How to interpret AQP errors?

[xuebinsu]

What do 0.0 and NaN mean in the error column? I notice that the reported error is 0.0 while the difference of the approximate result (returned by Verdict) and the true result (return by Spark SQL) is not 0. Why?

[pyongjoo]

Sorry for a late response. I just wanted to let you know that we are working on this issue.

Something we know right now is that our error estimation logic can produce unexpected results (such as NaN or 0.0) when the sample size is small. We will first clarify its root causes and will add some checks to prevent such cases.

[GaoleMeng]

The current progress in fixing this bug:
The cause of the bug is that the subsample size equals to one for estimation, which makes the stddev of this subsample value to be "Nan". We try to add two verdict_default properties to fix this bug:

verdict.error_bound.minimum_subsample_size = 10
This suggests that when the subsample group number is smaller than this number (default for 10), we set the error bound of the value to be -1 (which represent infinity)

verdict.error_bound.trust_error_bound = 0.1
This parameter suggest that when the error is out of 10% of the value, we set it to -1 (which represent infinity)

So before the user sees the actual error bound, we check and revise the value to be either -1 or an interpretable value.

[xuebinsu]

Thanks very much for your work! I'll test it.

[GaoleMeng]

Thx!
In fact, the code is still under review and not merged yet. We will update when we finished soon.