You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
precompute_online: feature names are about 80% of every vector blob #7075
Every pre-computed vector repeats the service's full feature name list, and for a wide service that is most of the blob. For a 90-feature service on Feast 0.67.0:
per entity
500-entity read
vector blob
4773 B, of which 3774 B are names identical for every entity
2.4 MB
per-view HMGET of the same 90 features
n/a
550 KB
So the single-read design moves 4x the bytes of the per-view path it replaces. Redis time per 500-entity read went from 2.0 ms to 0.4 ms with the single read, but the client spends it back parsing names.
Proposal, in #7077: an opt-in precompute_omit_names on FeatureService. Vectors written for such a service carry values, per-view timestamps, and a fingerprint of the service's ordered names. The fast path validates a nameless vector by fingerprint and takes names from the service; vectors with names keep working. Same service: 834 B per entity, 414 KB per 500-entity read. Needs one field on FeatureServiceSpec and one on PrecomputedFeatureVector.
Activity