Dataiku Checks Alerts and Notification

[lina]

Hi everyone,

I would like to receive a notification on Dataiku on the receiving errors in checks and metrics. Is this possible? If yes, how can I implement this?

For example, receive a notification once the value is greater than x

Thanks

[Marlan]

Hi @lina,

Couple of assumptions about what you are wanting to do. First, I assume you just want notification if a value is greater than x but you don't want to fail the process. Second, I assume you want to do this from a scenario. 

Is so, here are some notes on how to do this.

Define metrics and checks according to your needs. The check should have Soft maximum value set.

In a scenario that builds the datasets of interest, include a Compute metrics step followed by a Run checks step at the end. 

Then to receive an email when the check triggers a warning, use warningStepsCount > 0 as the Run condition in the mail reporter. You should receive an email when the value of the metric exceeds the value you specified in Soft maximum.

Note that scenarios that trigger Check warnings will have an outcome of SUCCESS unless the Run Checks step has the "Fail on warnings" box checked.

Marlan

[lina]

Thanks Marlan, is there a way to receive in-application notifications instead of emails though?

[Marlan · Accepted solution]

Hi @lina,

if you want a build to fail if a dataset doesn't meet checks, then yes. If you are after warnings, the answer is sort of.  

In either case, you'll need to turn on to auto-compute/run after build for both the metrics and the checks.

If the check is a regular check (value must be > minimum and/or < maximum) then a build of the dataset initiated from the flow that fails the check will fail and the error message will indicate this. So this works as one would expect.

If the check is set up as a warning (using soft minimum and/or soft maximum) then the build will succeed but if you check the job log, you'll see a message in orange saying that the checks produced a warning. Not ideal certainly as you'd probably not notice unless you were looking for it and if you had warning checks on multiple datasets then it'd be a pain to check all of the job logs.

It'd be better if there was a more clear / noticeable way of communicating that a job succeeded but that warnings were generated.

Marlan