IIUC the issue affects only the Flink Dynamic Sink, OTOH the fix effects
every user.

Maximilian Michels <[email protected]> ezt írta (időpont: 2025. nov. 21., P,
11:16):

> >>
> https://github.com/apache/iceberg/commit/c02d4676f38b834d6c672bb53abd34b07ebfddb9
> >I'm not sure that this is something which could be considered as a bugfix.
> >For me this is more like a feature which is not something we would like
> to have in a patch release
>
> IMHO it is a bugfix to classify append-only snapshots as APPEND
> snapshot instead of OVERWRITE, but I can see how it can be considered
> a feature as well. The larger problem of this is that OVERWRITE
> commits will not be read by Flink in streaming mode due to the lack of
> support for deletes. So this is quite a surprise to users that they
> cannot read back their appended data.
>
> -Max
>
> On Fri, Nov 21, 2025 at 5:02 AM Manu Zhang <[email protected]>
> wrote:
> >
> > Hi JB,
> >
> > We've already excluded Spark's parquet dependency from
> iceberg-spark-runtime[1], and I don't think we should go back and block
> variant/geometry support.
> > On the other hand, it's unlikely for Spark 4.0.x to bump to Parquet
> 1.16.x in a bug-fix release[2].
> > Hence, `useClassPathFirst=true` is the best solution I can see. We can
> add it to our documents.
> >
> > 1.
> https://github.com/apache/iceberg/blob/main/spark/v4.0/build.gradle#L80
> > 2. https://github.com/apache/spark/pull/52165#issuecomment-3240831583
> >
> > Regards,
> > Manu
> >
> > On Fri, Nov 21, 2025 at 2:23 AM Jean-Baptiste Onofré <[email protected]>
> wrote:
> >>
> >> Hi,
> >>
> >> While testing the 1.10.x and main branches, I encountered an issue
> >> regarding Parquet dependency versions that needs clarification.
> >>
> >> I noticed a mismatch in the Parquet versions used by Spark itself and
> >> the Iceberg Spark extension:
> >>
> >> - Spark 4.0.1 uses Parquet version 1.15.2.
> >> - Iceberg Spark 1.10.0 uses Parquet version 1.16.0.
> >>
> >> If I set spark.executor.userClassPathFirst=true, the execution is
> >> fine. However, with the default setting (useClassPathFirst=false),
> >> running table maintenance actions (such as expireSnapshots) results in
> >> a java.lang.NoSuchMethodError:
> >>
> org.apache.parquet.schema.LogicalTypeAnnotation$VariantLogicalTypeAnnotation
> >> org.apache.parquet.schema.LogicalTypeAnnotation.variantType(byte).
> >> This error originates within the Iceberg Spark
> >> ParquetWithSparkSchemaVisitor. So, I suspect a change in variant (and
> >> variant schema) in Parquet.
> >>
> >> This issue suggests there may be an incompatible change between
> >> Parquet 1.15.2 and 1.16.0. Since shading does not seem to resolve
> >> this, I wonder if we should enforce a consistent Parquet version
> >> across both Spark and the Iceberg extension to prevent such conflicts.
> >>
> >> Do you have any thoughts on how to best address this dependency
> >> mismatch for the upcoming 1.10.1 release?
> >>
> >> Regards,
> >> JB
> >>
> >> On Mon, Sep 22, 2025 at 10:29 AM Amogh Jahagirdar <[email protected]>
> wrote:
> >> >
> >> > Hey folks,
> >> >
> >> > Iceberg 1.10 was released 2 weeks ago and there was one issue around
> incorrect variant filtering reported that I think meets the criteria for a
> patch release. The fix PR is in (thank you Drew). I wanted to kick this
> discussion thread off in case folks had other issues in the 1.10 release
> that they think warrant a patch release.
> >> >
> >> > I also think this PR  is a good candidate for a patch release; this
> is for addressing a long-standing issue where closing the S3FileIO during
> an event like moving broadcast variables from memory to disk leads to an
> unexpected closing of the http client. There's still some discussion on the
> approach of the fix but there's general recognition that it's a legitimate
> issue, so I think it'd be ideal to get this in for a patch release as well.
> >> >
> >> > I've also created a milestone here.
> >> >
> >> > Thanks,
> >> >
> >> > Amogh Jahagirdar
>

Reply via email to