<p><strong>TLDR:</strong> PostgreSQL 19 is getting online checksum management, which is a long-overdue fix for clusters that were initialized without checksums and couldn’t enable them without a full reload. A sponsored benchmark puts Supermetal’s Iceberg sink at 13 minutes for a TPC-H SF=50 snapshot versus 90+ minutes for Flink and 3+ hours for Spark. PyMongo 4.13 ships a stable native async API, making Motor redundant. ClickHouse’s index pruning post is worth reading if you’re doing analytical queries on large datasets.</p>
<hr>
<h2>PostgreSQL 19: online data checksum enable/disable</h2>
<p>PostgreSQL 19 adds <code>pg_enable_data_checksums()</code> and <code>pg_disable_data_checksums()</code>, letting you toggle checksums on a live cluster without taking it offline or running <code>pg_checksums</code> in maintenance mode. A background worker marks buffers dirty so checksums get written on the next page flush, transitioning through an <code>inprogress-on</code> state. The function accepts <code>cost_delay</code> and <code>cost_limit</code> arguments to throttle the IO load, mirroring vacuum cost parameters. If you’re running a long-lived cluster that was initialized without checksums, this removes the last real blocker.</p>
<h2>Postgres-to-Iceberg: Supermetal vs. Flink, Kafka Connect, Spark</h2>
<p>This is a sponsored benchmark, so take it with appropriate skepticism, but the numbers are interesting. On identical single-node AWS hardware with TPC-H SF=50 data, Supermetal completed the snapshot phase in 13 minutes with no tuning. Flink needed aggressive fetch and split size tuning to land at 90-116 minutes. Kafka Connect came in at 120 minutes. Spark took over 3 hours, which the post attributes to its scale-out architecture struggling on a single node. The differentiator is Supermetal’s ability to switch Iceberg sink behavior between snapshot and CDC phases. All tools produced correct row counts.</p>
<h2>PyMongo 4.13 ships native async, Motor is now optional</h2>
<p>PyMongo 4.13 includes a stable <code>pymongo.async_</code> API, which means you no longer need Motor for async MongoDB access in Python. A Locust benchmark under 500 concurrent users shows async PyMongo at roughly 113 RPS and 2,800ms median latency versus sync PyMongo at 49 RPS and 8,700ms. If you’re running FastAPI with Motor, the migration path is straightforward and the dependency reduction is worth it.</p>
<h2>ClickHouse index pruning: three techniques on 243M rows</h2>
<p>A solid writeup covering ClickHouse’s primary index (granule-based binary search), lightweight projections introduced in 25.6 (sort key plus pointer, no full data duplication, combinable for intersection pruning), and skip indexes like minmax. All three are demonstrated on a 243 million row UK property sales dataset with actual query timings and EXPLAIN output. The lightweight projections section is the most useful part — they’re a cleaner alternative to full materialized projections when you just need efficient filtering on non-primary-key columns.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Google Cloud PostgreSQL contributions:</strong> Google highlighted upstream work on logical replication stability and upgrade processes, developed in collaboration with the PostgreSQL community.</li>
<li><strong>Do you even need a database?</strong> A Go/Rust benchmark shows in-memory hash maps hitting 97-169k req/s versus SQLite at 25k req/s, with the argument that SQLite on a single server can handle roughly 90 million DAU before you need something more.</li>
<li><strong>Laravel 13.5.0 Redis Cluster support:</strong> Fixes CROSSSLOT errors on AWS ElastiCache Serverless by automatically wrapping queue names in hash tags — a real pain point for anyone running Laravel queues on clustered Redis.</li>
<li><strong>TypeORM connection pooling:</strong> A reminder that failing to release transactions in TypeORM starves the pool; the post compares behavior against Prisma and Sequelize.</li>
<li><strong>API key design for sharded databases:</strong> A developer walks through routing requests to the correct shard without session cookies, landing on SHAKE256 with a 10-character base64url prefix for index efficiency and collision resistance.</li>
<li><strong>MySQL CDC guide:</strong> Covers trigger-based, timestamp, and binlog approaches, with binlog (ROW format) recommended for production; compares Debezium, Airbyte, Fivetran, AWS DMS, and DBConvert Streams.</li>
<li><strong>PlanetScale Postgres imports via pgcopydb:</strong> Uses PostgreSQL’s native COPY protocol with parallel table copies and CDC-based near-zero-downtime cutover; validated on multi-terabyte databases at up to 2 TB/hour.</li>
<li><strong>Azure Blob Storage smart tier GA:</strong> Automated tiering across hot, cool, and cold based on access patterns, with no retrieval or early deletion charges — just a per-object monitoring fee; objects under 128 KiB are excluded.</li>
<li><strong>InfluxDB 3 Explorer 1.7:</strong> Adds table management UI, guided migration from v1/v2/v3, and a visual transform interface for ingestion-time downsampling and column renaming.</li>
<li><strong>Context layer vs. semantic layer:</strong> A 522-query study found 38% improvement in SQL accuracy when AI agents were grounded in context layer metadata; the post argues semantic layer comes first, context layer on top, then expose both via MCP.</li>
<li><strong>Weaviate Shared Cloud GA on AWS:</strong> Now available in US East and EU Frankfurt alongside existing GCP support, with RBAC, immutable backups, and SOC 2/ISO 27001 compliance.</li>
</ul>


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

Databases Digest

A developer news roundup covering several database and backend topics. PostgreSQL 19 introduces online checksum enable/disable via background workers, removing the need to take clusters offline. A sponsored benchmark compares Postgres-to-Iceberg pipelines: Supermetal completes a TPC-H SF=50 snapshot in 13 minutes versus 90+ minutes for Flink and 3+ hours for Spark. PyMongo 4.13 ships a stable native async API, making the Motor driver optional, with significant RPS and latency improvements shown under load. A ClickHouse deep-dive covers primary index granule search, lightweight projections (new in 25.6), and skip indexes on a 243M row dataset. Also noted: Laravel Redis Cluster fix, TypeORM connection pool pitfalls, MySQL CDC approaches, PlanetScale Postgres imports via pgcopydb, and more.

PostgreSQL 19 enables checksums online, Postgres-to-Iceberg benchmark favors Supermetal