I've been looking through some stats in a customer's cluster and have noticed a lot of time is being spent performing scroll operations, according to indices.search.scroll_time_in_millis.
There is a legitimate use-case for this but I'm not sure what time actually gets reported so I'm not certain if the large amount of time spent doing them is really an issue.
I thought it could be the time spent doing the associated fetches but there's already indices.search.fetch_time_in_millis. Is scroll fetch split out? Does the scroll timeout play a part in how the number is reported?
Am I right in thinking the number never gets reset and will cover the lifetime of the index? Is there a reference I can see to give some more in-depth information on how all the stats in the API are calculated?
Thanks in advance,
George.