deriv() returns non-zero values even though my data hasn’t changed

[stapelberg]

Originally I thought this is a regression, but I see no proof of that — as per syslog, I’ve been using prometheus, version 0.18.0 (branch: stable, revision: f12ebd6) since the 15th of May, and the issue only now started to appear.

I have an alerting rule which looks at changes in state, see https://github.com/robustirc/robustirc/blob/669a278acd264e1e94c7ad8e7a6217be54585488/contrib/prometheus/robustirc_prometheus.rules#L19 for details.

This rule is now firing, even though there weren’t any changes in state.

When having a closer look, it turns out that deriv() returns values that are very close to 0 (e.g. -0.000000000006047549185814796), but not 0.

I ran the following two commands to capture the data and reproduce the issue:

wget -qO- http://localhost:9090/api/v1/query_range\?query\=raft_isleader\{instance\=\"ridcully\"\}\&start\='2016-05-19T18:00:00Z'\&end\='2016-05-19T22:00:00Z'\&step\=15s > /tmp/values.txt
values.txt

wget -qO- http://localhost:9090/api/v1/query_range\?query\=deriv\(raft_isleader\{instance\=\"ridcully\"\}\[10m\]\)\&start\='2016-05-19T20:00:00Z'\&end\='2016-05-19T22:00:00Z'\&step\=15s > /tmp/deriv.txt
deriv.txt

I would have expected deriv() to return 0 in this specific case.

Am I holding it wrong or is this a bug? If the former, do you have any suggestions to improve my alerting rule?

[brian-brazil]

That sounds like floating point inaccuracy, not much we can do about it.

Have you considered using the changes function?

[stapelberg]

Ah, changes does return 0. Does changes not involve floating point operations, or why does it not suffer from the same inaccuracies as deriv?

[beorn7]

On 20 May 2016 at 08:48, Michael Stapelberg notifications@github.com wrote:

Does changes not involve floating point operations, or why does it not suffer from the same inaccuracies as deriv?

Everything in Prometheus is floating point. changes, however, does
something really simple, while deriv does
https://en.wikipedia.org/wiki/Simple_linear_regression

There are way more chances of running into accuracy issues there.

Björn Rabenstein, Engineer
http://soundcloud.com/brabenstein

SoundCloud Ltd. | Rheinsberger Str. 76/77, 10115 Berlin, Germany
Managing Director: Alexander Ljung | Incorporated in England & Wales
with Company No. 6343600 | Local Branch Office | AG Charlottenburg |
HRB 110657B

[stapelberg]

Thanks for the explanation. I’m closing this issue as we found an alternative solution (changes) and can’t do much about floating point inaccuracies.

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.