Data Streaming Summit 2026 Spotlights Pulsar 5.0 and Ursa

A smaller conference with a bigger argument

StreamNative hosted the Data Streaming Summit at San Francisco's Hotel Nikko on October 7–8. Attendance was modest, with tickets listed between roughly $107 and $268 on Eventbrite.26 The company is the main commercial backer of Apache Pulsar. This year it renamed the event "The Data Streaming + Agent Infra Conference" and split the program into three tracks: streaming engines, streaming for agents, and agent runtimes and governance.19 Speakers came from Apple, OpenAI, Uber and RisingWave, and AWS, RisingWave and Ververica were the top-tier sponsors.2

The summit itself was small, but the questions it raised are among the most contested in data infrastructure this year. Who owns the path from a live event to a lakehouse table? Should streaming remain a separate layer at all? Snowflake and Databricks spent the first half of 2026 answering those questions their own way. Read together, this week's announcements look like the independent streaming vendors' reply.

What shipped: Pulsar 5.0 and its scalable topics

The biggest database-infrastructure release at the event was Apache Pulsar 5.0. The Apache community published it on October 5 as a long-term support release, ready for production, following two milestone builds.32 Its main feature is "Scalable Topics," a new topic type that grows and shrinks with load while keeping messages for each key in order. It ships with a new Java client API, and Oxia is now the recommended metadata store for new clusters.32

This fixes an old problem with partitioned logs. In Kafka and in Pulsar's existing partitioned topics, keys are assigned to partitions by modulo hashing. Adding partitions moves keys to different partitions, which breaks per-key ordering, and partition counts can never be reduced.31 Pulsar 5.0 replaces fixed partitions with key-range segments. These split automatically when they get busy and merge when traffic drops, and consumers finish reading an old segment before moving to the segments that replace it.32 Yaroslav Tkachenko attended and called scalable topics the release's biggest innovation. He noted that the design borrows from LinkedIn's Northguard and from Pravega.5 StreamNative's own write-up credits Northguard specifically for a constraint on which segments can merge with each other.31

The limits deserve attention. Scalable topics do not support geo-replication or replicated subscriptions in 5.0. Only the Java v5 client supports them for now, and Python and Go Functions cannot use them.32 StreamNative's May roadmap had already said that some current Pulsar strengths, especially geo-replicated subscriptions, would take more releases to catch up.31 Most teams should take that at face value: upgrade clusters now for the performance gains on existing topics, and wait for later releases before moving multi-region workloads to the new model.32

Tkachenko also highlighted Oxia as Pulsar's replacement for ZooKeeper, saying no one wants to run ZooKeeper in 2026.5 The project is careful on this point. ZooKeeper is still fully supported for regular topics, and a migration framework moves clusters to Oxia without downtime.3238

Ursa adds Kafka Connect, and the connector model shifts

Tkachenko found StreamNative's other keynote announcement more surprising: support for the Kafka Connect API in Ursa.5 Ursa is the company's diskless streaming engine. It uses object storage as its main data store, and its background compaction process already pushes data out to systems like OpenSearch and ClickHouse.5 Running Kafka Connect connectors this way changes the usual pipeline. Instead of source to Kafka to sink, data goes from source to object storage to sink, a change Tkachenko argued has a range of benefits.5

The announcement also reverses part of StreamNative's earlier pitch. When it launched Ursa For Kafka in April, the company said users would need no Kafka Connect, no materialization pipelines and no sink connectors, because Kafka topics would themselves be lakehouse tables.15 Adding Connect support does not abandon that idea. Instead, it accepts that companies have years of investment in connectors, and that not every destination is an Iceberg or Delta table. In our view, that is a sensible correction. Telling customers to drop their connectors was always more marketing than migration plan.

StreamNative has built this story in stages. In April it announced Lakestream, an architecture where stream storage writes straight to Iceberg or Delta Lake. A catalog layer connects to Databricks Unity Catalog, Snowflake Horizon Catalog and AWS S3 Tables, and stateless servers handle the Kafka and Pulsar protocols.20 It claims cost cuts of up to 95% by removing replication across availability zones, tested at 5 GB/s of sustained throughput. Those are vendor figures, but the architecture behind them won VLDB 2025's Best Industry Paper award.15 The company also promised to open-source Ursa and key Lakestream components. A September 23 blog post titled "Open-Sourcing Ursa & UFK" suggests that has started.1516

Streamhouse: rivals agree on a name

The summit also showed the new "Streamhouse" coalition on stage, including Ververica in the keynote. Tkachenko read Streamhouse as the industry's answer to how streaming should supply fresh, reliable, governed context to AI systems.5 StreamNative joined the Streamhouse Working Group in mid-September as a founding member, alongside Aiven, Confluent, Redpanda and Ververica. The group's goal is a vendor-neutral definition of an architecture for real-time applications, analytics and agents.41 Ververica coined the name and has made it free for anyone to use under a binding trademark commitment.41

The framing is meant to sound cooperative: the lakehouse analyzes the business, and the Streamhouse runs it, on the same object storage, table formats and catalogs.41 Still, the founding list stands out. It includes nearly every major independent streaming vendor, including Confluent, which IBM now owns after an acquisition that closed in December 2025.2141 Snowflake and Databricks are not on it. Our reading is that this is a defensive alliance presented as an open standard. Streaming specialists are trying to agree on a shared vocabulary before the lakehouse platforms take over the category.

One detail differs between accounts. Tkachenko wrote that StreamNative Cloud "now ships with" a RisingWave-powered SQL notebook.5 StreamNative's September post described the RisingWave-powered SQL Workspace as being in private preview.41 Buyers should treat it as an early feature, not a finished one.

The lakehouse giants have already moved

The urgency makes sense given what happened in June. At Snowflake Summit, the company announced Datastream, a managed streaming service that speaks the Kafka wire protocol. Topics land as governed Snowflake or Iceberg tables and inherit access controls, masking, lineage and Time Travel.27 Datastream is still headed for private preview.2628 Futurum's Brad Shimmin said the goal is clearly to absorb the streaming layer into the data platform and go after what Snowflake describes as a $128 billion market.28 Snowflake also made Apache Iceberg V3 generally available, with two-way read and write through Horizon Catalog, which is built on Apache Polaris.2224

Databricks took a different route with Zerobus Ingest. Producers push events straight into Delta tables through a serverless API, and a serving layer called Lakehouse RT promises query responses in under a second.21 Its Kafka-compatible APIs are in beta, while the rest of Zerobus is generally available.23 Release notes show Databricks continuing to expand streaming-table features, including generally available AUTO CDC flows in Databricks SQL and external Iceberg and Delta read access to pipeline streaming tables in public preview.29

Analysts mostly agree on the narrow point and differ on what it means. Kai Waehner argues that using the Kafka protocol to feed a lakehouse is a different job from running Kafka as a company's event backbone. He notes that "Kafka-compatible" often means a working on-ramp without full Kafka semantics underneath.23 Volt Active Data agrees that Kafka is not going away. It focuses on the commercial motive instead: every hop between producer and platform is a place a competitor can sit, so taking over ingestion also brings governance and billing inside the platform's walls, with lock-in risks for customers.21

The verdict

The two camps are converging on the same technical ground: open table formats on object storage, governed through catalogs. They differ on which side should own the protocol layer. Snowflake and Databricks want Kafka-compatible ingestion to end at their platforms. StreamNative and the Streamhouse group want streams to stay a separate, open layer that any engine can read.2141

On the evidence so far, the independent vendors still have the stronger technical case for multi-consumer, operational workloads. Pulsar 5.0's elastic, order-preserving topics are something no lakehouse ingestion API tries to offer.3223 The platforms, however, have the stronger distribution argument for the many teams that run Kafka only to feed one warehouse.21 The next test comes at Confluent's Current conference in San Francisco on November 4–5, where the Kafka ecosystem's largest vendor gets its own chance to answer.10