Skip to main content

DDL

MilvusQL's CREATE TABLE ... WITH (...) and CREATE INDEX ... USING <method> WITH (...) have no base-SQL equivalent, so they're exposed through SQLAlchemy's standard mechanism for dialect-specific DDL options — the same <dialect>_<option> prefixed-kwarg pattern mysql_engine=... uses.

Table options​

from sqlalchemy import Table, Column, BigInteger, String

items = Table(
    "items", metadata,
    Column("id", BigInteger, primary_key=True, autoincrement=True),
    Column("category", String(64)),
    milvusql_shards=1,
    milvusql_consistency_level="Bounded",
    milvusql_partition_key="category",
)

Renders WITH (shards=1, consistency_level='Bounded', partition_key='category') on CREATE TABLE.

A single-column primary key with autoincrement=True renders with an inline PRIMARY KEY AUTO_INCREMENT keyword on the column, in addition to the standard trailing PRIMARY KEY (id) constraint SQLAlchemy's base DDL compiler always emits — both are present in the output. This is unconditional, not configurable.

Index options​

from sqlalchemy import Index

Index(
    "idx_emb", items.c.embedding,
    milvusql_using="HNSW",
    milvusql_with={"metric_type": "COSINE", "M": 16, "ef_construction": 200},
)

Renders CREATE INDEX idx_emb ON items (embedding) USING HNSW WITH (metric_type='COSINE', M=16, ef_construction=200).

String values in WITH (...) are single-quoted — MilvusQL follows ANSI SQL here, where "..." means a quoted identifier and '...' means a string literal. (A double-quoted value in a property parses as exp.Var, not exp.Literal — confirmed directly against sqlglot-milvus.)

BM25 full-text: Computed() for the generated SPARSEVEC​

GENERATED ALWAYS AS (BM25(<text column>)) — the one server-side generated column Milvus supports — needs no dialect-specific construct: SQLAlchemy's own Computed() already renders it, since the base DDLCompiler appends a generated-column clause to any column carrying one:

from sqlalchemy import BigInteger, Column, Computed, Table, Text
from milvusql_sqlalchemy.types import SPARSEVEC, VECTOR

docs = Table(
    "docs", metadata,
    Column("id", BigInteger, primary_key=True, autoincrement=True),
    Column("content", Text),
    Column("content_sparse", SPARSEVEC(), Computed("BM25(content)")),
    Column("embedding", VECTOR(768)),
)
Index(
    "idx_fts", docs.c.content_sparse,
    milvusql_using="SPARSE_INVERTED_INDEX",
    milvusql_with={"metric_type": "BM25"},
)

A plain sa.Text column is Milvus's analyzer-enabled full-text input (TEXT in MilvusQL) — see MilvusQL Concepts → Full-text search for what TEXT/BM25_SCORE/MATCH ... AGAINST do underneath.

Loading a collection​

LOAD TABLE/RELEASE TABLE have no Table/Index-level equivalent to attach kwargs to — issue them directly:

with engine.begin() as conn:
    conn.exec_driver_sql("LOAD TABLE items")

Alembic​

Importing the dialect registers a milvusql-specific DefaultImpl with Alembic automatically, no separate setup call needed — this turns Alembic's otherwise-opaque KeyError for an unregistered dialect into a working engine.

Alembic migrations still cannot run end to end against Milvus

Alembic's own alembic_version bookkeeping table (version_num VARCHAR(32) PRIMARY KEY) has no vector column, and Milvus refuses to create a collection with zero vector fields. alembic upgrade, alembic downgrade and alembic revision --autogenerate therefore all fail with

NotSupportedError: CREATE TABLE 'alembic_version' has no VECTOR/SPARSEVEC column --
Milvus requires at least one vector field per collection.

raised client-side before any RPC. This is a Milvus limitation, surfaced as an explicit error rather than papered over with a hidden pad column. (milvusql-django takes the opposite approach for its own bookkeeping table — see Django → Schema & Migrations.)

Migrations run without a wrapping transaction — Milvus has no multi-statement rollback (the same reason Connection.rollback() and do_rollback() behave the way they do throughout this dialect), so Alembic's transactional_ddl is False here rather than promising a rollback this backend can't perform if a migration fails partway through.