drill-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Adam Gilmore <dragoncu...@gmail.com>
Subject Re: Parquet pushdown filtering
Date Tue, 08 Dec 2015 12:59:51 GMT
That makes sense​​, yep.  The problem is I guess with my implementation.  I
will iterate through all Parquet files and try to eliminate ones where the
filter conflicts with the statistics.  In instances where no files match
the filter, I end up with an empty set of files for the Parquet scan to
iterate through.  I suppose I could just pick the schema of the first file
or something, but that seems like a pretty messy rule.

Julien - I'd be happy to have a chat about this.  I've pretty much got the
implementation down, but need to solve a few of these little issues.


On Fri, Dec 4, 2015 at 5:22 AM, Hanifi GUNES <hanifigunes@gmail.com> wrote:

> Regarding your point  #1. I guess Daniel struggled with this limitation as
> well. I merged few of his patches which addressed empty batch(no data)
> handling in various places during execution. That said, however, we still
> could not have time to develop a solid way to handle empty batches with no
> schema.
>
> *- Scan batches don't allow empty batches.  This means if a
> particular filter filters out *all* rows, we get an exception.*
> Looks to me, you are referring to no data rather than no schema here. I
> would expect graceful execution in this case. Do you mind sharing a simple
> reproduction?
>
>
> -Hanifi
>
> 2015-12-03 10:56 GMT-08:00 Julien Le Dem <julien@dremio.com>:
>
> > Hey Adam,
> > If you have questions about the Parquet side of things, I'm happy to
> chat.
> > Julien
> >
> > On Tue, Dec 1, 2015 at 10:20 PM, Parth Chandra <parthc@apache.org>
> wrote:
> >
> > > Parquet metadata has the rowCount for every rowGroup which is also the
> > > value count for every column in the rowGroup. Isn't that what you need?
> > >
> > > On Tue, Dec 1, 2015 at 10:10 PM, Adam Gilmore <dragoncurve@gmail.com>
> > > wrote:
> > >
> > > > Hi guys,
> > > >
> > > > I'm trying to (re)implement pushdown filtering for Parquet with the
> new
> > > > Parquet metadata caching implementation.
> > > >
> > > > I've run into a couple of challenges:
> > > >
> > > >    1. Scan batches don't allow empty batches.  This means if a
> > particular
> > > >    filter filters out *all* rows, we get an exception.  I haven't
> read
> > > the
> > > >    full comments on the relevant JIRA items, but it seems odd that we
> > > can't
> > > >    query an empty JSON file, for example.  This is a bit of a blocker
> > to
> > > >    implement the pushdown filtering properly.
> > > >    2. The Parquet metadata doesn't include all the relevant metadata.
> > > >    Specifically, count of values is not included, therefore the
> default
> > > >    Parquet statistics filter has issues because it compares the count
> > of
> > > >    values with count of nulls to work out if it can drop it.  This
> > isn't
> > > >    necessarily a blocker, but it feels ugly simulating there's "1"
> row
> > > in a
> > > >    block (just to get around the null comparison).
> > > >
> > > > Also, it feels a bit ugly rehydrating the standard Parquet metadata
> > > objects
> > > > manually.  I'm not sure I understand why we created our own objects
> for
> > > the
> > > > Parquet metadata as opposed to simply writing a custom serializer for
> > > those
> > > > objects which we store.
> > > >
> > > > Thoughts would be great - I'd love to get a patch out for this.
> > > >
> > >
> >
> >
> >
> > --
> > Julien
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message