Hi, Lu.
Thank you for your proposal. I believe it is very helpful for joins to improve
performance. After reading the Flip, I have the following questions and look
forward to your reply.
1. I would like to confirm what happens when an Iceberg table is not a
partitioned table but still supports SupportsPartitioning. Will
SupportsPartitioning#outputPartitioning return empty?
2. In KeyGroupedPartitioning, could you explain why partition values need to be
ordered?
3. Can you outline the relationship between the parallelism of join and the
data sizes and number of partitions of the left and right tables?
4. Nit: Regarding the Flink flip, it is not necessary to specifically list out
the changes that Iceberg needs to adapt to. If needed, these can be added as an
appendix at the end.
5. In the "Partition Mismatch Problem" section, how do we inform table B that
it needs to create an empty partition?
6. In the "Join Keys and Partition Keys Relationship" section, I noticed that
when the join keys are a superset of the partition keys, c1 is still used as
the partition key. Do you mean we will treat “A.c2 = B.c2” as a normal
condition and apply this filter after the join?
7. Considering the "Correlated Partition Number", could we have something
similar to `resolveCommonPartitioning` that allows the source to assist in
merging two partitionings when their partition numbers are inconsistent or
their partition values differ?
8. In the "Compatibility, Deprecation, and Migration Plan" section, I noticed
you mentioned that “The feature will be opt-in through configuration, ensuring
no impact on existing workloads.” However, I did not find the relevant
configuration in the public interface—could you please elaborate on this?
9. Is it possible to eliminate the exchange between the source and join using
FlinkExpandConversionRule? This seems more general and could benefit other
operations like rank, group agg etc. as well.
--
Best！
Xuyang
At 2025-06-26 06:36:38, "Lu Niu" <qqib...@gmail.com> wrote:
>Hi devs,
>
>I’d like to start a discussion on FLIP-535: Support Storage Partition Join
>in Flink:
>
>https://docs.google.com/document/d/1bX37wpwivO3fS4WrR9tZ0pDF92Mpux6pZE54Tw8czzU/edit?tab=t.0#heading=h.goqvqbz0cyel
>
>
>The goal is to support storage partition join in Flink batch mode for
>iceberg tables to avoid unnecessary shuffles.
>
>It would be great if someone could help to copy the contents to a FLIP
>page. I don’t have the permission.
>
>Looking forward to your feedback.
>
>Best
>
>Lu
