<p><strong>TLDR:</strong> PostgreSQL 18’s native uuidv7() function makes time-ordered UUIDs a one-line change, with a 50M-row benchmark showing bulk inserts 10x faster than v4. ClickHouse’s internal logging platform LogHouse crossed 431 PiB and 1.59 quadrillion rows across three clouds, with architectural details worth stealing. PostgreSQL 18 and 19 are also adding native temporal table support, closing a long-standing gap against SQL:2011. On the tooling side, SQL Server 2025 extends regex support to varchar(max) LOB inputs across all seven T-SQL regex functions.</p>
<hr>
<h2>PostgreSQL 18 ships native uuidv7(), 50M-row benchmark shows 10x faster bulk inserts than v4</h2>
<p>Random UUIDv4 primary keys scatter inserts across B-tree pages, causing splits, cache misses, and WAL amplification at scale. PostgreSQL 18’s uuidv7() puts the timestamp in the most significant bits, so inserts append at the right edge of the index like a sequential bigint. A published 50M-row benchmark shows bulk inserts completing in roughly 1.8 minutes versus 20 minutes with v4, indexes about 25% smaller, and range scans 3x faster. The one real tradeoff: v7 embeds a decodable millisecond timestamp, so don’t expose it as a public identifier when creation time is sensitive. Adoption is a one-line DEFAULT change for new tables. <a href="https://daily.dev/posts/OboTHMuJs" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>ClickHouse LogHouse scales to 431 PiB and 1.59 quadrillion rows across three clouds</h2>
<p>ClickHouse’s internal logging platform grew 23x, from 19 PiB to 431 PiB, across 30+ regions on AWS, GCP, and Azure. The architecture uses geosharding with isolated cells for write scalability, Async Inserts to avoid TOO_MANY_PARTS errors, and a three-level Distributed table hierarchy that hides topology from query authors. A sharding key dictionary sourced from system.clusters enables optimize_skip_unused_shards to prune irrelevant cells, keeping region-filtered queries under 300ms even cross-continent. Peak ingestion hits 80 GiB/s and 190 million rows/second across 36 cells. <a href="https://daily.dev/posts/aIoh2yc6E" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>PostgreSQL 18 and 19 add native temporal table support, closing SQL:2011 gaps</h2>
<p>PostgreSQL 18 introduced temporal primary and unique keys via WITHOUT OVERLAPS and temporal foreign keys via the PERIOD clause, enforced through GiST indexes. PostgreSQL 19 Beta 1 adds UPDATE/DELETE … FOR PORTION OF, which lets you surgically modify application-time history while automatically preserving unaffected time periods. This covers the booking and scheduling use cases that previously required application-layer workarounds. System-versioned tables and FOR SYSTEM_TIME AS OF are still missing, so full SQL:2011 compliance isn’t there yet. <a href="https://daily.dev/posts/Ypp4BfCv5" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>SQL Server 2025 CU5 extends regex to varchar(max) LOB inputs across all seven T-SQL functions</h2>
<p>Previously only three scalar regex functions accepted LOB inputs; the other four required truncating to 8,000 bytes or chunking in application code. SQL Server 2025 CU5 and Azure SQL now support varchar(max) and nvarchar(max) inputs up to 2 MB across all seven functions: REGEXP_LIKE, REGEXP_COUNT, REGEXP_INSTR, REGEXP_REPLACE, REGEXP_SUBSTR, REGEXP_MATCHES, and REGEXP_SPLIT_TO_TABLE. The engine uses RE2 (linear-time, no backtracking), which matters for large inputs. Exceeding the 2 MB limit raises an error rather than silently truncating. <a href="https://daily.dev/posts/qZd0Xt4vf" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>PostgreSQL GUC deep dive: enable_material vs enable_memoize:</strong> Memoize (added in PostgreSQL 14) is a keyed in-memory cache for nested-loop joins with low-cardinality outer sides — it uses work_mem × hash_mem_multiplier as its budget and never spills to disk, unlike Materialize which buffers an entire child plan’s output and can spill. EXPLAIN (ANALYZE) exposes Hits, Misses, Evictions, and Overflows to diagnose whether either node is helping or hurting; the real fix for misbehaving plans is almost always better statistics, not disabling the GUC permanently. <a href="https://daily.dev/posts/xwoJI4kKb" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>NVIDIA GQE achieves 7.5x aggregate speedup over DuckDB on TPC-H SF1000:</strong> NVIDIA’s open-source GPU Query Engine hits 7.5x aggregate speedup over DuckDB on CPU at TPC-H SF1000, with up to 25.5x on individual queries, using cuDF, nvCOMP, and pipelined CUDA streams; partition pruning alone skips 31% of data for a 1.43x end-to-end speedup. <a href="https://daily.dev/posts/mrYGJ3PrG" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>PostgreSQL: too many tables bloat CacheMemoryContext to 221 MB per connection:</strong> A test with 20,000 tables shows per-backend CacheMemoryContext growing from ~4 MB to 221 MB after querying all tables, with no size limit — mitigations include idle_session_timeout to recycle connections and keeping object counts moderate. <a href="https://daily.dev/posts/vPPcBovr4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Tinybird lightweight deletes now in beta, completing in milliseconds via ClickHouse hidden row mask:</strong> The new POST /v1/datasources/{name}/delete endpoint uses ClickHouse’s native lightweight DELETE (a hidden _row_exists mask) instead of rewriting disk parts, dropping delete time from minutes or hours to milliseconds to seconds — useful for GDPR erasure and bad-ingest cleanup. <a href="https://daily.dev/posts/QbljNl0mZ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Databricks LTAP architecture eliminates CDC pipelines by transcoding Postgres WAL to Parquet at the PageServer:</strong> Databricks co-founder Reynold Xin describes LTAP, which transcodes Postgres row data into Parquet/Delta/Iceberg at the PageServer materialization step so both Postgres and Lakehouse engines read a single copy of data — analytical queries fetch the current LSN from Postgres and merge un-materialized recent changes, removing ETL lag and data drift without a separate CDC pipeline. <a href="https://daily.dev/posts/gSuP5T8Nr" 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 18's native uuidv7() function makes time-ordered UUIDs a one-line change, with a 50M-row benchmark showing bulk inserts 10x faster than v4. ClickHouse's internal logging platform LogHouse crossed 431 PiB and 1.59 quadrillion rows across three clouds, with architectural details worth stealing. PostgreSQL 18 and 19 are also adding native temporal table support, closing a long-standing gap against SQL:2011. On the tooling side, SQL Server 2025 extends regex support to varchar(max) LOB inputs across all seven T-SQL regex functions.

PostgreSQL 18 uuidv7 lands, ClickHouse LogHouse hits 431 PiB