David Yaffe and Johnny Graettinger are extending Estuary's exactly-once design across millisecond streams and bulk loads as the rollout begins.
By RuntimeWire Staff · Published
Primary source: PR Newswire
Why it matters #
Estuary's founders are betting that data integrity will become a control layer for autonomous software, because speed only magnifies errors when agents act on partial updates.
David Yaffe and Johnny Graettinger, the founders of Estuary, are rolling out a Rust-based runtime designed to keep complete database transactions intact as data moves into AI agents, operational software and warehouses. Estuary said in an August 26th announcement that new pipelines use the runtime by default, while existing pipelines are moving over without customer intervention.
The release gives Yaffe and Graettinger a new buyer for a problem they have been working on since adtech's high-volume years: systems that need fresh data cannot afford to receive an incomplete version of it. Before founding Estuary in 2019, the pair worked together at Invite Media and later built Arbor, a people-based data marketplace acquired by LiveRamp in 2016. Their infrastructure had to process roughly 20 million advertising requests per second with millisecond response times, according to their account of Estuary's origins.
Graettinger responded at Arbor by building Gazette, an open-source streaming broker intended to remove the need for separate batch and streaming systems. That architecture became the foundation for Estuary. The Rust rewrite applies the same founding idea to a wider operating range, with autonomous software replacing human dashboard users as the most demanding consumer.
The transaction is the product
Estuary captures changes from databases and other sources, stores them in durable logs and materializes the resulting collections into destinations. The new runtime is supposed to preserve the boundary around related changes, so downstream software receives a complete transaction or receives nothing from that transaction.
That distinction matters when software acts on the data immediately. A dashboard can display a temporarily incomplete number and wait for the next refresh. An inventory agent, fraud system or automated workflow can make an irreversible decision before the missing records arrive.
Estuary says the new runtime keeps its transactional guarantees across work ranging from millisecond streams to multi-hour loads, and from kilobytes to hundreds of petabytes. Estuary also claims each worker can process as much as 200 GB per hour, with workers added as volume rises. Those performance figures are self-reported and are not independently benchmarked in the materials Estuary published.
The guarantee also depends on the destination. Estuary's materialization documentation explains that checkpoints and view updates can commit together for end-to-end exactly-once behavior when a destination supports transactions. A connector can also preserve that behavior through an idempotent apply operation. Direct writes into a non-transactional destination may provide at-least-once semantics instead.
That qualification narrows the broadest reading of Estuary's announcement. The runtime can maintain transaction boundaries through its own processing, while the final delivery guarantee still reflects what the receiving system and connector can support.
Together AI shows the initial use case
Estuary is using Together AI's inference analytics pipeline to show why complete streaming data has become a commercial issue. Together AI instrumented inference services to publish structured events into AWS, then used Estuary to capture those streams and materialize them into warehouse tables.
Julia Zhang, a senior software engineer on Together AI's inference platform, built the first iteration in roughly one week, according to Estuary's customer case study. The resulting data supports analysis of utilization, serving economics and service-level commitments, while retaining request history beyond the shorter window used by Together AI's operational monitoring tools.
The example concerns Estuary's broader platform rather than an independently measured deployment of the new runtime. It still illustrates the workload Yaffe and Graettinger are pursuing: high-volume machine activity that must arrive quickly enough for operational use and completely enough to trust.
A 2026 MIT Technology Review Insights study produced with Google Cloud found that organizations' AI systems could access an average of 45% of enterprise data. Fifty-five percent of surveyed executives said their existing data systems were preventing them from scaling agentic AI, according to the Google Cloud report summary. Estuary's release cites a 300-executive survey from the same research program in which 83% reported some agent use, while 10% described adoption as widespread.
Those figures explain the timing. Model access is becoming easier, while production deployments still depend on scattered databases, warehouses, software applications and event streams. The data-integration layer gets harder once an agent is permitted to act rather than draft an answer for a person to approve.
A founder bet funded for enterprise deployment
Yaffe's path through Boeing, Yahoo, Invite Media, Google, Arbor and LiveRamp gave him repeated exposure to systems where latency has direct economic value. Graettinger supplied the underlying streaming architecture through Gazette and now leads Estuary's infrastructure development as CTO. Estuary packages that history into a managed product spanning change data capture, streaming and batch movement.
Investors funded the enterprise push before this runtime release. Estuary raised a $17 million Series A on October 21st, 2025, led by M13, with participation from FirstMark and Operator Partners, according to Estuary's Series A announcement. Combined with FirstMark's earlier $7 million financing, Estuary has disclosed at least $24 million in funding. Estuary has not disclosed a valuation.
The new runtime enters a market where agent language has spread quickly across data infrastructure. RisingWave 3.0 added exactly-once delivery and positioned its streaming database around agent applications in July. Confluent is building streaming agents and real-time context products on Kafka and Flink. Airbyte, Fivetran, and Artie are pursuing overlapping workloads from different architectural starting points.
Estuary's pitch rests on one system carrying both the live stream and the historical load while preserving transaction boundaries. The Rust runtime makes that argument more concrete, provided Estuary can reproduce its claimed throughput and integrity across customer environments rather than controlled workloads.
For Yaffe and Graettinger, agents are the latest expression of the same infrastructure problem they encountered at Arbor. Software needs a current view of the world, and it needs to know that the view did not stop halfway through a transaction. Autonomous action raises the cost of getting either requirement wrong.