Re: [DISCUSS] KIP-1238: Multipartition for TopologyTestDriver in Kafka Streams

[Matthias J. Sax]

Thanks.


MJS8 (follow up): For `equalsIgnorePartition()`, it seems we could use 
`TestRecord<K, V>, instead of undefined generic types? -- If the types 
don't match, we cannot really expect that both records are equal to each 
other to begin with, and the test code seems to be incorrect to begin 
with, and should not compile? -- Or is the test scenarios I miss, for 
which `TestRecord<?, ?>` would be required?



MJS8 (a). Edit suggestion. The KIP says:

Existing code that does not set a partition explicitly will produce partition = 
-1 on both sides of an equality check, preserving backward compatibility.

This applies to input TestRecords. It might be helpful to also add a 
sentence for output TestRecords for existing code, which runs in 
single-partition mode: all `TestRecords` would get partition=1, so it's 
also backward compatible.



MJS11. The KIP says:

Topics not declared via declareTopic() are automatically assigned 1 partition. For example, if 
declareTopic("a", 3) is declared but topic "b" is not, topic "b" will have 1 partition. 
The driver will still operate in multi-partition mode since topic "a" was declared with more than 1 partition.

Do I read this right, that we would allow to have not declared topics in 
multi-partition mode? Just wondering if we can actually implement it? My 
understanding was, but I could be wrong, that for multi-partition mode, 
it would only work correctly if all topics got declared? At least, I 
would believe all input topics need to be declared upfront? I can see 
that output topic don't need to be declared and can safely default to 
having one partition.

I am totally ok with supporting not declared topis, just want to double 
check that the KIP does not say something we cannot build.




-Matthias


On 6/11/26 1:14 PM, Sébastien Viale wrote:
Hi all,

Thank you for the follow-up.

MJS6 (follow up)

We decided to go with the builder as the single entry point, with implicit
mode detection based on declared partition counts (at least one topic
declared with more than 1 partition triggers multi-partition mode)

Indeed, we think that a single entry point is simpler for users.

We also propose to deprecate TopologyTestDriver constructors in order to
avoid having to ways to declare TDD

I have updated the KIP accordingly.


MJS8

You are right, I replaced Object by TestRecord


MS10: Nit

Fixed also


cheers !




Le lun. 8 juin 2026 à 02:16, Matthias J. Sax <mj...@apache.org> a écrit :

Thanks.



MJS6 (follow up): The KIP says

Multi-partition mode is implicitly enabled when at least one topic is
declared with a partition count greater than 1.

and

If all topics are declared with a partition count of 1, or no topics are
declared at all, the driver behaves identically to the legacy constructors

I am wondering if this make it more complex for users to reason about
the mode? Would it be simpler to use singe-partition mode only for `new
TopologyTestDriver` and multi-partition mode always when the new builder
is used?

For this case, we would be in multi-partition mode even if all declared
topics have 1 partition. -- And if the builder is used, we would always
require all topics (even if all have 1 partition) to be declared.

We would also disable the `getStateStore()` methods w/o a partition
parameter when the builder is used, and only support them for `new
TopologyTestDriver` (which is the only way to use single-partition mode).

Given that we don't propose to deprecate `new TopologyTestDriver`, I
believe this design would be cleaner and easier to understand for users?
Two modes, and two different entry points:

   - new TopologyTestDriver == single-partition mode
   - Builder == multi-partition mode

For this case, we would never deprecate `new TopologyTestDriver` I guess?

If we want to support single-partition mode using the builder, I would
propose to deprecate `new TopologyTestDriver` with this KIP. No need to
keep it? Or is there a reason why we would defer this to a follow-up KIP?



MJS8: Given that `equalsIgnorePartition()` is newly added method, it
could take `TestRecord` as parameter? No need to go with `Object` (only
required for `equals()` to use `Object` as we inherit it).



MS10: Nit.

Only after build() can records be piped and outputs be read.

This is obviously true, because only `build()` create the TTD object
allowing to create `TestInput|OutputTopics`. Guess it's an relict from
the previous version of the KIP w/o the builder class.






On 6/5/26 12:09 AM, Sébastien Viale wrote:
Hi all,

MJS6
Actually yes, the single-partition mode will behave the same as
multi-partition mode with one single partition for all topics
So yes, the TopologyTestDriver constructors could be deprecated / removed
in the future.

MJS7:
Good catch indeed.

We can use -1 as default the default value to clearly indicate that no
explicit partition was set, making the routing strategy unambiguous

Concretely:

     - Constructors without explicit partition should default to -1.
     - In pipeInput:
        - if partition == -1: delegate to TopologyTestDriverPartitioner
        - if partition >= 0: use it directly (with range validation)

equals() / hashCode() methods include the partition field:

     - existing code will now also produce -1 when no partition is set
     explicitly
     - backward compatibility for equality checks is preserved

Finally, -1 never appears in output records (readRecordsToList()), since
those always carry the resolved partition, it only exists as an input
sentinel value indicating “no explicit partition”.

I'll update the KIP accordingly.
thanks

Le jeu. 4 juin 2026 à 21:08, Matthias J. Sax <mj...@apache.org> a écrit
:

Thanks for the update.


Few more questions:

MJS6: Would it be simpler to (in the long run) drop single-partition
mode all together? We would only keep it now for backward compatibility
if `new TopologyTestDriver()` is used, but would always use
multi-partition mode when then new builder is used?

It seems to add complexity to keep single-partition mode for the builder
case?

If we say, the new builder always does multi-partition mode, and we
deprecate/remove `new TopologyTestDriver()` later, single-partition mode
would get removed, too.

Or is there something specific about single-partition mode that makes it
worth keeping for the new builder case? I would believe, with the only
exception that multi-partition mode does create more tasks, if all
topics are defined with a single partition, the record piping would be
the same as in single-partition mode? Or is there a difference?



MJS7:
Update equals() and hashCode() to include the partition field. For
records created via existing constructors, partition defaults to 0,
preserving backward compatibility for existing assertEquals calls.

Should it default to `-1` (or `Optional.empty()`) instead of `0`? Making
the default `0` seems to not work with "multi-partition mode" but break
it -- as it would skip the hashing/partitioning, as `0` would be set
explicitly?

And for `-1` as default and topics with a single partition, `-1` will
get hashed/mapped to partition-0 correctly.




-Matthias


On 6/4/26 5:47 AM, Sébastien Viale wrote:
Hi,

Thanks for the feedback and suggestions.


MJS1:

I fixed the typo `woraround` => workaround`


MJS2:

I fixed the inconsistency:

      - removed duplicate entry for getKeyValueStore
      - added missing *WithHeaders variants for state store APIs


MJS3:

The existing usages of TopologyTestDriver remain fully unchanged. In
particular, current code relying on the default behavior (i.e., no
explicit
topic declaration) will continue to work as before.

Topic declaration is only required when a test explicitly opts into
multi-partition behavior. Otherwise, topics continue to be created
implicitly with the current semantics.

We will introduce a TopologyTestDriverBuilder class:

public class TopologyTestDriverBuilder {
       public TopologyTestDriverBuilder(Topology topology);
       public TopologyTestDriverBuilder withConfig(Properties config);
       public TopologyTestDriverBuilder withInitialWallClockTime(Instant
initialWallClockTime);
       public TopologyTestDriverBuilder declareTopic(String topicName,
int partitions);
       public TopologyTestDriver build();
}

In this model, multi-partition mode is implicitly enabled when at least
one
topic is declared with a partition count greater than 1.

As a result, the explicit withMultiPartitionMode() is no longer needed.

The existing TopologyTestDriver constructors remain unchanged for
backward
compatibility.

If it is ok for you all, deprecation of those constructors is
intentionally
deferred to a follow-up KIP to keep the scope of this change focused.


MJS4:

For this KIP, we wanted initially to make the partitioning behavior
deterministic and test-controlled.

Supporting partitioner.class would indeed be a natural extension, and
aligns well with production behavior. We propose deferring this to a
follow-up KIP to keep this one focused

However, we will implement the internal routing logic with an internal
TopologyTestDriverPartitioner that implements Partitioner. Which would
provide a natural extension point for future support of custom
partitioners.


MJS5:
-

Your proposal seems like the best compromise:
- - TestRecord.equals() and TestRecord.hashCode() will be updated to
include the partition field.
- For existing code using the current constructors (without a partition
argument), the partition defaults to 0, preserving backward
compatibility
for existing assertEquals() assertions.

- - A new utility method TestRecord.equalsIgnorePartition(TestRecord
other) will
be added for cases where the test author does not want to assert on
partition placement.
- This allows use with
assertTrue(expected.equalsIgnorePartition(actual)) when
partition is irrelevant to the test.
-
- The KIP has been updated.
-
- Cheers


Le mer. 3 juin 2026 à 22:56, Matthias J. Sax <mj...@apache.org> a
écrit
:

Thanks for the great discussion and KIP refinements.

Couple of comments:


MJS1: (nit) there is a typo `woraround`


MJS2: the KIP lists `getKeyValueStore` twice, and it's missing the
newly
added `*WithHeader` variants.


MJS3: About the builder pattern. In general I like it, but it seems
the
applied patter is a little bit odd, and also not backward compatible
(or
maybe I misunderstand the proposal)?

Existing code, would only call `new TopologyTestDriver()`, and it
seems
this code would break, and would need to get updated to add the
dedicated setup phase? W/o declaring the topics explicitly, it seems
later calls to `createInputTopic` and `createOutputTopic` would fail
(at
least on what the KIP says I believe; could also be a phrasing problem
only)?

I also think, that we would need to have some "builder helper class".
Adding the methods to `TopologyTestDriver` can leave the object in an
weird state. A classic builder pattern would suggest, that we add

public class TTDBuilder { // or similar/better name

      public TTDBuilder(Topology topology);

      TTDBuilder withConfig(Properties config);
      TTDBuilder withInitialWallClockTime(Instant
initialWallClockTime);

      TTDBuilder declareTopic(String topicName, int partitions);

      TopologyTestDriver build();
}

For this case, I don't think that we even need
`withMultiPartitionMode()`?

We would also deprecate the existing constructors `new
TopologyTestDriver`, which we still need to support thought. For `new
TopologyTestDriver()` we would just accept all topic names for
`createInputTopic` and `createOutputTopic` and create these topic
with a
single partition. For this case, we would also still allow to use the
existing `getStateStore()` methods, which we might also want to
deprecate?

(Maybe this is already proposed, but the KIP itself is not 100% clear
about it -- at least to me.)



MJS4: about partitioning; if the passed `Properties` configure
`partitioner.class`, I think we could just use it allowing to plugin a
global customer partitioner? Could also be a follow up KIP (maybe
including the ability to define a customer partitioner per input
topic).

In general, I am wondering if we should implement the discussed
partition-routing strategy just as an internal `TDDPartitioner extends
Partitioner` (of course, this is an implementation detail). This would
setup the code nicely to support `partitioner.class` config.



MJS5: About testing and the new `partition` field on `TestRecord`. I
think that both `equals()` and `hashCode()` should be considered. For
existing code, all `TestRecord`s get set partition=0 and thus it's
backward compatible. For new code, checking the expected partition
sound
desirable?

I still believe that for some tests, the partition is not necessary.
For
this case, we could add
`TestRecord.equalsIgnorePartition(TestRecord)`,
and users can use `assertTrue()` for this case?




-Matthias




On 6/1/26 10:46 AM, Sébastien Viale wrote:
Thanks for your response.

BB1:
I have removed the contradictory statement.

LB5:
OK, I will keep the KIP as it is.

I will call for the vote

Cheers

Le lun. 1 juin 2026 à 16:33, Bill Bejeck <bbej...@apache.org> a
écrit
:

Hi Sebastian,

Thanks for the KIP! This will be a great, much needed addition.
I realize I'm a bit late to the discussion.  Overall things look
good
to me
and I have one nit comment.  I also think the KIP is ready for a
vote.

Regarding the treatment of null keys I agree with Lucas's comment
for
the
treatment of null keys (round-robin) distribution.
In  "Partition Routing" section point 2 seems to reflection this
change
but
then then it concludes with "a null key goes to a default partition
(0
in
the test driver)" . I find this a little contradictory, can we
either
clarify or remove the statement?

Thanks,
Bill

On Mon, Jun 1, 2026 at 8:33 AM Lucas Brutschy via dev <
dev@kafka.apache.org>
wrote:

Hi Sebastian,

Thanks for the response!

LB5: I was thinking of this solution but I think it introduces more
problems than it solves. Long-term, I feel it would be a footgun if
people try to compare test records _with_ partitions. And the
current
implementation with a reasonable default partition should already
be
backward compatible. There are other possible solutions, like a
custom
assertion or an operation to "wipe" the partition, but they are
probably not worth the effort. As I mentioned before, I don't see a
better solution to this -- I would just keep the KIP as it is. I
mostly included it in the discussion to make people aware and maybe
spark an idea in someone.

Cheers,
Lucas