13 October 2026
52 new posts across 9 database and data engineering sources — 8 worth your time.
Chasing that 32 ms tail latency in ClickHouse
An investigation into 32ms tail latency spikes in ClickHouse revealed that idle periods caused TCP congestion windows to reset. Small, unbatched metadata packets in the Native protocol were overflowing these windows, a problem eventually solved by batching the packets.
ClickHouse · 13 October 2026 · Read the original →
Alexander Ioffe: Why is my Whale Query 1000x Slower in Prod? Databases hate skew, each hates it differently
A comparison of how Postgres and SQL Server handle extreme data skew, where 10% of customers own 80% of data. While a single session setting can rescue Postgres performance, SQL Server requires a different access-path fix to avoid a 1,000x slowdown.
Planet PostgreSQL · 13 October 2026 · Read the original →
What gets TOASTed in Postgres
A deep dive into the TOAST mechanism in Postgres, demonstrating the exact 2,001-byte threshold that triggers LZ4 compression. The post explains how Postgres moves large values out of the main row to maintain performance and storage efficiency.
PlanetScale · 13 October 2026 · Read the original →
Christophe Pettus: All Your GUCs in a Row: array_nulls (Special Update)
A warning regarding the removal of the `array_nulls` parameter in PostgreSQL 20. Because `pg_dump` does not explicitly set this parameter, restoring backups into databases where it was previously disabled can lead to silent corruption of text arrays.
Planet PostgreSQL · 13 October 2026 · Read the original →
Igor Suhorukov: Vectorized Query Execution in PostgreSQL 19: Arrow, ORCA and Flight SQL in a PostgreSQL Extension
An analysis of porting Apache Cloudberry to PostgreSQL 19, identifying the row-based executor as the primary bottleneck. Even with MPP implementations, the overhead of passing every value through the function manager (fmgr) prevents Postgres from matching the speed of columnar engines.
Planet PostgreSQL · 13 October 2026 · Read the original →
A DuckDB Database with No Data in It
An architectural pattern using DuckDB where the database file contains no actual data, only view definitions. This allows the database to act as a lightweight metadata layer (only a few hundred kilobytes) pointing to massive Parquet datasets on object storage.
DuckDB · 13 October 2026 · Read the original →
How we migrated Clera’s 500 GB production database to ClickHouse Managed Postgres overnight
Details the overnight migration of a 500 GB production database with over 500 tables to a managed Postgres environment. The transition successfully reduced CPU utilization from a constant 100% to a stable 10-20% range.
ClickHouse · 13 October 2026 · Read the original →
CockroachDB CTO on Distributed Database Design
Peter Mattis, Cockroach Labs CTO, discussed B-trees, automatic sharding, and strong consistency via Raft in distributed databases. He highlighted the latency tradeoff with cross-region replication and advised against chatty read-write patterns, emphasizing batching reads in parallel.
Cockroach Labs · 13 October 2026 · Read the original →
Still optimizing, still migrating, still debating architectures. The work persists.
8 stories, every Tuesday
Published here every week. Follow by RSS to get it as it lands.