<p><strong>TLDR:</strong> DuckLake 1.0 hit production-ready status today, bringing SQL-backed lakehouse metadata and data inlining that benchmarks at 926x faster aggregations than Iceberg with Polaris. Google launched BigQuery Graph in preview, adding native ISO GQL graph analytics on top of existing relational tables. Amazon DocumentDB got in-place major version upgrades from 5.0 to 8.0, with up to 7x lower query latency. A detailed post-mortem on Uber’s LedgerStore makes the case that DynamoDB’s consumption pricing made the whole thing an $8M mistake.</p>
<hr>
<h2>DuckLake 1.0 reaches production-ready status</h2>
<p>DuckLake 1.0 is out, and the core architectural bet — store all lakehouse metadata in a SQL database instead of scattered JSON files on object storage — is now stable and versioned. The headline feature is data inlining: small writes go directly into the catalog database rather than creating Parquet files, which eliminates the small files problem that plagues streaming workloads on Iceberg and Delta Lake without requiring compaction jobs. Benchmarks against Iceberg with Apache Polaris show 926x faster aggregation queries and 105x faster ingestion. Multi-engine support covers DuckDB, DataFusion, Trino, and Spark, and MotherDuck has a managed preview available now. DuckDB 1.5.2 ships alongside it as the reference implementation.</p>
<h2>BigQuery Graph launches in preview</h2>
<p>Google launched BigQuery Graph in preview today, letting you model property graphs directly on top of existing BigQuery tables using ISO GQL — no data duplication, no separate graph database. You declare nodes and edges via DDL, then query with standard SQL or GQL, including multi-hop traversals and vector/full-text search integration. Federated queries against Spanner Graph are supported, which is the interesting part: operational graph data in Spanner, analytical graph workloads in BigQuery, shared schema. Curve is cited as a real-world customer claiming ~£9.1M in fraud savings.</p>
<h2>Amazon DocumentDB gets in-place upgrades from 5.0 to 8.0</h2>
<p>Amazon DocumentDB now supports in-place major version upgrades directly from 5.0 to 8.0 — no new cluster, no endpoint changes, no index rebuilds. Version 8.0 brings up to 7x lower query latency, up to 5x better storage compression, vector index builds up to 30x faster, and enhanced text search via text index v2. The upgrade is available via Console, SDK, or CLI at no additional cost in all regions where DocumentDB 8.0 is supported. If you’ve been putting off that upgrade because of the migration overhead, that excuse is gone.</p>
<h2>Uber’s LedgerStore: the $8M DynamoDB lesson</h2>
<p>A detailed breakdown argues that Uber’s LedgerStore — frequently cited as a DynamoDB success story — was actually abandoned within 2-3 years because 11 million daily trips generating hundreds of millions of writes made the consumption-based pricing untenable, with an estimated $8M bill. The migration to an internal DocStore-based system required building a CDC streaming framework called Flux on top of it, compounding the waste. The core argument is that engineers optimized for the conference talk, not the economics, and that cost modeling needs to be a first-class part of system design — not an afterthought after you’ve already presented at re:Invent.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Meilisearch Launch Week:</strong> Shipped self-serve sharding and replication, enterprise SSO/SAML with SCIM, a Search Performance Inspector that breaks down query latency by component, and two experimental features — Document Joins and Dynamic Search Rules for pinning results by query condition.</li>
<li><strong>SingleStore MRL embeddings:</strong> Benchmarks on a 10M-row dataset show 128-dim IVF_FLAT indexes reduce memory by 87% and increase throughput 6.6x while maintaining 96% Recall@5, only 1.8% below the full-vector baseline.</li>
<li><strong>Debezium + Oracle LogMiner lag fix:</strong> A team at Pepkor NexTech traced 95-minute LogMiner fetch times to a nightly truncate-and-load script generating 476 million redo operations on tables not even in the capture list; replacing it with Oracle External Tables and a MERGE-based PL/SQL procedure cut max query time to ~12 minutes.</li>
<li><strong>ClickHouse migration at ELO:</strong> Migrating from Elasticsearch to ClickHouse for 22 million daily transactions reduced storage from 12 TB to 2 TB and cut annual infrastructure costs 87%, with 5-9x query performance improvement and sub-2-second dashboard refresh latency.</li>
<li><strong><a href="http://monday.com" target="_blank" rel="noopener nofollow">monday.com</a> ID generation:</strong> Migrated board item ID generation from MySQL auto-increment to a DynamoDB-backed ID Generator Service using atomic range leasing and double-buffering; recommends UUIDv7 for anyone starting fresh.</li>
<li><strong>Redb 4.1:</strong> Claude authored the dynamic read/write cache partitioning implementation in this Rust embedded key-value store, resulting in up to 1.5x speedup on some benchmarks.</li>
<li><strong>Apache Polaris graduates to TLP:</strong> The open-source Iceberg catalog is now an Apache Top-Level Project and has adopted Quarkus as its runtime, with 10ms startup in native mode.</li>
<li><strong>MotherDuck Agent Skills:</strong> 17 SKILL.md-based instructions open-sourced for teaching coding agents the right DuckDB SQL dialect, connection patterns, and tenant isolation — installable via a single npx command.</li>
<li><strong>Mercedes-Benz Delta Sharing:</strong> Built a cross-cloud data mesh between AWS and Azure using Delta Sharing and Deep Clone, reducing egress costs 66% on initial data products and up to 93% annually, improving data freshness from weekly to every two days.</li>
<li><strong>Claude Code token reduction:</strong> Structured MCP context (skills, semantic exit codes, full topology in ~500 tokens) cut Claude Code token usage 2.8x compared to Supabase’s MCP server — 3.7M vs 10.4M tokens for the same app build.</li>
<li><strong>PostgreSQL 19:</strong> Tech Watch notes reduced EXPLAIN ANALYZE timing overhead in the upcoming release.</li>
<li><strong>Hibernate ORM 7.3:</strong> Adds a KeyType enum for loading entities by natural id and a new @NaturalIdClass annotation for composite natural ids, plus Jackson 3 support.</li>
<li><strong>ggsql alpha:</strong> Posit announced a grammar-of-graphics visualization tool using SQL syntax with DuckDB as the backend — no R or Python runtime required, designed for SQL-native users.</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: DuckLake 1.0 reaches production with SQL-backed lakehouse metadata and 926x faster aggregations than Iceberg with Polaris; Google launches BigQuery Graph in preview with native ISO GQL on existing tables; Amazon DocumentDB adds in-place upgrades from 5.0 to 8.0 with 7x lower latency; and a post-mortem argues Uber's LedgerStore was an $8M DynamoDB mistake due to consumption pricing at scale. Also covered: Meilisearch sharding, SingleStore embeddings benchmarks, ClickHouse migration wins, Apache Polaris graduating to TLP, and more.

DuckLake 1.0 ships, BigQuery Graph lands in preview