Description
In the event of an outage in the ML node there is a skew between the time of the model snapshot and the time of the datafeed/data counts.
When the job gets started up again the datafeed starts feeding data starting from time x, while the model is at time y, and y < x. The difference between y and x can be worst case 3 hours.
The datafeed sends a skipTime control message to notify autodetect about the skew, notifying that it starts data from time x instead of y.
Still, for the time window [y, x] autodetect writes model plots, usually 'zero' results (actual == 0), this is technically correct but has side-effects: On the plugin side the empty results can overwrite already existing model plots written before the outage with actuals !=0.
Therefore we suspect a bug in skipTime being not effective for model plots.
Needs repro+debug.
Related issue: elastic/elasticsearch#31867