Thanks Sophie, Woah, that’s a nasty bug. Rather than working around it, I think we should just fix it. I’ll leave some comments on the Jira.
It doesn’t seem like it should be this KIP’s concern that some serdes might be incorrectly written. Were there other practical concerns that you had in mind? Thanks, John On Tue, May 19, 2020, at 19:10, Sophie Blee-Goldman wrote: > I like this "fancy idea" to just flip the to/from bytes but I think there > are some practical limitations to implementing this. In particular > I'm thinking about this issue > <https://issues.apache.org/jira/browse/KAFKA-8159> with the built-in signed > number serdes. > > This trick would actually fix the problem for negative-negative queries > (ie where to & from are negative) but would cause undetectable > incorrect results for negative-positive queries. For example, say you > call #range with from = -1 and to = 1, using the Short serdes. The > serialized bytes for that are > > from = 1111111111111111 > to = 0000000000000001 > > so we would end up flipping those and iterating over all keys from > 0000000000000001 to 1111111111111111. Iterating in lexicographical > order means we would iterate over every key in the space *except* for > 0, but 0 is actually the *only* other key we meant to be included in the > range query. > > Currently we just log a warning and return an empty iterator when > to < from, which is obviously also incorrect but feels slightly more > palatable. If we start automatically converting to reverse queries we > can't even log a warning in this case unless we wanted to log a warning > every time, which would be weird to do for a valid usage of a new > feature. > > All that said, I still like the idea overall. Off the top of my head I guess > we could add a store config to enable/disable automatic reverse iteration, > which is off by default? > > Thanks for the KIP! This will be a nice addition > > Sophie > > > On Tue, May 19, 2020 at 3:21 PM John Roesler <vvcep...@apache.org> wrote: > > > Hi there Jorge, > > > > Thanks for the KIP! > > > > I think this feature sounds very reasonable. > > > > I'm not 100% sure if this is "too fancy", but what do you think > > about avoiding the enum by instead allowing people to flip > > the "from" and "to" endpoints? I.e., reading from "A" to "Z" would > > be a forward scan, and from "Z" to "A" would be a backward one? > > > > Thanks, > > -John > > > > On Tue, May 19, 2020, at 16:20, Jorge Quilcate wrote: > > > Hi everyone, > > > > > > I would like to start the discussion for KIP-617: > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-617%3A+Allow+Kafka+Streams+State+Stores+to+be+iterated+backwards > > > > > > Looking forward to your feedback. > > > > > > Thanks! > > > Jorge. > > > > > > > > > Attachments: > > > * 0x5F2C6E22064982DF.asc > > >
