History and trends are the two ways of storing collected data in Zabbix.
Whereas history keeps each collected value, trends keep averaged information on hourly basis and therefore are less resource-hungry.
You can set the storage period for how long the history will be kept:
Any older data will be removed by the housekeeper.
The general strong advice is to keep history for the smallest possible number of days and that way not to overload the database with lots of historical values.
Instead of keeping a long history, you can keep longer data of trends. For example, you could keep history for 14 days and trends for 5 years.
You can get a good idea of how much space is required by history versus trends data by referring to the database sizing page.
While keeping shorter history, you will still be able to review older data in graphs, as graphs will use trend values for displaying older data.
Trends is a built-in historical data reduction mechanism which stores minimum, maximum, average and the total number of values per every hour for numeric data types.
You can set the storage period for how long the trends will be kept:
Trends usually can be kept for much longer than history. Any older data will be removed by the housekeeper.
Zabbix server accumulates trend data in runtime in the trend cache, as the data flows in. Server flushes trends into the database (where frontend can find them) in these situations:
To see trends on a graph you need to wait at least to the beginning of the next hour (if item is updated frequently) and at most to the end of the next hour (if item is updated rarely), which is 2 hours maximum.
When server flushes trend cache and there are already trends in the database for this hour (for example, server has been restarted mid-hour), server needs to use update statements instead of simple inserts. Therefore on a bigger installation if restart is needed it is desirable to stop server in the end of one hour and start in the beginning of the next hour to avoid trend data overlap.
History tables do not participate in trend generation in any way.
Also restarting server might result in the precision loss of unsigned data type average value calculations for the current hour.