Weaviate ships query profiling: one flag pinpoints slow-query bottlenecks inline

victorialslocum · x · 2026-08-27

Weaviate has released Query Profiling. The author notes that "the query is slow" isn't a diagnosis — the delay could come from filter evaluation, the vector index, reading objects off disk, or keyword scoring; without knowing the actual cause, every fix is a guess.

Usage is lightweight: add one opt-in flag on a single query and the breakdown comes back inline in the response:

python

returnmetadata=MetadataQuery(queryprofile=True)

Output is organized by shard and only shows stages that actually ran. totaltook is the wall clock for that shard's search; everything else explains where the time went. E.g. totaltook: 48.2ms, objectstook 36.8ms, vectorsearchtook 8.4ms means filters and search are cheap and you're disk-bound hydrating objects — look at page cache, etc. The existing slow query logs were good at passively catching regressions across a fleet, while this new feature targets debugging one specific query.

Original post →

More from coding & agent

coding & agent channel →