A partition physically or logically divides a collection into subsets of vectors. When you target a specific partition in operations such as insert or query, DashVector scopes the operation to that partition only, reducing the search space and improving query performance.
Key concepts
| Concept | Description |
|---|---|
| Partition naming | Each partition has a unique name within its collection |
| Shared schema | All partitions in a collection share the same vector dimensions, vector data types, distance metrics, and field definitions |
| Default partition | Every collection includes a default partition that cannot be deleted. Operations without a specified partition target this default |
| Partition limit | A collection supports up to a fixed number of partitions. For the exact limit, see Limits |
| Management | Create and delete partitions through API operations |
When to use partitions
Partitions work best when your data has a clear categorical dimension that aligns with how you query it. Each query targets one partition, so the partitioning field should match the most common query filter.
Good candidates for partitioning:
| Use case | Partitioning field | Why it works |
|---|---|---|
| E-commerce image search | Product category (shoes, skirts, pants) | Users search within a known category, so each query hits one partition |
| Video surveillance | Date (one partition per day) | Data has a fixed 30-day retention window; create partitions daily and delete them when they expire |
| Trademark detection | Trademark structure (text, graphic, numeric, letter) | Queries target a specific trademark type |
| Multilingual Q&A | Language (Chinese, English, French) | Queries match the user's language |
| Multi-tenant SaaS | Customer ID | Provides physical data isolation between tenants at low cost |
When partitions are not the right fit:
Small datasets: The overhead of managing partitions outweighs the performance gain.
No clear partitioning field: If queries frequently span multiple partitions, performance is worse than using a single partition.
Prerequisites
Before you begin, make sure that you have:
A DashVector cluster. See Create a cluster
An API key. See Manage API keys
The latest DashVector SDK. See Install DashVector SDK
Manage partitions
All examples on this page use the following client and collection setup:
import dashvector
import os
# Get credentials from environment variables
client = dashvector.Client(
api_key=os.environ.get('DASHVECTOR_API_KEY'),
endpoint=os.environ.get('DASHVECTOR_ENDPOINT')
)
# Create a collection (skip if one already exists)
client.create(name='understand_partition', dimension=4)
collection = client.get('understand_partition')| Placeholder | Description |
|---|---|
DASHVECTOR_API_KEY | Your DashVector API key |
DASHVECTOR_ENDPOINT | Your cluster endpoint URL |
Create a partition
collection.create_partition(name='shoes')Describe a partition
Retrieve metadata for a specific partition:
ret = collection.describe_partition('shoes')
print(ret)List partitions
List all partitions in a collection:
partitions = collection.list_partitions()
print(partitions)Insert documents into a partition
Pass the partition parameter to route documents to a specific partition. Without it, documents go to the default partition.
collection.insert(
('1', [0.1, 0.1, 0.1, 0.1]),
partition='shoes'
)Query within a partition
Run a vector similarity search scoped to a single partition:
docs = collection.query(
vector=[0.1, 0.1, 0.2, 0.1],
partition='shoes'
)
print(docs)Delete documents from a partition
Remove specific documents from a partition by ID:
collection.delete(ids=['1'], partition='shoes')Get partition statistics
Check the document count and other metrics for a partition:
ret = collection.stats_partition('shoes')
print(ret)Delete a partition
Delete a partition and all documents it contains. The default partition cannot be deleted.
collection.delete_partition('shoes')Use cases
E-commerce image search
A cross-border e-commerce platform stores 20 million clothing product images. Products fall into predefined categories such as shoes, skirts, and pants. Each category maps to a partition. When a user searches by image, the system identifies the product category (either from user input or a classification model) and queries only the corresponding partition.
Video surveillance with time-based partitions
A video surveillance system extracts frames from 1,000 cameras in an industrial park, identifies vehicle features, and imports the vectors into DashVector. Data is retained for 30 days. Create one partition per day and delete partitions as they expire.
Trademark infringement detection
A trademark agent maintains a database of 50 million trademarks divided into nine structural categories: text, graphic, numeric, letter, and others. Each category corresponds to a partition. When checking for infringement, the agent queries only the relevant category partition, narrowing the search space significantly.
Multilingual knowledge base
An international e-commerce team maintains a knowledge base in Chinese, English, and French. Each language maps to a partition. The system embeds the user's question and queries the partition that matches the detected language, returning answers in the correct language.
Multi-tenant data isolation
An e-commerce SaaS provider offers image search capabilities to multiple small and micro businesses. Each customer maps to a partition within a single collection. This provides physical data isolation between tenants while keeping infrastructure costs low.