Strange variance in calculated items

[evgen]

I have been trying to get some calculated items working in the manner that one would rationally expect them to work and am running into strange variance on the calculated values that I cannot explain.

For example (using 1.8.4):

Counters A, B, & C are delta counters that are refreshed every 30 seconds, they are closely in sync which is why I used them for my testing.

1) Creating a calculated counter that was basically ((last("CounterA")+last("CounterB")+last("CounterC "))/3 should provide an average but instead gave a value that jumped wildly from way below to way above the three individual counters.

2) Thinking that perhaps there was a minor race condition being developed with the calculation I changed the checks to be last("CounterX",0,120) but still had the same problems.

3) Thinking that perhaps my formula was somehow wrong and just wanting to see this work without the variance I changed the calculated value to be just last("CounterA") and instead of seeing a mirror of counter A I saw the same wild swings as from the other counters (copy of the graph generated from this test attached to show the most degenerate case.)

I could understand if test case 1 involved some tricky bits that I needed to compensate for, but there is no reason for test case 3 to have problems.

Any ideas?

[Rudd]

I am doing exactly what you describe in scenario 1. There are some small deviations I put down to rounding and some timing differences when the values get calculated. Zabbix 1.8.4.

I notice you are graphing md0. Apologies in advance but you are not actually monitoring /dev/md0 somewhere as well, and graphing that value instead of the calculated item?