Skip to content
This repository was archived by the owner on Nov 21, 2026. It is now read-only.
This repository was archived by the owner on Nov 21, 2026. It is now read-only.

Benchmarks: run Milvus as Zilliz recommends (3.0.3 Compose standalone, bulk import, seal and index wait, recorded output shape) #458

Closed

Description

@tae898

Zilliz (Jael Gu, 2026-10-22) reviewed our Milvus setup note of 2026-10-19. Their points: test Milvus 3.0.3 (latest 3.0 release); double-check the image tag and running digest in the Compose YAML; keep Compose defaults and try Storage V3 (V2 is the default), recording the storage mode per result; the HNSW settings are a fine common starting point but equal ef is not equal recall, so report recall with latency; prefer bulk import over row-by-row inserts; make sure all segments are sealed and fully indexed before timing searches; record requested output fields and filters.

This changes the Milvus arms of the dense and sparse lanes: 3.0.3 on the vendor's Compose standalone with every image pinned by digest, Storage V2 as the headline and V3 as a one-switch sensitivity run, ingest by bulk import (Parquet through RemoteBulkWriter into Milvus's MinIO, one import job), an explicit seal, index and load wait before any timed search with each wait recorded, and the search shape (no output fields, no filter) recorded on every row. Existing Milvus rows are 3.0.1 single-container rows and are not edited; the arms re-run in the comparator pass after the 2026-10-27 subset.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions