Thanks Alieh.
I like Lucas' idea about "format" vs "type".
For the downgrade path, KIP-1271 says:
Downgrades are not supported; the only way to achieve one is by deleting the
local store and rebuilding it from the changelog.
While KIP-1285 only say:
The same as in KIP-1271, downgrades are not supported.
Might be good to align both?
One last question: Should we call out that
`Materialized.as(KeyValueBytesStoreSupplier)` (same for window and
session case), will also allow to enable a headers store on a per-store
basis, similar to the plain-kv / ts-case? Ie, if one uses
Materialized.as(
Stores.persistentTimestampedKeyValueStoreWithHeaders(...)
)
The corresponding DSL operator would also use a header store,
overwriting whatever is configured via `StreamsConfig`.
While it's unfortunately not well documented, it is currently possible
to pass in both types of supplier, and the DSL would "bridge" between
both store types. Thus, it's currently possible, to disable ts-stores in
the DSL via
Materialized.as(
Stores.persistentKeyValueStore(...)
)
The determining factor is, if the passed-in supplier implements
`TimestampedBytesStore` or not and we added a similar marker interface
`HeaderBytesStore` with KIP-1271, to we can do the same thing for header
stores.
This should not imply any more code changes I believe, but should work
this way automatically base on KIP-1285 proposal. I think it's mainly an
documentation question.
-Matthias
On 2/22/26 6:16 AM, Alieh Saeedi via dev wrote:
Thanks Lucas for the feedback.
LB1) dsl.store.type makes sense to me, but I’m not attached to it. I can
change it (and already have changed it) to dsl.store.format if we prefer
that naming.
LB2) Since Kafka Streams does not have a timestamped session store, we
don’t have any DslSessionParams constructor with isTimestamped. What we
are deprecating in the KIP is the older constructor that does not take a
DslStoreType (now DslStoreFormat). To keep things cleaner, we only keep a
single constructor that also includes DslStoreFormat. For session stores,
DEFAULT corresponds to the regular session store, and HEADERS corresponds
to SessionStoreWithHeaders. This detail is already mentioned in the KIP.
LB3) Because this KIP does not introduce any new state store formats, the
upgrade and downgrade paths are identical to those described in KIP-1271.
I’ve added this note to the KIP as well.
Thanks,
Alieh
On Fri, Feb 20, 2026 at 9:08 PM Lucas Brutschy <[email protected]>
wrote:
Hi Alieh,
Thanks for the KIP. A few thoughts below.
LB1: The config name dsl.store.type feels a bit ambiguous since we
already have dsl.store.suppliers.class (which controls RocksDB vs.
in-memory). dsl.store.type could be confused with that distinction.
Would something like dsl.store.format be clearer, since this is really
about the value format (timestamped vs. timestamped+headers)?
LB2: Regarding DslSessionParams: the current code entirely lacks an
isTimestamped field or constructor parameter. The KIP shows
deprecating a constructor with isTimestamped on DslSessionParams, but
that constructor does not exist today. Could you clarify whether the
KIP intends to add and immediately deprecate the constructor, or
whether the old DslSessionParams constructor should simply remain
as-is and a new overload with DslStoreType should be added alongside
it?
LB3: Regarding compatibility, the KIP mentions that switching from
DEFAULT to HEADERS triggers KIP-1271 upgrade paths. It would be good
to clarify whether switching back from HEADERS to DEFAULT is also
supported (i.e., is it a two-way migration or one-way only?). If it is
one-way, that must be documented explicitly.
Thanks,
Lucas
On Fri, Feb 20, 2026 at 10:30 AM Alieh Saeedi via dev
<[email protected]> wrote:
Hey Matthias,
Thanks for the suggestions. I agree with all of them. They all make sense
to me. I updated the KIP accordingly.
Bests,
Alieh
On Thu, Feb 19, 2026 at 8:33 PM Matthias J. Sax <[email protected]>
wrote:
Thanks for the KIP. Great follow up to KIP-1271.
Nit: The `DSL_STORE_TYPE_DOC` should be `private` and I also think we
can improve the wording -- but that's more of am impl detail, not
really
critical for the KIP itself.
How about:
"Controls the state store type used by the DSL store supplier (see
config 'dsl.store.suppliers.class'). 'default' uses timestamped stores.
'headers' uses the headers-aware stores."
MJS1: Should we deprecate the existing `DslKeyValueParams(final String
name, final boolean isTimestamped)` constructor, and make the newly
introduced "type" a mandatory parameter? (Cf MJS2 below)
MJS2: Should we add new enum `DslStoreType` with DEFAULT and HEADERS
value, and change the constructor
from
- DslKeyValueParams(final String name, final boolean isTimestamped,
final boolean isHeadersEnabled)
to
- DslKeyValueParams(final String name, final DslStoreType storeType)
This would more align to the newly added String config (and the KIP
calls out that we did not want to add a boolean config, what makes
sense
to me), and would subsume the existing `isTimestamped` parameter on the
existing constructor which we could remove (Cf MJS1)
[We have a similar thing for `group.protocol` config.]
If yes, wondering if we should also deprecate `isTimestamped()` and not
add `headersEnabled()` but instead add `storeType()` getter?
Similar for window/session Dsl*Parameter classes.
-Matthias
On 2/18/26 3:30 AM, Alieh Saeedi via dev wrote:
Hi all,
I’d like to start a discussion on a new KIP, KIP-1285: DSL Opt-in
Support
for Header-Aware State Stores.
KIP-1271 introduced headers-aware state stores in Kafka Streams.
KIP-1285
proposes adding corresponding DSL-level opt-in support, so that DSL
applications can take advantage of these headers-aware stores while
keeping
existing applications and topologies unchanged by default.
You can find the KIP here:
https://cwiki.apache.org/confluence/display/KAFKA/KIP-1285%3A+DSL+Opt-in+Support+for+Header-Aware+State+Stores
Feedback would be greatly appreciated.
Best,
Alieh
