How to use spike kill?

[knebb]

Hi,

I have seen several topics regarding the spike kill and most of them end in "was configured somehow wrong, do not use if you do not know what you are doing".

So I am trying to unterstand this feature.
I have the following graph with two wrong data entries causing a graph spike:

cacti_graph_200.png (21.73 KiB) Viewed 1011 times

Whatever "Spike Kill" option I choose in the WEb-GUI, it does simply: nothing.
DRy run tells me:

Code: Select all

NOTE: NO Standard Deviation found in '/usr/share/cacti/rra/digitemp_digi_out_temp_478.rrd'

So my question is about what I can configure or change so this feature will kill such spikes.

Any hints here?

/KNEBB

[TheWitness]

Much of this is knowing how standard deviations and variances work. You should do some research.

[knebb]

Well,

I am aware how they work. But it does not at all explain why the spikes do not get killed here. If you like we can do some calculations:

Taking just this graph into account (24hrs) and without knowing the exact numbers we can assume most of the values are around 2 to 5 degree. With one value every five minutes I am at approximately 1,437 items (subtracted the two spikes and the one missing value as seen in the graph). To simplify let us assume these items only would show 5 or 2 degree in an (more or less) equal distribution. BTW: variance is much higher with this simplifying assumption compared to the real values.

So we cann assume an average expectation value of ~3.5 degrees calculated by: ( 5⋅718 + 2⋅719 + 0⋅1 + 28.69⋅2 ) / 1440 ~ 3.53
Then let's calculate the variance based on this formula: ∑i(xi−μX)2⋅P(X=xi) which results in an variance of ~3.13. The standard devation the is the square root of the variance so we are at ~1.77 degrees.

Well, that's absolutely expected so far. Every value in the range of 3.53 - 1.77 = 1.76 and 3.53 + 1.77 = 5.3 would be considered as "correct". Values below 1.76 or above 5.3 should be considered as "unusual" (or: as spike).

So back from theory:
Obviously the two values with ~ 28.69 degrees are "unusual" and should be killed. But the are not.
Which values takes the algorithm into account? Which values should I change here to get theses spikes killed?

What is the meaning of "3 (-200) Standard Deviations)", "Variance Percentage (100%-500000%)" and "Variance Outliers (3-1000 High/Low Samples)"?

Any answer except "42"?

Thanks!

/KNEBB

[TheWitness]

What does the dry run show in the reports?

[knebb]

HI,

unfortunately the graph has moved since my post so the values are merged anyways. But the spikes can still be ssen:

cacti_graph_201.png (26.06 KiB) Viewed 959 times

See the attached screenshots:
StdDev:

DryRunStdDev

cact1.png (33.03 KiB) Viewed 959 times

And variance:

DryRunVariance

cacti2.png (33.59 KiB) Viewed 959 times

The spike (much lower as before!) is still there....

[knebb]

Hi,

anyone with some further explanations here? I do not get the numbers shown in the dry run screen. Where are they coming from?

Thanks!

/KNEBB

[zuka1337]

Well, I've always had problems with spikekill..

I did my own version with python you can find it here https://github.com/zuka1337/Spikepyrrd

3 things:

• Read the code understand what's there, then use it.

• Before you tested make an backup of the original

• Pay attention to number 1 and 2.