tests. The guys from the Data Warehouse Department of Tinkoff Bank shared their in-memory (and not only) database testing results and observations on Hubra. Today, here is a brief extract of their performance tests of Exasol, Greenplum, ClickHouse, SAP HANA, MemSQL and Cloudera Impala.

​

​

DBMS requirements and use case

The current Greenplum database-based storage was not very satisfactory in terms of query execution speed. DBMS testing was performed to select a front-end storage database for a selective data set (up to 4 TB, with a possible increase). The target database had to have the following functionality:

• column-by-column data storage,

• horizontal scalability,

• the ability to perform local joins and use the "correct" distribution key in tables (ad-hoc user queries in 90% of cases are Select with joins (from 1 to 10) based on equality conditions and, sometimes, on the conditions of dates within the interval),

• efficient work with cache and large amount of available memory,

• Good integration with the BI-system SAP Business Objects (it will access the database, in addition to ad-hoc user queries)

• Reliable, preferably incremental, data import from Greenplum (this is Tinkoff's current main data warehouse and data will be loaded into the database from it).

• Window functions,

• Redundancy (ability to store multiple copies of data on different nodes),

• Ease of further expansion of the cluster,

• Parallel data loading.