<p><strong>TLDR:</strong> Databricks Lakebase now offers Git-style database branching for Postgres using copy-on-write, so you can spin up isolated environments in seconds without full data copies. ClickHouse released clickhousectl in beta, a version-manager-style CLI for both local and cloud that’s explicitly designed to work with AI coding agents. Databricks Research also published work on memory scaling for AI agents, showing accuracy jumping from near zero to 70%+ as agents accumulate external memory. On the data reliability front, a post on silent pipeline failures puts the average schema drift incident at $35,000 — a number worth keeping in mind if your validation story is thin.</p>
<hr>
<h2>Databricks Lakebase: Postgres database branching</h2>
<p>Lakebase now supports creating isolated Postgres branches in under a minute, regardless of database size, using a copy-on-write architecture with compute and storage separated. The obvious use cases are per-developer environments, per-PR isolation, and CI test runs — all without paying for full data copies. Non-production branches auto-scale to zero when idle, which makes it practical to create and throw them away freely. This is the kind of workflow that’s been standard in application code for years; it’s good to see it land at the database layer.</p>
<h2>ClickHouse releases clickhousectl CLI (beta)</h2>
<p>ClickHouse’s new official CLI handles both local instance management and ClickHouse Cloud — service creation, scaling, backups, API keys, and team management. It’s modeled after uv and pnpm as a version manager, supports project scaffolding, and outputs JSON for scripting. The explicit mention of AI agent support (including installing agent skills into coding assistants) is notable, though the docs flag a reasonable caution about granting agents destructive permissions. Worth evaluating if you’re managing ClickHouse deployments and tired of stitching together curl calls.</p>
<h2>Databricks Research: memory scaling for AI agents</h2>
<p>The MemAlign experiments on Databricks Genie Spaces show a clear pattern: agent accuracy rises from near 0% to over 70% and reasoning steps drop from around 20 to around 5 as the agent accumulates more external memory. The infrastructure stack they describe leans on serverless Postgres with vector search for storage, plus memory management pipelines covering bootstrapping, distillation, and consolidation. The governance angle — identity-aware ACLs, data lineage, auditability — is where this gets interesting for teams that actually have to ship this in production. The core idea, that the LLM becomes a swappable reasoning engine while the memory store carries the domain knowledge, is a meaningful architectural shift.</p>
<h2>Apache Iceberg architecture breakdown</h2>
<p>A solid deep-dive into Iceberg’s three-layer structure: the catalog layer as a thin atomic pointer, the metadata layer handling hierarchical query pruning via manifest lists and manifest files, and the data layer as open-format Parquet in object storage. The Copy-on-Write vs Merge-on-Read trade-off for deletes gets specific treatment, which is the decision that actually bites you at scale. Catalog options covered include Snowflake Polaris, AWS Glue, Nessie, and Hive Metastore. If you’re evaluating lakehouse formats or onboarding engineers to an existing Iceberg setup, this is a reasonable reference.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Snowflake + Salesforce architecture:</strong> A detailed walkthrough of using Snowflake as the intelligence backend for CRM with Salesforce as the UI layer — cursor-based pagination to stay within governor limits, a three-phase progressive loading strategy that cuts time-to-interaction by 98%, and multi-tier caching dropping repeat load times from 6.5s to under 2s.</li>
<li><strong>Visual EXPLAIN in Tabularis:</strong> An open-source database client adding interactive execution plan graphs for PostgreSQL, MySQL, MariaDB, and SQLite, with AI-powered analysis via OpenAI, Anthropic, or Ollama; still on a feature branch but plan comparison is on the roadmap.</li>
<li><strong>Silent pipeline failures:</strong> Schema drift incidents average $35,000 each and poor data quality costs organizations $12.9–15M annually; the post covers JSON Schema validation, schema registries, and consumer-driven contract testing with Pact as the layered defense.</li>
<li><strong>Drizzle vs Prisma 2026:</strong> Prisma 7’s TypeScript rewrite dropped bundle size from ~14MB to ~1.6MB with a 9x cold start improvement; the practical split is Drizzle for edge runtimes and SQL-fluent teams, Prisma for standard relational CRUD and teams less comfortable writing raw SQL.</li>
<li><strong>Google Cloud Postgres contributions:</strong> Google summarized H2 2025 Postgres work including automatic conflict detection for active-active replication, logical replication extended to sequences, and pg_upgrade improvements that reduce upgrade times for large deployments.</li>
<li><strong>Redis/Valkey security coordination:</strong> A screenshot from the Redis core team chat surfaced suggesting a security issue flagged by AI tooling, with a mention of coordinating a release with Valkey — details are thin but worth watching.</li>
<li><strong>dotConnect for PostgreSQL 9.1:</strong> Devart’s .NET data provider now supports Entity Framework Core 10 and pgvector VECTOR and HALFVEC types.</li>
<li><strong>Real-time data products pattern:</strong> A post using Chipotle’s kitchen model as an analogy argues for continuously pre-transforming operational data into reusable domain objects; Materialize is the SQL platform presented as implementing this.</li>
<li><strong>Orange ORM:</strong> An open-source Node.js/TypeScript ORM supporting PostgreSQL, MySQL, MariaDB, SQLite, MS SQL, Oracle, SAP ASE, PGlite, and Cloudflare D1, with no code generation required and Row Level Security integration for Postgres.</li>
</ul>


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

Databases Digest

A roundup of database and data engineering news: Databricks Lakebase adds Git-style Postgres branching using copy-on-write for instant isolated environments; ClickHouse releases clickhousectl beta CLI with AI agent support; Databricks Research shows AI agent accuracy jumping from near 0% to 70%+ with external memory scaling; Apache Iceberg's three-layer architecture is broken down with CoW vs MoR trade-offs; plus notes on Drizzle vs Prisma 2026, silent pipeline failures costing $35K per schema drift incident, Google Cloud Postgres contributions, and several ORM and database tooling updates.

Databricks Lakebase ships Postgres branching, ClickHouse drops an official CLI