• Reference

Dynamic Schema

Schema design is crucial for Milvus data processing. Before inserting entities into a Milvus collection, clarify the schema design and ensure that all data entities inserted afterward match the schema. However, this limits Milvus collections, making them similar to tables in relational databases.

Dynamic schema enables users to insert entities with new fields into a Milvus collection without modifying the existing schema. This means that users can insert data without knowing the full schema of a collection and can include fields that are not yet defined.

Create collection with dynamic schema enabled

To create a collection using a dynamic data model, set enable_dynamic_field to True when defining the data model. Afterward, all undefined fields and their values in the data entities inserted afterward will be treated as pre-defined fields. We prefer to use the term "dynamic fields" to refer to these key-value pairs.

With these dynamic fields, you can ask Milvus to output them in search/query results and include them in search and query filter expressions just as they are already defined in the collection schema.

from pymilvus import connections, Collection, FieldSchema, CollectionSchema, DataType, utility

connections.connect(host='localhost', port='19530')

# 1. define fields
fields = [
    FieldSchema(name="id", dtype=DataType.INT64, is_primary=True, auto_id=True),
    FieldSchema(name="title", dtype=DataType.VARCHAR, max_length=512),
    FieldSchema(name="title_vector", dtype=DataType.FLOAT_VECTOR, dim=768)
# 2. enable dynamic schema in schema definition
schema = CollectionSchema(
        "The schema for a medium news collection", 

# 3. reference the schema in a collection
collection = Collection("medium_articles_with_dynamic", schema)

# 4. index the vector field and load the collection
index_params = {
    "index_type": "AUTOINDEX",
    "metric_type": "L2",
    "params": {}


# 5. load the collection

Insert dynamic data

Once the collection is created, you can start inserting data, including the dynamic data into the collection.

Prepare data

To demonstrate the use of dynamic schema, we have prepared a dataset from Kaggle containing the articles published on from Jan 2020 to August 2020.

In this section, we need to prepare the dataset as follows:

import json
import pandas as pd

# read the raw dataset and convert it to JSON
df = pd.read_csv('New_Medium_Data.csv')
df.to_json('New_Medium_Data.json', orient='records')

# convert the vector field values into real numbers
def m(row):
    row.update({'title_vector': list(map(float, row['title_vector'][1:-1].split(', ')))})
    return row

with open('New_Medium_Data.json') as f:
    data_rows = json.load(f)
    data_rows = map(m, data_rows)
    data_rows = list(data_rows)

Then we should use data_rows as a handler to demonstrate the use of a collection with dynamic schema enabled.

Insert data

You can insert this dataset into the collection we have just created.

# 6. insert data

# Call the flush API to make inserted data immediately available for search

print("Entity counts: ", collection.num_entities)

# Output
# Entity counts:  5979

Search with dynamic fields

If you have created a collection with dynamic field enabled and inserted data with dynamic fields into, index, and load the collection, you can use dynamic fields in the filter expression of a search or a query as follows:

# Use the vector field of the first entity as the query vector.
result =
    param={"metric_type": "L2", "params": {"nprobe": 10}},
    expr='claps > 30 and reading_time < 10',
    output_fields=["title", "reading_time", "claps"],

for hits in result:
    print("Matched IDs: ", hits.ids)
    print("Distance to the query vector: ", hits.distances)
    print("Matched articles: ")
    for hit in hits:
            "Title: ", 
            ", Reading time: ", 
            ", Claps", hit.entity.get("claps")

# Output:
# Matched IDs:  [442005795759615782, 442005795759615816, 442005795759613616]
# Distance to the query vector:  [0.36103832721710205, 0.3767401874065399, 0.4162980318069458]
# Matched articles: 
# Title:  The Hidden Side Effect of the Coronavirus , Reading time:  8 , Claps 83
# Title:  Why The Coronavirus Mortality Rate is Misleading , Reading time:  9 , Claps 2900
# Title:  Coronavirus shows what ethical Amazon could look like , Reading time:  4 , Claps 51

If the key of a dynamic field contains characters other than digits, letters, and underscores (e.g. plus signs, asterisks, or dollar signs), you need to include the key within $meta[] as shown in the following code snippet when using it in a boolean expression.

expr='$meta["#key"] in ["a", "b", "c"]'

What's next

Supported data types Boolean express rules JavaScript Object Notation (JSON)