Recent
- News briefDT-2026-0202
OpenRouter's request count grew 33-fold in under two years
A ClickHouse case study puts OpenRouter at 19.7 billion requests a month and 110 trillion tokens a week, logged into one store.
2 minSource: ClickHouse
- News briefDT-2026-0201
Ordinary views leaked through their statistics
DuckDB's team lists changes that missed the 2.0.0 announcement, from a Python appender to views that hide what they filter out.
2 minSource: DuckDB
- BenchmarkDT-2026-0200
Eight fewer cores and 46 percent more throughput
DuckDB ran the full TPC-H workload at scale factor 1,000 on NVIDIA's Vera against an Intel Xeon 6 baseline.
2 minSource: DuckDB
- Field reportDT-2026-0199
Group on integers, join the strings later
DuckDB documents the star-schema trick for string-heavy aggregations and is honest about when it is not worth it.
2 minSource: DuckDB
- News briefDT-2026-0198
The adapter moved into the tool
dbt v2 is the first release shipping a DuckDB adapter built in, ending five years in which the most popular local setup was a community package.
2 minSource: DuckDB
- News briefDT-2026-0197
SQL table functions, written in Java
DuckDB's Java client can now register table functions in pure Java, so any JVM-reachable source becomes a SQL table without a C++ extension.
3 minSource: DuckDB
- News briefDT-2026-0196
Ten seconds to one millisecond
DuckDB v2.0 retires its PostgreSQL-derived parser for a PEG one. The headline number comes from a query that was never going to run anyway.
3 minSource: DuckDB
- News briefDT-2026-0195
Four JSON functions that move work back into SQL
DuckDB v2.0 adds RFC 7396 patching, canonical ordering and null stripping — with a measured 123× on one of them against Python.
2 minSource: DuckDB
- News briefDT-2026-0194
DuckDB 2.0 turns the embedded engine into a server
A preview lands ahead of the autumn release: network serving, triggers, a new parser and a storage format that breaks the old one.
3 minSource: DuckDB
- News briefDT-2026-0193
A third of ICML was reproduced in three weeks
A community effort ran 2,226 papers through agents and judged 35,908 individual claims. Roughly a quarter did not survive.
3 minSource: Hugging Face
- AnalysisDT-2026-0192
Chunking is still the whole game
Teams swap vector databases hoping for better recall. The gains almost always came from how the text was split.
9 min
- Market briefDT-2026-0187
Hybrid search became the default, quietly
Pure dense retrieval is now the exception in production systems.
5 min
- Field reportDT-2026-0181
Embedding drift is a migration problem, not a model problem
Changing the embedding model invalidates the index. Few teams plan for it.
7 min
- BenchmarkDT-2026-0176
Metadata filters cost more than expected at scale
Pre-filtering and post-filtering behave very differently once the corpus passes a hundred million vectors.
6 min
- AnalysisDT-2026-0170
Warehouses are absorbing vector search
For teams already running one, a separate vector store is getting harder to justify.
6 min