TIL thank you.
In the case of aggregate pushdown, can't an engine use inexact min/max? I
understand it will have to scan at least one row group. To be exact it will
have to scan all row groups that match the inexact min (or max) across all
row groups. If we consider that the optimization is intere
>
> Another option is to keep the complexity, make inexact the default (and
> thus not pay bytes for it on the wire) and allow engines to emit exact
> BINARY if they so desire.
I would argue that - for correctness - inexact *has to be* the default, as
legacy writers will not write it. So you canno
Update: The Spark community has successfully closed the vote [1]
to move Variant spec to Parquet. I will start a formal vote on the Parquet
side, though we have already received sufficient binding votes in this
thread.
[1] https://lists.apache.org/thread/pkybo148j6qyn2wsjnmyrhqs3crn9b89
Best,
Gan
No, the proposed approach comes from a quick math from all
received responses which is acceptable by most people.
If people have strong opinions to use a dedicated repo for variant
spec and impl, would it be good to express "+1 for dedicated repo"
in this vote as well?
On Tue, Sep 10, 2024 at 10: