<p><strong>TLDR:</strong> Coupang replaced native database sequences across 100+ services migrating to DynamoDB, hitting 50,000 sequences/second for under $1,000/month. A real-world RAG bug report makes a strong case for hybrid search over pure vector similarity, with recall jumping from 72% to 94% on a 10M-row corpus. NetBox shipped TurboBulk using PostgreSQL’s COPY protocol to hit 24,000 objects/second versus the 100-200/second you get through a standard ORM. Toasty, an async Rust ORM from the Tokio team, landed on <a href="http://crates.io" target="_blank" rel="noopener nofollow">crates.io</a> with SQL and DynamoDB support.</p>
<hr>
<h2>Coupang’s two-tier sequence replacement at 50K/sec</h2>
<p>When Coupang migrated from a relational database to DynamoDB, they had to replace native sequences for 100+ teams without breaking anything. The solution is a thick client SDK embedded in each service that handles roughly 99% of requests from local memory, backed by a server-side cache, with DynamoDB only seeing about 0.1% of actual requests via conditional writes. A sliding window algorithm handles refill timing so cache exhaustion never blocks user requests. The Orders team migrated 12 services in 3 weeks with zero downtime and fewer than 50 lines of code changed per service. The key insight worth stealing: most teams didn’t actually need gap-free or globally ordered sequences, which turned a hard distributed coordination problem into a straightforward caching design.</p>
<hr>
<h2>Hybrid search fixes the RAG stale-document problem</h2>
<p>A customer-support RAG agent returning outdated policy documents after a laptop return exposed something worth paying attention to: vector similarity alone cannot enforce recency, tenant isolation, or document scope. The fix is combining vector distance with SQL WHERE clauses, JOINs, and GROUP BY in a single query. On a 10M-row corpus, hybrid search pushed Recall@5 from 72% to 94% and dropped the stale document rate from 23% to under 1%, at a latency cost of only 15-30ms. The post also argues against the “vector sidecar” pattern of running a separate vector database alongside your primary relational store, pointing to sync pipelines and consistency windows as operational debt that native vector support eliminates.</p>
<hr>
<h2>NetBox TurboBulk: PostgreSQL COPY protocol at 24K objects/sec</h2>
<p>NetBox Labs shipped TurboBulk for their cloud and enterprise tiers, bypassing the ORM entirely and going straight to PostgreSQL’s native COPY protocol and bulk SQL statements. The standard REST API processes objects one at a time through the ORM at 100-200 objects/second. TurboBulk hits over 24,000 objects/second. In real-world testing against tens of millions of objects, a full export finished in about 10 minutes versus an estimated 23+ hours through the REST API, using under 500MB of memory. It supports JSONL and Parquet export, async writes with insert/upsert/delete, three validation modes, and dry-run mode.</p>
<hr>
<h2>Toasty async ORM for Rust lands on <a href="http://crates.io" target="_blank" rel="noopener nofollow">crates.io</a></h2>
<p>Toasty, built by the Tokio team, is an async ORM for Rust that supports both SQL and NoSQL backends, with DynamoDB as the first NoSQL target. It works as an application-level query engine that decouples your application schema from the database schema, which makes conditional updates and other patterns that are awkward in raw SQL easier to express. It’s a 0.x release with breaking changes expected, but the foundation is stable enough to build on. Worth watching if you’re running Rust services that need to talk to DynamoDB without hand-rolling conditional write logic.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>pg_textsearch v1.0:</strong> Tiger Data (Timescale) released a new PostgreSQL extension bringing BM25-ranked keyword search under a permissive Postgres license, positioning it as an alternative to ParadeDB’s AGPL-licensed offering.</li>
<li><strong>Estuary PostgreSQL CDC updates:</strong> New release adds read-only CDC mode (no watermarks table writes), standby replica capture for PostgreSQL 16+, IAM auth for AWS/GCP/Azure, and per-partition stream capture.</li>
<li><strong>IVF vs HNSW in Milvus:</strong> Benchmarks across 1M to 100M vectors show HNSW wins on recall and latency under 50M vectors; IVF_SQ8 and IVF_PQ are better for larger datasets or memory-constrained environments. A two-stage IVF_PQ → HNSW re-rank pattern is also covered.</li>
<li><strong>Inbox pattern with PostgreSQL:</strong> Solid implementation walkthrough using <code>ON CONFLICT DO NOTHING</code> for idempotent inserts and <code>FOR UPDATE SKIP LOCKED</code> for safe concurrent batch processing, with <code>UNNEST</code> to minimize round-trips on batch updates.</li>
<li><strong>AI workloads vs. legacy database architectures:</strong> Agentic query patterns (dozens of concurrent queries in rapid succession) are exposing mismatches in systems built for batch reporting, pushing teams toward Postgres + ClickHouse splits and open table formats like Iceberg.</li>
<li><strong>SageMaker Data Agent:</strong> AWS added interactive charting, Snowflake SQL analytics, and AI-driven materialized view recommendations to SageMaker Unified Studio notebooks via natural language prompts.</li>
<li><strong>Database selection framework:</strong> A practical breakdown of how Netflix (Cassandra, 3M writes/sec), Instagram (PostgreSQL with PgBouncer and read replicas), and Twitter (Redis fan-out-on-write) chose databases based on access patterns rather than popularity.</li>
</ul>


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

Databases Digest

A curated roundup of database engineering stories: Coupang replaced native DB sequences across 100+ services with a two-tier caching SDK hitting 50K sequences/second for under $1K/month on DynamoDB. A real-world RAG bug case shows hybrid search (vector + SQL filters) boosting Recall@5 from 72% to 94% on a 10M-row corpus. NetBox's TurboBulk uses PostgreSQL's COPY protocol to reach 24K objects/second versus 100-200/second through a standard ORM. Toasty, an async Rust ORM from the Tokio team, launched on crates.io with SQL and DynamoDB support. Also covered: pg_textsearch BM25 extension, Estuary PostgreSQL CDC updates, IVF vs HNSW benchmarks in Milvus, the inbox pattern with PostgreSQL, and a database selection framework drawing on Netflix, Instagram, and Twitter architectures.

Coupang kills DB sequences at scale, hybrid search fixes broken RAG pipelines