Austin Godber
Austin Godber
The first step here is to create a semi-realistic recovery scenario with real data and see if it fails. We may need to do this on internal dev clusters but...
Maybe using chaosmesh in KIND is the way to go.
There were a few other edge cases Kimbro and I discussed yesterday: What happens on jobs with: * multiple slicers? * more slices than the length of the slice queue?...
Your tests are being performed on the master branch, this problems happened on `v2.6.3` or `v2.9.2`. I regret not capturing that info at the time of writing this issue. It's...
We have had a second occurrence of this issue on Feb 21st.
The specific scenario we have been doing this manually in is a case where we have an ES cluster of a fixed size and we write daily indices into it....
Can we constrain it so that `percentageKept` never goes to zero or below something?  If I am understanding this correctly, I am taking the cumulative sum of the bytesThisWindow,...
> Can we constrain it so that percentageKept never goes to zero or below something? I don't like the idea of some periods having zero records, that misrepresents that window...
I think this is sufficiently done for now, it's just a matter of me deploying a convincing test.
Without traceability to which specific records in the kafka case, I think the benefit of this is pretty marginal.