adriangb commented on PR #20481: URL: https://github.com/apache/datafusion/pull/20481#issuecomment-4013608124
> In fact I think a well tuned system should be able to keep the CPU busy almost _ALL_ of the time (what this PR does makes it much better but there is more): That's a good way to look at it! I expect to keep the CPU always feed will require either infinitely fast IO or a lot of pre-fetching / wide IO (with the consequence of higher memory consumption and possibly wasted work). So maybe *fully solid* is a bit of a stretch goal, but a good target to have! > I actually think we could do this fairly efficiently (with relatively low ovehead) using the RowSelection API -- aka fetch the data for a single Row Group but then potentially break that row group up into several smaller morsels (e.g. 10-20 batches worth of rows) 🤔 I was playing around with the opposite: the idea that we'd pack multiple row groups into a morsel if they are small. I feel that ideally a morsel would be defined in terms of ~ units of work (which is difficult because we'd have one metric for both CPU and IO work; maybe "bytes to scan" is good enough) and not row groups or rows. So if row groups are small we pack several into 1 morsel. If row groups are large we split them up into multiple morsels using RowSelection. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
