<p><strong>TLDR:</strong> Snowflake quietly shipped directory imports for UDFs and stored procedures, which removes a real papercut for anyone managing Python dependencies or ML models on a stage. Datadog published a detailed post-mortem on how they replaced a slow Postgres search workload with a CDC-to-Kafka pipeline, and it’s one of the more honest engineering writeups you’ll see. Google Cloud dropped a summary of their PostgreSQL upstream contributions for the second half of 2025, including active-active conflict detection and logical replication of sequences. ProxySQL is splitting into three release tracks — Stable, Innovative, and AI — which is a reasonable way to stop forcing production users to absorb experimental changes.</p>
<hr>
<h2>Snowflake directory imports for UDFs and stored procedures</h2>
<p>Snowflake now lets you point the IMPORTS clause at a folder instead of listing individual files. That means you can reference a stage directory with a trailing slash or pull directly from a SnowGit repo ref like <code>@repo/branches/main/src/app/</code>. For anyone who’s maintained a Python UDF with a handful of helper modules and a serialized model file, this is a genuine quality-of-life fix. It also makes CI/CD pipelines cleaner since you can deploy by pushing to a branch rather than re-uploading and re-listing files.</p>
<hr>
<h2>Datadog’s Postgres-to-Kafka CDC migration</h2>
<p>Datadog’s Metrics Summary page was hitting 7-second p90 latency because it was running search workloads against a transactional Postgres database — joins across 82K metrics and 817K configuration rows. The fix was CDC via Debezium streaming WAL changes into Kafka, with Avro serialization and a Schema Registry enforcing backward compatibility for schema evolution. They used Temporal to automate pipeline provisioning, which turned a one-off fix into a reusable platform now supporting Postgres-to-Postgres, Postgres-to-Iceberg, Cassandra, and cross-region Kafka. The honest admission that they accepted replication lag as a deliberate tradeoff for resilience is the kind of detail that makes this worth reading.</p>
<hr>
<h2>Google Cloud’s PostgreSQL upstream contributions</h2>
<p>Google Cloud published a summary of their Postgres contributions from July through December 2025. The headline items are automatic conflict detection for active-active replication and logical replication of sequences — both long-standing gaps for anyone running multi-primary setups. They also fixed a self-deadlock in DROP SUBSCRIPTION and optimized pg_upgrade for large objects and logical slots. Work in progress includes a conflict log table, parallel pg_dump for large tables, and pipe-based parallel export. Dilip Kumar was selected for the PGConf.dev 2026 Program Committee.</p>
<hr>
<h2>ProxySQL splits into three release tracks</h2>
<p>ProxySQL is introducing Stable, Innovative, and AI release tracks. Stable is for production reliability, Innovative gets new features faster, and AI is for AI-related capabilities. This is a sensible move — the alternative is either holding back features to protect conservative users or shipping experimental stuff into production environments. No specific release dates or feature lists were announced yet, but the direction is clear.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Microsoft Fabric Deltaflow:</strong> Eliminates manual CDC connector setup by transforming raw Debezium feeds into analytics-ready events with automatic schema evolution, currently in preview for Azure SQL, SQL Managed Instance, SQL VM, and PostgreSQL.</li>
<li><strong>ClickHouse BYOC on Google Cloud GA:</strong> Fully managed ClickHouse running in your own GCP account, with control plane at ClickHouse and data plane on GKE with GCS storage connected via Tailscale; includes SharedMergeTree, auto-scaling, and ArgoCD-managed upgrades.</li>
<li><strong>pg_ivm 1.14:</strong> Fixes two bugs in incremental view maintenance for outer join views — one affecting cross joins, one affecting NULL values in join-condition columns.</li>
<li><strong>Amazon RDS for Oracle cross-account snapshot sharing:</strong> Now works for instances with additional storage volumes (up to 256 TiB across four volumes), preserving the original storage layout when shared across accounts.</li>
<li><strong>Connection pool sizing talk:</strong> Applies Little’s Law and Kingman’s Law to pool sizing, lands on (CPU cores × 2) + spindle count as a max, and includes a banking case study where 640 configured connections replaced 200 actually needed — the 75-80% utilization ceiling before wait times go exponential is the number to remember.</li>
<li><strong>SQL assertions in Oracle:</strong> A conference talk covers Oracle’s implementation of SQL assertions from the 1992 standard — declarative multi-row, multi-table constraints enforced by the engine rather than application code, with deferrable validation and partial validation support, though aggregation and deep nesting are not yet supported.</li>
<li><strong>Supabase docs over SSH:</strong> <code>supabase.sh</code> exposes Supabase docs as a virtual markdown filesystem so AI coding agents can grep and browse without hallucinating stale APIs.</li>
<li><strong>RxDB 17.0.0:</strong> Adds Google Drive and OneDrive replication plugins for serverless cross-device sync, a WebMCP plugin for AI agent access, and an Expo Filesystem storage for React Native that outperforms SQLite.</li>
<li><strong>Chess in pure SQL:</strong> Uses conditional aggregation with MAX/CASE inside a CTE to pivot piece positions into an 8x8 grid; the same pattern works for calendars and heatmaps.</li>
<li><strong>HyperLogLog turns 12:</strong> The Redis HyperLogLog implementation was released April 1st, 2014 and widely assumed to be an April Fools’ joke — both because of the name and because it claimed to improve on a Google paper.</li>
<li><strong>ClickStack RBAC:</strong> HyperDX’s rebranded ClickStack now supports role-based access control with three permission levels across dashboards, saved searches, alerts, and notebooks, with row-level controls still requiring manual ClickHouse configuration.</li>
<li><strong>Appwrite and MongoDB partnership:</strong> Appwrite’s self-hosted edition now spins up a MongoDB Community Edition container as a managed database option, with MongoDB on Appwrite Cloud and a more idiomatic NoSQL API on the roadmap.</li>
</ul>


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

Databases Digest

A curated roundup of database and data engineering news. Snowflake now supports directory imports for UDFs and stored procedures, letting you point the IMPORTS clause at a folder or Git repo branch instead of listing individual files — a quality-of-life improvement for Python UDFs and CI/CD pipelines. Datadog published a detailed engineering writeup on replacing a slow Postgres search workload (7-second p90 latency) with a CDC-to-Kafka pipeline using Debezium, Avro, and Temporal, now reused as a platform for multiple replication targets. Google Cloud summarized their PostgreSQL upstream contributions for H2 2025, including active-active conflict detection and logical replication of sequences. ProxySQL is splitting into Stable, Innovative, and AI release tracks to separate production-safe changes from experimental features. Additional items cover Microsoft Fabric Deltaflow, ClickHouse BYOC on GCP, pg_ivm 1.14 bug fixes, connection pool sizing using Little's Law, RxDB 17.0.0, and more.

Snowflake adds directory imports, Datadog's CDC replication story is worth reading