<p><strong>TLDR:</strong> PostgreSQL 19 is shaping up to be a significant release for I/O-heavy workloads: io_uring support lands as a new io_method option, and online checksum conversion via pg_enable_data_checksums() finally removes the need to take clusters offline. ClickHouse Managed Postgres published a concrete solution to PgBouncer’s single-threaded ceiling, hitting ~336k TPS on a 16-vCPU box by running a fleet of processes sharing a port via so_reuseport. Notion’s vector search team cut embedding costs by 90% and infrastructure costs by 50% through chunk hashing, metadata splitting, and a move to Ray for unified CPU/GPU processing. Mycli 2.0.0 ships with breaking changes and a handful of features worth knowing about if you use MySQL from the terminal.</p>
<hr>
<h2>PostgreSQL 19 adds io_uring as a native io_method</h2>
<p>PostgreSQL 19 introduces io_uring as an io_method option alongside the existing worker method. Instead of dedicated IO worker processes issuing synchronous pread64 calls, io_uring submits reads directly to the kernel’s submission queue and pulls completions asynchronously, keeping multiple reads in-flight simultaneously. The practical win is hiding I/O latency through prefetching on predictable access patterns like Sequential Scan and Bitmap Heap Scan. One monitoring gotcha: with io_uring and many parallel workers, load average can spike above 25 while CPUs sit nearly idle — your alerting thresholds may need revisiting. <a href="https://daily.dev/posts/OzZBu63EC" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>PostgreSQL 19 enables online data checksum conversion with pg_enable_data_checksums()</h2>
<p>Until now, enabling checksums on an existing cluster meant taking it fully offline and running pg_checksums — potentially hours or days of downtime for large databases. Postgres 19 adds pg_enable_data_checksums(), which spawns background workers per database, marks pages dirty for checksum writes on next flush, and WAL-logs the work so standbys convert in parallel. A new enum-based data_checksums parameter exposes four states (off, inprogress-on, on, inprogress-off) for monitoring progress. Checksums are also required by pg_rewind for HA failover, so this removes a real operational barrier for existing clusters. <a href="https://daily.dev/posts/Whx7zsHXw" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>ClickHouse Managed Postgres runs a PgBouncer fleet via so_reuseport, hits ~336k TPS on 16 vCPUs</h2>
<p>PgBouncer is single-threaded by design, which means a single process caps out at one core regardless of how many you have. ClickHouse’s approach: run one PgBouncer process per core, all bound to the same port via so_reuseport, letting the kernel load-balance incoming connections across the fleet. On a 16-vCPU AWS c7i.4xlarge, the fleet reaches ~336k TPS vs ~87k for a single process — roughly 4x — while using ~52% of available CPU vs ~9%. Query cancellation is handled through process peering, forwarding cancel requests to whichever process owns the session. The connection budget is split across processes to avoid oversubscribing Postgres. <a href="https://daily.dev/posts/dIq3cL9Xo" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Notion scaled vector search 10x at one-tenth the cost</h2>
<p>Notion’s vector infrastructure now handles 10x the original capacity at 10% of the original cost. The biggest levers: chunk hashing to skip re-embedding unchanged content on edits, splitting chunks into metadata and textual content so permission changes don’t trigger full re-embedding (70% data volume reduction), moving to a serverless architecture (50% cost reduction), and migrating to Ray for unified CPU/GPU processing on the same node (anticipated 90% embedding cost reduction). Kafka handles real-time updates while Apache Spark batch jobs manage chunking at scale across tens of thousands of workspaces daily. <a href="https://daily.dev/posts/lkqA2n49A" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Mycli 2.0.0 released with breaking changes:</strong> Mycli v2.0.0 ships fuzzy history search via fzf, an integrated /llm command, shell-like query-result redirection, system keyring password storage, checkpoint/resume for batch execution, and SSL modernization — check the changelog before upgrading if you have scripts depending on existing behavior. <a href="https://daily.dev/posts/RBJmGOuVk" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cloudflare’s internal lakehouse: billing workloads account for 53% of all queries:</strong> Cloudflare published details on Town Lake, its internal lakehouse built on Apache Trino, Apache Iceberg, and R2 object storage, which processed 91,760 billing-related queries from 324 employees in a measured period — 53% of total query volume — with a default-closed governance model requiring automated PII scanning before any new dataset gets access. <a href="https://daily.dev/posts/WKTGQiOUh" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Extralite 3.0.0 switches to WAL and enables foreign keys by default:</strong> Extralite 3.0.0 breaks backward compatibility by enabling WAL journaling, synchronous=NORMAL, and foreign key constraints as defaults, and blocks loading SQLite extensions via raw SQL — extensions must now go through Database#load_extension; a legacy mode exists for existing code. <a href="https://daily.dev/posts/o2uFbxTrC" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>enable_parallel_append: when a missing Parallel Append is a statistics problem, not a GUC problem:</strong> PostgreSQL’s enable_parallel_append controls Parallel Append nodes that spread workers across partitions or UNION ALL branches simultaneously, but a missing Parallel Append is often a costing or stale-statistics issue — use it as a diagnostic probe, fix root causes like outdated stats or parallel cost thresholds, then restore the default. <a href="https://daily.dev/posts/kJOWvXP32" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Prisma Compute vs Vercel: $98 vs $236 for 20M requests/month with 2 TB egress:</strong> At 20M requests/month with 2 TB egress and one seat, Prisma Compute (public beta) comes in at ~$98 vs Vercel Pro’s ~$236, with Prisma cheaper on CPU (half the price), memory (43% less), and egress (6x less) — though Prisma Compute has no established free tier and limited ecosystem maturity. <a href="https://daily.dev/posts/WZFjpQ6xq" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


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

Databases Digest

PostgreSQL 19 is shaping up to be a significant release for I/O-heavy workloads: io_uring support lands as a new io_method option, and online checksum conversion via pg_enable_data_checksums() finally removes the need to take clusters offline. ClickHouse Managed Postgres published a concrete solution to PgBouncer's single-threaded ceiling, hitting ~336k TPS on a 16-vCPU box by running a fleet of processes sharing a port via so_reuseport. Notion's vector search team cut embedding costs by 90% and infrastructure costs by 50% through chunk hashing, metadata splitting, and a move to Ray for unified CPU/GPU processing. Mycli 2.0.0 ships with breaking changes and a handful of features worth knowing about if you use MySQL from the terminal.

Postgres 19 gets io_uring and online checksum conversion, PgBouncer scales 4x with so_reuseport