Each node stores its histograms as running totals since it started.
deltaHistograms lined up every node's samples by time and diffed them in
one list, so in a cluster it subtracted one node's total from another's.
The timing averages came out wrong: on production the delivery attempt
time read 14.7 s over the last hour ("1 stage running slow") when the
nodes' own figures give 0.7 s.
Samples now carry nodeId (inbuxa-server, x:Metric), and totals are diffed
per metric and node. Without nodeId (older servers, single node) it works
as before.
jcoffey-dev