Github user thunterdb commented on the pull request:
https://github.com/apache/spark/pull/11553#issuecomment-193889292
If you use 0.0 for the relative error, it is going to return the exact
quantiles. However in this case, there will be no data compression and the
algorithm will essentially run distributed merge-sort. The cost of doing this
may be prohibitive.
The current implementation is already approximate, so it makes sense to
pick a relative error > 0. It looks like we do not offer any guarantee in the
approximation already anyway. I suggest we automatically pick the relative
error as follows:
> target_error = min(0.1, 1.0 / (alpha * num_buckets))
for some value of alpha (1-10): the run time is proportional to the number
of buckets, and the precision increases with the number of buckets required.
If this is considered too complicated, then a fixed value such as 1e-2 or
1e-3 looks fine to me as well.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
