Anya Sage

In a benchmark against Ignition’s existing SQL Historian on Postgres, the new Enterprise Historian Module for TimescaleDB ingested 10x more data, queried it 10x faster, and used 10x less storage. Full SQL access throughout. Travis Cox (@InductiveAuto) and Doug Pagnutti (@TigerDatabase) walk through how Ignition and TimescaleDB fit together, plus TimescaleDB for Ignition: unlimited tags, one price, optional HA cluster at no extra charge. Thursday Oct 22, 9am PDT / 4pm GMT. Register and you'll get the recording either way. tsdb.co/ignition-webinar-x

Storing vectors is easy. Retrieving the right evidence is harder. Introducing searchgres.js: BM25, semantic + hybrid search in Postgres, with hierarchy, metadata, regex + time filters. Filtering belongs inside retrieval. tsdb.co/searchgres-x

The HNSW build took 42 hours. We indexed 100M Cohere embeddings with pgvector. On a 128 GiB machine, the build couldn’t finish. On a 512 GiB machine (4x the cost), it took 42 hours and served 95% recall in 8 ms. Today we’re releasing pg_vectorsearch, an open-source Postgres extension with a new index called PRISM. On the 128 GiB machine, the one HNSW couldn’t build on, PRISM built in 1h 40m and serves 95% recall in 6 ms. The index is 25 GB to HNSW’s 180 GB. It holds at smaller scale too. At 1M vectors, PRISM serves 99% recall in 15 ms on 2 GiB of RAM. HNSW on twice the memory takes 392 ms. How it works: PRISM groups similar vectors into partitions and stores them as compressed codes of about 1 bit per dimension. A query finds the nearest partitions and scans them sequentially, which keeps the CPU prefetcher ahead and lets SIMD score many vectors at once. Graph hops are unpredictable, so the CPU spends most of its time waiting on memory. It builds directly on existing pgvector columns. No rewrite, no copy, and your pgvector queries run unchanged. It’s an alpha. Heavy churn needs a manual rebalance, deletes don’t reuse space yet, and filtered streaming search is on the GA roadmap. Requires Postgres 18+. Run it on your data and tell us where it falls short: tsdb.co/ugw9oocq

A remote inverter drops off for 3 days, then flushes every buffered reading with its original timestamp. Those rows land in aggregate buckets materialized days ago. A 12-hour lookback never sees them. 5 refresh policy shapes for solar and wind SCADA: tsdb.co/cagg-refresh-x