<p><strong>TLDR:</strong> ClickHouse hit $250M ARR with a 3x year-over-year growth rate and launched Claude-powered agents at Open House 2026. Apache Iceberg 1.11.0 shipped with server-side scan planning, built-in table encryption, and Spark 4.1 support. Snowflake committed $6B to AWS over five years, leaning on Graviton for cost-efficient inference workloads. Databricks Lakebase added Always-On pricing and a Change Data Feed in public preview.</p>
<hr>
<h2>ClickHouse hits $250M ARR and ships multi-stage distributed query execution</h2>
<p>ClickHouse crossed $250M in annualized revenue, tripling year-over-year, with 4,000 customers including Anthropic, Meta, and Capital One. The company also shipped multi-stage distributed query execution in experimental preview on ClickHouse Cloud — the new model uses exchange operators (Shuffle, Broadcast, Gather, Scatter) to repartition intermediate data between stages, fixing the two worst bottlenecks in the old parallel replicas model: every node building a full right-side hash table for joins, and high-cardinality aggregations bottlenecked on a single coordinator. TPC-H SF100 benchmarks on 8 nodes show 7.4x scaling for scan-heavy queries and 2-4.5x speedups on multi-join queries. A cost-based optimizer is planned. They also dropped CostBench, an open benchmark claiming 23x better cost-performance than the nearest cloud data warehouse competitor — worth reading the methodology before taking that number at face value.</p>
<h2>Apache Iceberg 1.11.0 released</h2>
<p>The headline feature is server-side scan planning, which shifts metadata traversal from the query engine to the REST catalog. For large tables with deep snapshot histories, this meaningfully reduces the work query engines have to do before they can start reading data. Built-in table encryption arrives with envelope encryption and a three-tier key hierarchy, with Google KMS support out of the box. Spark 4.1 gets MERGE INTO schema evolution, Flink 2.1 gets the new DynamicIcebergSink for multi-table routing, and a new File Format API standardizes reads and writes across Parquet, Avro, and ORC. GCS users get footer prefetching and threaded VectoredIO via the Analytics Core library integration.</p>
<h2>Snowflake commits $6B to AWS, bets on Graviton for inference</h2>
<p>Snowflake signed a five-year, $6B AWS deal covering both Graviton ARM instances and GPU-accelerated EC2 for AI training and inference. The strategic logic is straightforward: Cortex AI inference workloads are CPU-bound enough that Graviton’s lower cost frees up budget for the GPU capacity Snowflake actually needs. The deal expands Snowflake’s AWS footprint to 10 new regions including sovereign cloud regions in Europe. Lifetime AWS Marketplace revenue has now crossed $7B. More details expected at Snowflake Summit, June 1-4 in San Francisco.</p>
<h2>Databricks Lakebase adds Always-On pricing and Change Data Feed</h2>
<p>Databricks introduced Always-On pricing for Lakebase: disable scale-to-zero, set a minimum compute unit baseline, and after 24 hours of continuous use you automatically get a 25% discount on that baseline with no long-term commitment required. An additional 50% promotional discount runs through January 2027. Separately, Lakebase Change Data Feed hit public preview — enabled at the project level in under a minute, the feed lands in Unity Catalog Managed Tables, positioning the operational database as a native Bronze layer in the Medallion Architecture without separate CDC connectors.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Redpanda SQL is GA:</strong> Postgres-compatible MPP query engine (built on Oxla, acquired 2025) runs inside existing BYOC clusters on AWS, querying live streaming topics and Iceberg tables in a single statement with no ETL pipeline required.</li>
<li><strong>Neon/Lakebase resilience post:</strong> Stateless Postgres compute backed by zone-redundant storage means no hot standbys; a May 2026 AWS us-east-1 outage affected only ~13% of databases due to cell-based regional isolation.</li>
<li><strong>Cache Aware Scheduling coming in Linux 7.2:</strong> Benchmarks on AMD Zen 5 show measurable gains for PostgreSQL and Valkey workloads by colocating tasks sharing data within the same LLC domain.</li>
<li><strong>Airtable’s Milvus architecture:</strong> One partition per customer base for isolation, hierarchical capping at 400 collections × 1,000 partitions per cluster, HNSW indexing, and hot/cold offloading to handle 75% idle bases at any given time.</li>
<li><strong>Valkey + Mem0 for agent memory:</strong> Full pipeline walkthrough covering FT.CREATE with HNSW, HSET with scoped TAG filters, and hybrid retrieval — claims up to 90% token cost reduction vs. full-context approaches.</li>
<li><strong>DoorDash offline LLM personalization:</strong> Batch LLM calls produce carousel definitions stored as embeddings in Milvus; no LLM in the request path. Blue/green collection swaps for zero-downtime refreshes. A/B results showed $0.47 increase in per-user spend over three weeks.</li>
<li><strong>Snowflake Workload Identity Federation:</strong> Snowflake is deprecating password-based auth for service users; WIF lets workloads authenticate via AWS IAM roles, GCP service accounts, or Azure managed identities with no long-lived secrets.</li>
<li><strong>Hibernate Search 8.4.0.Final:</strong> Adds Elasticsearch 9.4 and OpenSearch 3.6 support, Hibernate ORM 7.4.0.Final compatibility; StatelessSession with listener-triggered indexing now throws an explicit exception instead of producing confusing silent failures.</li>
<li><strong>TigerFS:</strong> Mounts PostgreSQL as a FUSE/NFS filesystem — every file is a row, every directory a table, every file operation an atomic transaction. Interesting for AI agent workflows that prefer filesystem interfaces.</li>
<li><strong>Azure DocumentDB blog launch:</strong> Microsoft’s MongoDB-compatible vCore service (rebranded from Cosmos DB for MongoDB at Ignite 2025) now has a dedicated blog covering migration guidance and open-source ecosystem updates.</li>
<li><strong>ULIDs for Rails:</strong> Timestamp-prefixed IDs avoid B-tree fragmentation from random UUIDs; four-step implementation using the <code>ulid</code> gem and an ActiveRecord Concern.</li>
<li><strong>PostgreSQL zero-downtime migrations:</strong> ACCESS EXCLUSIVE locks on ALTER TABLE can queue thousands of queries on a hot table; expand-and-contract patterns and online DDL are the practical alternatives.</li>
<li><strong>AI coding agents and Spark performance:</strong> Agents tend to reach for exact percentile calculations on terabyte-scale datasets without understanding the cost, and sometimes silently coerce column types to suppress errors while producing wrong results.</li>
</ul>


Query optimization, schema design, and data store coverage across PostgreSQL, MongoDB, Redis, and more.

Databases Digest

ClickHouse reached $250M ARR with 3x year-over-year growth and shipped multi-stage distributed query execution in experimental preview, showing 7.4x scaling on scan-heavy TPC-H benchmarks. Apache Iceberg 1.11.0 released with server-side scan planning, built-in envelope encryption with a three-tier key hierarchy, Spark 4.1 MERGE INTO schema evolution, and Flink 2.1 multi-table routing. Snowflake signed a $6B five-year AWS deal leveraging Graviton ARM instances for cost-efficient Cortex AI inference workloads. Databricks Lakebase added Always-On pricing with automatic 25% discounts after 24 hours of continuous use and launched Change Data Feed in public preview for native CDC into Unity Catalog. Additional coverage includes Redpanda SQL GA, Neon/Lakebase resilience architecture, Airtable's Milvus partitioning strategy, DoorDash offline LLM personalization with Milvus, PostgreSQL zero-downtime migration patterns, and more.

ClickHouse triples to $250M ARR, Iceberg 1.11 ships server-side scan planning