The Rows & Columns Summit ran on September 22, 2026 at the Contemporary Jewish Museum in San Francisco, a single track from ten in the morning to half past five. It describes itself as a practitioner-first conference on the architecture question that won't go away: OLTP and OLAP, together or apart. It is organised by an independent board of three and backed by ClickHouse, with a blind CFP review.
Andy Pavlo, Head of ClickHouse Labs, opened the day. After lunch Hannes Mühleisen, co-founder and creator of DuckDB, gave the talk whose title is the thesis of the whole conference: Nobody Knows What OLTP Is: DuckDB Moves to the Middle. Between them a panel on convergence put Nikita Shamgunov of Databricks, Sailesh Krishnamurthy of Google and Yury Izrailevsky of ClickHouse on the same stage, and Fatma Ozcan of Google closed.
Russell Spitzer, Principal Engineer and Apache Iceberg PMC member at Snowflake, made the interoperability argument concrete. He started where he came from: the DataStax years building Cassandra's analytics integrations, including Apache Pig, and the way they used to move data between a transactional store and an analytical one, which was to issue a select star broken into many bounded queries. That is still how the Spark JDBC driver works. It has no atomicity, because records arrive from different moments of the scan, and no transactional semantics, so a job that fails halfway through a batch of upserts cannot be wound back. Iceberg is the answer to that specific problem, not a general-purpose faster thing.
Bonnie Xu of OpenAI gave the talk with the numbers. Nearly the whole company uses their data platform: over 600 petabytes processed a day across roughly 70,000 datasets, with about 200,000 queries run in production daily, and growing fast. When ChatGPT launched the question was how many weekly active users there were. Now it is how many instant checkout users are on Chat Pro in Japan — the same shape of question, far more nuanced, and today it costs five Slack threads and two meetings. The hard part is not the query, it is table discovery at that scale, where similarly named tables hold different cohorts, team-specific views, and columns added two weeks ago. So they built an internal data agent to navigate it.
Alexy Khrabrov also recorded an interview on the floor with Arun Sharma, founder of LadybugDB and Ladybug Memory. LadybugDB continues Kuzu, the embedded graph database developed at the University of Waterloo whose team was acquired by Apple: a well-regarded codebase with published research behind it and no community, which Sharma has spent eleven months building one around. Before that, five years at Google and then Facebook, where he built a graph indexing system sitting beside the world's largest MySQL cluster, listening to its write-ahead log and built on RocksDB as its very first user, and in 2018 prototyped what today looks like Amazon DSQL.