Problem

Currently histograms are reported using a bunch of manually created counters with names like __1ms, __2ms, __10ms and a label histogram="ExecutionTime" by which you can identify the bucket. This format is supported only in Solomon. Example (prometheus format):

__001ms{component="server", request="ReadBlocks", histogram="ExecutionTime", } 1
__1ms{component="server", request="ReadBlocks", histogram="ExecutionTime", } 0
__2ms{component="server", request="ReadBlocks", histogram="ExecutionTime", } 4
__3ms{component="server", request="ReadBlocks", histogram="ExecutionTime", } 1

The grafana is waiting for histogram in the following format (with "le" label that means "less or equal" bucket):

ExecutionTime_bucket{component="server", request="ReadBlocks", le="0.001", } 1
ExecutionTime_bucket{component="server", request="ReadBlocks", le="1", } 1
ExecutionTime_bucket{component="server", request="ReadBlocks", le="2", } 5
ExecutionTime_bucket{component="server", request="ReadBlocks", le="3", } 6

Solution

monlib already has all that we need - special class THistogram that can report histogram in format supported both in Solomon and Grafana. One need to support reporting histogram using this class. To provide backward compatibility it is proposed to leave the old format (with multiple counters). It would be useful to add two new parameters in diagnostics config that will manage histogram report mode:

• ReportHistogramAsSingleCounter, with default false value
• ReportHistogramAsMultipleCounters, with default true value