pypi / pymilvus

pymilvus

What changed in pymilvus 2 to 3, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.

3 breaking1 security15 new APIs and options1 performance6 behavior and defaults11 fixes and other

See which of these apply to your code, ranked, in your agent:

npx @unstale/cli init

pymilvus 332 changes

Empty binary-vector lists now raise a parameter error instead of leaking an IndexError.
3.0.1release notes
Rejects oversized or non-positive target_size values.
2.6.17, 3.0.1release notes release notes
Adding a function field now requires bound index parameters.
3.0.1release notes

Uploads use safer per-upload options, lower default concurrency, byte-level progress, and request-local concurrent-upload state.
3.0.1release notes

Adds builders for Bloom filters and Roaring bitmaps for membership filtering.
3.0.2release notes
Adds APIs for observing the compaction lifecycle.
3.0.2release notes
Adds scalable upload policies for Volume.
3.0.2release notes
Adds a method for altering collection schemas.
3.0.1release notes
Adds role and user descriptions, including this role-alteration method.
3.0.1release notes
Adds TEXT handling across schema validation, result decoding, and BulkWriter paths.
3.0.1release notes
Adds a method for dropping function fields.
3.0.1release notes
Adds synchronous and asynchronous support for data-salvage workflows.
3.0.1release notes
Adds support for external field mappings and an add-field fallback for external collections.
3.0.1release notes
Adds FunctionChain APIs and forwards every specified stage during search.
3.0.1release notes
Adds this method to MilvusClient and AsyncMilvusClient.
3.0.1release notes
Adds support for MinHash function fields.
3.0.1release notes
Supports dynamically adding struct fields.
3.0.1release notes
Adds an expression helper for XGBoost.
3.0.1release notes

Array query results are materialized as standard Python lists.
2.6.17, 3.0.1release notes release notes
Null values are allowed in this column during insert.
2.6.17, 3.0.1release notes release notes
Role descriptions are now included in the method's output.
3.0.1release notes
Float inputs are now accepted for FP16 and BF16 vectors.
3.0.1release notes
Exposes error classification; only connection-setup RPC failures are mapped to CONNECT_FAILED.
3.0.1release notes
Query and Search Iterator V1 and V2 implementations move to the client layer, with ORM compatibility preserved and ORM deprecation warnings aligned.
3.0.1release notes

Configured connection timeouts are preserved during recovery.
3.0.1release notes
Sync and async reconnects swap channels atomically, protect active RPCs, and cover retired-channel reclamation.
3.0.1release notes
Adds search aggregation and correctly forwards the group_by_fields parameter.
3.0.1release notes
External filtering is preserved when creating search iterators.
3.0.1release notes
Booleans are rejected as sparse-vector values or indices, with improved validation errors.
3.0.1release notes
Supports partial updates and preserves struct-array fields reconstructed from schema dictionaries.
2.6.17, 3.0.1release notes release notes
BulkWriter supports struct-vector fields and skips payloads for null struct rows and null struct-vector subfields.
3.0.1release notes

Supports struct vector fields; remote upload path prefixes are preserved on Windows.
2.6.17release notes
Async schema results are unpacked before collection optimization runs.
2.6.17release notes
Primary-key, partition-key, and clustering-key metadata stays synchronized when fields are added to a schema.
2.6.17release notes
Boolean values used as sparse-vector indices are rejected.
2.6.17release notes