davisp commented on issue #866: Fix incorrect index selection when sort specified URL: https://github.com/apache/couchdb/pull/866#issuecomment-335570620 Couple things to unpack in the discussion: First, for #816 I think that's actually the opposite, its not whether it needs a new version bump but how many versions should be be back porting this fix to. A database that gives bad answers to queries is more a bug that needs to be patched backwards, not something we cut the cord on and just bump version numbers. There is of course the issue around notifying users that a pending change may break their application (technically fix, but users have a different perspective :), so this is definitely something we should talk about. We'll also want to update the documentation as we've long said that using column prefixes was a suggestion for sharing indexes so that'll need to be fixed. For @tonysun83's example of this changing a query from something that uses _all_docs to something that uses an existing index I don't think that's an issue. Originally I would have ixnayed it because of the lack of range restriction on A and B but we've already relaxed that by having an _all_docs default. Also the behavior here is different than the issue introduced by #816 in that previously Tony's example would have been an incorrect sort returned where the new version changes that (ie, same set of docs would be returned both places, just now its sorted). Which I think is quite a bit different than subtley different sets of documents returned. Also, if a user specifies a sort and isn't detecting the wrong sort order based on the _all_docs fall back then I'm failing to see how they'd be relying on that sort, and thus fixing it seems like not an issue. For #816 with the _all_docs fallback is that also not more of a concern about performance rather than an app that stops working? Ie, all queries will continue to work and be "fixed" but some unknown number of queries will suddenly switch to using _all_docs and become much slower? Granted suddenly slower performance that lasts until an index can be built isn't nothing either. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
With regards, Apache Git Services
