I suppose they are rounded in the graph, rely on a graph to see if the engine has improved performance is not the best also because the node could be overloaded. I can not say more than this

While the Y-axis shows 0.5 seconds resolution, the data is truncated to seconds. The test in builds 33 & 34 have run for 1.8 and 1.9 seconds, but both are displayed on the graph as 1.0s.

I don't see why there should be any truncation / conversion to integers. Isn't it possible & simpler to just use the floating point data from the report in the graph? I assume here the truncation to integers is non intentional and just an oversight, but maybe there is a good reason for the truncation... ? I think (but not 100% sure) that there was floating point data in the previous style of duration graphs.

Having the actual floating point values plotted would allow to properly see performance improvements of our core algorithms in these graphs. These performance improvements are typically in the range of 10% - 20% and are now invisible due to the truncation to seconds.

graph is built by junit plugin by using the report generated by xunit plugin. From the attached image the scale is 0.5 second and it's clear that it show an appromimation because you could have a test that can takes 1.01 seconds and other one 1 minutes and the differences of 0.1 are invisible in the graph.