<p><strong>TLDR:</strong> Cash App completed a migration of its two largest database systems to PlanetScale, covering 400TiB and 3-4 million QPS across 400 shards. Microsoft announced Azure HorizonDB, a new managed Postgres offering with compute-storage separation and some aggressive TPS numbers. DoorDash published a detailed writeup on replacing behavioral signals with LLM-generated embeddings, with production results worth reading. Databricks also had a busy day, shipping Agent Bricks GA and a native Lovable connector.</p>
<hr>
<h2>Cash App migrates 400TiB and 4M QPS to PlanetScale</h2>
<p>Cash App moved off self-hosted Vitess clusters onto PlanetScale’s managed platform. The systems run 3-4 million QPS across 400 shards with 400TiB of data — not a small lift. The main drivers were operational complexity from maintaining a forked Vitess version and the maintenance burden that came with it. Post-migration they’re reporting reduced overhead, better performance stability during peak loads, and improved query visibility through PlanetScale’s dashboard. This is a useful data point for anyone still running self-hosted Vitess at scale and wondering whether the managed path is worth it.</p>
<h2>Azure HorizonDB decouples compute and storage for Postgres</h2>
<p>Microsoft announced HorizonDB, a managed Postgres offering that separates compute and storage, scaling to 3,072 cores and 128TB. The headline numbers: over 9,000 TPS for OLTP workloads versus roughly 4,000 TPS with open-source Postgres and Patroni, and nearly 270,000 TPS read throughput across three replicas under a single load-balanced endpoint. There’s also a new DiskANN vector index that consolidates filtering into a single operation, claiming around 3x better performance than the prior version. Sub-millisecond multi-zone commit latency is the other claim worth stress-testing in practice.</p>
<h2>DoorDash: data quality beats model choice for embeddings</h2>
<p>DoorDash’s engineering team replaced sparse behavioral signals with LLM-generated merchant and item profiles to power content-first embeddings across food, grocery, retail, and gifting. The key finding is blunt: data quality from LLM profiles delivered +31% item similarity gains, while switching to a better encoder alone gave +6%. Production results include a 3.65% reduction in null search rate, +0.66% search CVR, +2.4% homepage order rate, and offline precision@10 going from 68% to 85%. They used gemini-embedding-001 with 256-dim MRL and an incremental Metaflow-based refresh pipeline. Worth reading if you’re building any kind of retrieval system — the lesson about where to invest effort is pretty clear.</p>
<h2>Databricks ships Agent Bricks GA and a Lovable connector</h2>
<p>Databricks announced general availability of Agent Bricks, its enterprise agent platform, claiming 70% higher accuracy than standard RAG by using Unity Catalog metadata for business context. The platform supports Cursor, Codex, Claude Code, LangGraph, and OpenAI Agents SDK through a single API. Separately, they launched a native connector for Lovable that lets business users build data-driven apps directly on Databricks data using plain English, with data staying inside the Databricks security perimeter and queried at runtime — no ETL pipeline, no replication.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Mintlify migrates analytics to ClickHouse Cloud:</strong> Moved from PostHog, cut dashboard load times from tens of seconds to under one second using incremental materialized views, reduced costs by 60%, and reports a 30% NPS improvement. Migration took two weeks using dual-write through Amazon MSK via ClickPipes.</li>
<li><strong>Trendyol builds self-service Postgres migration platform:</strong> Used <a href="http://Temporal.io" target="_blank" rel="noopener nofollow">Temporal.io</a> to replace manual DBA-driven migration scripts, handling WAL bloat, sequence sync, and safe cutover with a hybrid pg_dump + logical replication approach. Teams can now trigger migrations autonomously.</li>
<li><strong>OpenDuck open-sources MotherDuck’s architecture:</strong> Self-hostable DuckDB with differential storage, hybrid execution, and a DuckDB-native catalog via StorageExtension. Gateway is Rust, protocol is two gRPC RPCs returning Arrow IPC batches, MIT licensed.</li>
<li><strong>Tigris bucket forking for agent workloads:</strong> Copy-on-write forks let you run 50 isolated agents against a 50GB dataset while paying for one copy of the data. Combined with LanceDB and DuckDB reading Lance files directly over S3.</li>
<li><strong>Neo4j deprecates APOC batching:</strong> <code>apoc.periodic.iterate</code> is being superseded by native Cypher’s <code>CALL { … } IN TRANSACTIONS</code>, which adds proper memory tracking, EXPLAIN support, and an ON ERROR RETRY option added in Neo4j 2025.03. APOC Core is now maintenance-only.</li>
<li><strong>Tinybird expands to AWS Sydney:</strong> Second Asia-Pacific region after Hong Kong, with full feature parity including Kafka/S3 connectors and real-time materialized views.</li>
<li><strong>Aurora DSQL PHP connector released:</strong> Automates IAM token generation, SSL config, and connection pooling for PDO_PGSQL, with opt-in OCC retry and exponential backoff.</li>
<li><strong>dltHub toolkit eval:</strong> A controlled 12-run experiment showed a base Claude Code agent leaked credentials in 100% of runs and skipped sampling; the workbench-guided agent passed all behavioral checks at 58% higher cost per run ($2.21 vs $1.40).</li>
<li><strong>DuckDB internals course published:</strong> 15-week university curriculum from Tübingen covering memory management, ART indexing, vectorized execution, and query optimization — slides available on GitHub.</li>
</ul>


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

Databases Digest

A roundup of major database and data infrastructure news: Cash App completed a migration of 400TiB and 3-4M QPS across 400 shards from self-hosted Vitess to PlanetScale's managed platform. Microsoft announced Azure HorizonDB, a managed Postgres service with compute-storage separation claiming 9,000+ TPS for OLTP and sub-millisecond multi-zone commit latency. DoorDash published production results showing LLM-generated embeddings improved item similarity by 31% versus 6% from a better encoder alone. Databricks shipped Agent Bricks GA with Unity Catalog integration and a native Lovable connector. Additional items cover Mintlify's ClickHouse migration (60% cost reduction), Trendyol's self-service Postgres migration platform using Temporal.io, OpenDuck open-sourcing MotherDuck's architecture, Neo4j deprecating APOC batching, and a DuckDB internals university course.

Cash App moves 400TiB to PlanetScale, Azure HorizonDB enters the managed Postgres race