Feature-store demo: online serving + offline training, one copy, no skew
You are looking at a public feature tablefeature rows as 64-byte containers, durable on disk in one Semurg substrate. Ask it to serve features online (point reads) and scan them offline (training) from one copy and it runs the
real query live on this box, on data you can
download and re-run yourself ▸.
Click an example in the agent rail, or ask for the online read, the offline training scan, or both together. Every run is measured end to end, on
this box, on the real dataset. Nothing is pre-computed.
A feature row lives once as a 64-byte container, durable on disk. Online serving is a seqlock point read off the RAM-ring fanned across every core; offline training is a mass-scan fold over the same rows. No offline warehouse plus separate online cache, no materialization pipeline, no online/offline skew. Measured live, cold / warm / hot.
the feature-platform leader
TWO STORES THAT DRIFT APART. An offline warehouse for training and a separate online cache for serving, joined by a materialization pipeline that copies between them and an ever-present risk of online/offline skew.
semurg, one engine
One copy, skew impossible. The online seqlock read and the offline batch scan touch the identical bytes, so there is no warehouse plus cache pair, no materialization job copying between them, and no drift between training and serving features.
The second store, the copy pipeline and the skew it causes all collapse into one.
Feature rows folded online (serving) and offline (training) over the one copy. Measured live on this box. Re-run SERVE yourself.
ask the agent to SERVE, the live online (serving) + offline (training) numbers render here
Honest: a managed feature platform serves online point reads from a tuned in-memory store and answers offline training scans from a columnar warehouse, and each of those, on its own tuned path, is fast. We are not pitting our cold against their hot. The win here is COLLAPSE, not raw speed: we eliminate the second system (offline warehouse plus a separate online cache), the materialization job that copies offline features into the online store, and the online/offline skew that copy can introduce. The feature row lives once, durable on disk, and the online seqlock read and the offline batch scan touch the identical bytes, in the one engine that also serves graph, full-text, vectors and key-value.