hi, do you have a reproducer?

mmm in cloud? no but I can try to make.. I opened ticket here

ok, I see. 6000 selected columns explains it: questdb generates java bytecode to speed up the query, but there is a limit on max. size of the bytecode. with 6000 columns you exceed this limit.

😞 6000 is just the beginning 😉 can it be increased somehow?

in theory - yes. but it would be more complicated than just increasing a constant. as the the limit is inherited from the JVM platform. chances are there is a better solution for what you are trying to achieve. I understand you treat AVG as some sort of a checksum. are these 6000 columns really physical? or it’s a result of many many nested LT JOINs?

Is this limit set by number of characters in the select query? what is the limit exactly? I will try to construct a query that don't exceed the limit

no, the limit is not directly related to SQL text length. each grouped column contributes some amount of bytes to the generated bytecode size. by halving intervals you could try to find how many column you can group in one go. I have you say you are really good in stressing questdb and coming up with unusual testing scenarios:))

I know ;] you guys should start paying me 😉

one consideration:

sum()

(and hence

avg()

) on floating point numbers is not guaranteed to be associative. in other words:

(1.0 + 1.1) + 1.2

does not necessary yield precicely the same result as

1.0 + (1.1 + 1.2)

as anyone with basic math knowledge could expect:) (well, it does in this case, but you get the point). this is not QuestDB-specific, it’s due to how floating points numbers are represented internally inside a computer. this may or may not be an issue for your use-case.

hmm as long as its doing it consistently across all calculations i think its ok

to calculate “column similarity”? what would be the advantage compared to

avg()

?

that it also captures some of the differentially within the results of each column?

yes. but you can find another example where stddev will match while average will be different.

yeah right 😕 there is no hash function?

nope. feel free to open a feature request, that could be a nice addition!