Il giorno ven 21 set 2018 alle ore 20:03 Enrico Olivelli <eolive...@gmail.com> ha scritto: > > > > Il ven 21 set 2018, 19:45 Julian Hyde <jh...@apache.org> ha scritto: >> >> Well, maybe we can. >> >> But look at the premise of the question: "Problem with 'assert' >> ignored in production, leading to inconsistent query results". The >> problem is not with the assert. > > > Yes actually the bug is not clear. > > We have a merge join with 2 unsorted inputs. I don't know if the problem is > in the planner (why it had choosen a merge join and not a simple joon ) or in > the translation from planner output to actual plan > > I will be back with news next week. > But I will start a new thread. > > Thank you for you comments and suggestions. > I will evaluate if removing the assert with introduce too much cost > (comparing large strings for instance)
This is the a patch, the assertion is an "integer < 0" check so it is very cheap in Java, I think it is worth https://issues.apache.org/jira/browse/CALCITE-2591 I have started in another thread the discussion about the actual bug which led to such situation Thank you Enrico > > Enrico > >> On Fri, Sep 21, 2018 at 10:23 AM Vladimir Sitnikov >> <sitnikov.vladi...@gmail.com> wrote: >> > >> > Julian>We can't afford to check assumptions in performance-critical code. >> > >> > We definitely can afford checks that take a couple of CPU cycles to >> > perform. >> > >> > Vladimir > > -- > > > -- Enrico Olivelli