Time Series Data - HBase Design Patterns - page 73

The UID table schema
A separate, small table called tsdb-uid stores UID mappings, both forward and
reverse. Two columns exist, one called name that maps a UID to a string and another
called id that maps strings to UIDs.
Here, we will learn to keep the lookup tables in the same HBase. A small table can
be directed to be served from the memory, so you can use HBase for its in-memory
database capabilities too. Additional information can be gleaned from the OpenTSDB
design documentation found at http://opentsdb.net/docs/build/html/user_
guide/backends/hbase.html .
For practical purposes, if your NoSQL database is Cassandra and not HBase, you
will find that there is a Cassandra implementation of TSDB, called KairosDB , which
can be found at https://code.google.com/p/kairosdb .
Summary
We learned the important NoSQL and HBase design principles related to storing
and accessing time series data. We described how to utilize time-based keys, how to
avoid region hotspotting, how to properly balance tall and narrow rows versus wide
rows, and we also saw how to glean additional information from other systems on
top of HBase, such as OpenTSDB. All this together should make us masters at storing
and using time series data in HBase and in NoSQL databases in general.
In the next chapter, we will deal with the most common design principle, that
is, denormalization. We will also discuss how to store all the objects for a user,
popularity contest, and how to store tags efficiently.