<p><strong>TLDR:</strong> PostgreSQL 19 is adding a native REPACK command that replaces the VACUUM FULL/CLUSTER combo for reclaiming table bloat. Databricks just GA’d Real-Time Mode for Spark Structured Streaming, with early adopters reporting sub-100ms P99 latency. There’s also a solid writeup on the Outbox Pattern in Go and PostgreSQL worth reading if you’re building event-driven systems. SpacetimeDB is making noise with 1000x Postgres performance claims that deserve some scrutiny.</p>
<hr>
<h2>PostgreSQL 19 introduces REPACK</h2>
<p>PostgreSQL 19 is adding a built-in REPACK command that consolidates VACUUM FULL and CLUSTER into one cleaner operation. In the examples shown, an 8.5GB bloated table (after deleting 80% of rows) shrinks to ~1.6GB. It supports ordering by a specified index, includes a new <code>pg_stat_progress_repack</code> view, and accepts VERBOSE and ANALYZE options. The current implementation acquires the same locks as VACUUM FULL, but the design explicitly leaves room for a future CONCURRENTLY mode — which is the part that actually matters for production workloads.</p>
<h2>Databricks Real-Time Mode for Spark Structured Streaming is generally available</h2>
<p>Databricks GA’d Real-Time Mode (RTM) for Apache Spark Structured Streaming on Databricks Runtime 16.4+, enabled via a single config flag. RTM replaces microbatching with continuous data flow, and Databricks claims latency comparable to Flink in benchmarks. Coinbase, DraftKings, and MakeMyTrip are cited as early adopters reporting 80%+ latency reductions and sub-100ms P99s for fraud detection and ML feature pipelines. If the benchmark holds up in practice, this removes a real reason to run a separate Flink cluster alongside Spark.</p>
<h2>Outbox Pattern in Go and PostgreSQL</h2>
<p>A thorough implementation walkthrough covers the full Outbox Pattern using pgx and PostgreSQL: storing pending events in an outbox table within the same transaction as business data, then polling with a relay using <code>FOR UPDATE SKIP LOCKED</code> for concurrent-safe processing. The post covers at-least-once delivery semantics, idempotency strategies for consumers, and a higher-throughput alternative using PostgreSQL logical replication via the WAL. If you’ve been duct-taping this together yourself, it’s worth comparing against.</p>
<h2>SpacetimeDB’s 1000x Postgres claim</h2>
<p>SpacetimeDB is circulating benchmark results claiming over 1000x faster performance than PostgreSQL. The architectural explanation is straightforward: application code and database run on the same machine (no network round-trips), and queries run against in-memory data with disk persistence in the background — essentially Redis with a query engine. The 1000x number is real in the narrow scenario they’re measuring, but it’s comparing a co-located in-memory system against a networked disk-based one. Not a fair fight, and not a drop-in replacement.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Operational ontologies in Snowflake:</strong> A practical guide to storing hierarchical data (BOMs, supplier tiers) as node-edge tables and precomputing transitive relationships with closure tables built from recursive CTEs, turning multi-hop graph traversals into flat indexed SQL lookups that run under a second for 40,000+ component BOMs.</li>
<li><strong>Text-to-SQL agent analysis:</strong> A 50-question BIRD-Bench sample run through Claude Opus 4.5 with MotherDuck MCP found single-shot answers succeed 91% of the time while iterative loops succeed only 64% — and agents that struggle tend to fail completely rather than partially.</li>
<li><strong>DynamoDB architecture deep dive:</strong> Covers the full request path through five internal services (request router, auth, metadata service, global admission controller, storage nodes), quorum-based writes, WAL archival to S3, and the MEMSDS in-memory metadata cache that prevents thundering herd problems.</li>
<li><strong>KeeperDB launch:</strong> Keeper Security added vault-embedded database access to KeeperPAM, supporting MySQL, PostgreSQL, Oracle, and SQL Server with session recording and a proxy option for existing database clients. Official launch at RSA Conference 2026.</li>
<li><strong>Amazon RDS Custom for SQL Server OS updates:</strong> Customers can now view and schedule OS updates for RDS-provided engine versions via <code>describe-pending-maintenance-actions</code> and <code>apply-pending-maintenance-action</code> APIs, available across all supported regions.</li>
<li><strong>Snowflake graph reasoning without a graph database:</strong> Closure tables built from recursive CTEs handle multi-hop relationship queries (part_of, suppliedBy, locatedIn) in flat SQL, with Cortex Analyst sitting on top for natural language querying.</li>
<li><strong>Reactive streaming with Project Reactor:</strong> A step-by-step build of a non-blocking SSE endpoint using Spring WebFlux and Reactive MongoDB, with explicit backpressure via <code>onBackpressureLatest()</code> and a clear warning that calling <code>.block()</code> anywhere destroys the scalability model.</li>
<li><strong>SQL transformation discipline:</strong> A case for treating dbt/SQLMesh transformation layers like software — modular, version-controlled, tested, with structured modeling layers (raw, staging, intermediate, marts) and no business logic buried in BI tools.</li>
<li><strong>PGConf.dev 2026:</strong> Scheduled for May 19-22 in Vancouver, with community-led sessions added this year and events marking the 30th anniversary of the PostgreSQL project.</li>
</ul>


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

Databases Digest

PostgreSQL 19 is adding a native REPACK command that consolidates VACUUM FULL and CLUSTER into a single operation, shrinking bloated tables significantly, with a future CONCURRENTLY mode planned. Databricks has GA'd Real-Time Mode for Spark Structured Streaming on Runtime 16.4+, replacing microbatching with continuous flow and achieving sub-100ms P99 latency for early adopters like Coinbase and DraftKings. A detailed walkthrough covers implementing the Outbox Pattern in Go with PostgreSQL using pgx, FOR UPDATE SKIP LOCKED, and WAL-based logical replication. SpacetimeDB's 1000x Postgres performance claim is contextualized as a narrow benchmark comparing co-located in-memory architecture against a networked disk-based system. Additional items cover DynamoDB internals, text-to-SQL agent accuracy, Snowflake graph queries with recursive CTEs, and reactive streaming with Spring WebFlux.

PostgreSQL 19 gets a REPACK command, Databricks Structured Streaming hits millisecond latency