We might need different types of metrics. Some of them could be simple counters. A good example is errors. We might want to count them and react when their numbers reach certain thresholds. That in itself might not be enough, and we should be able to differentiate errors depending on a function or part of the service that produced them.
How about more complex metrics? Response times are another good example.
We might need a metric that will provide request response times. That might lead us towards having something like resp_time 0.043 as the metric. It has a name (resp_time) and value in seconds (0.043). If we'd implement a metric like that, we'd soon discover that we need more. Having the information that the system responses are slow does not give us a clue which part of it is misbehaving. We need to know the name...
