Mismatch of chi_squared_cdf() result from Presto Java (within epsilon) about velox HOT 4 OPEN

Why are we marking this as mismatch in the first place? Floating point numbers are compared with delta within epsilon, no?

from velox.

CC: @Krishna-Prasad-P-V

from velox.

Why are we marking this as mismatch in the first place? Floating point numbers are compared with delta within epsilon, no?

@Yuhta Unfortunately, our result comparison methods in QueryAssertions.cpp didn't compare this result with epsilon because the rest non-floating-point columns doesn't form unique keys. This is a limitation of our QueryAssertions, so it would be good to extend it.

from velox.

Here is where the condition for epsilon comparison was not satisfied in QueryAssertions:

QueryAssertions requires every row to have distinct combination of values at non-floating-point columns so that it can find the corresponding rows between the expected and actual result vectors. This condition was not satisfied in the fuzzer iteration of chi_squared_cdf(), so epsilon comparison was not used. One solution would be to add an additional "row_number" bigint column to the input vector and project this column to the result as well. Then this column would be able to distinctly identify the corresponding rows between the expected and actual result vectors and hence enable the epsilon comparison.

from velox.