Learn
How to Pick the Right Vector Database for Your Use Case

How to Choose a Vector Database: Qdrant Cloud vs. Zilliz Cloud

Oct 17, 20249 min read

Compare Qdrant Cloud and Zilliz Cloud (fully managed Milvus) in this in-depth benchmark, cost, and features comparison.

Read the entire series

This Qdrant comparison was last updated on October 17, 2024. To provide you with the latest findings, this blog will be regularly updated with the latest benchmark figures.

Not all vector databases are made equal: unveiling the distinction

In the fast-paced world of artificial intelligence and machine learning, vector databases have become essential for managing and querying high-dimensional data. Last year, we saw a significant shift as vector search use cases moved from prototype to production, changed market dynamics, and raised the bar for database performance and scalability. As the demand for these specialized databases grows with real-world, large-scale deployments, it’s essential to understand that not all solutions are created equal. This article will compare two leading players in the space: Qdrant and Zilliz Cloud, each tackling production scale vector search in very different ways.

Qdrant: Optimized for Medium-Scale Deployments

Qdrant has gained recognition for its focus on high-performance vector search, particularly for medium-sized datasets:

HNSW Optimization: Qdrant has done an excellent job promoting the strengths of the Hierarchical Navigable Small World (HNSW) algorithm, known for its low latency, high throughput, and high recall rates.
Cost-Effective Features: The platform has introduced several features to enhance cost-effectiveness, including quantization and memory-mapped (mmap) file I/O, which can significantly reduce memory usage and improve performance for certain workloads.
Performance Sweet Spot: Qdrant's architecture is particularly well-suited for datasets up to approximately 100 million vectors, where it can maintain impressive performance metrics.

However, Qdrant faces challenges when scaling to larger datasets:

Scalability Limitations: For datasets exceeding 100 million vectors, Qdrant's performance may begin to degrade, and costs can become prohibitive due to its reliance on in-memory indexing.
Resource Intensity: As dataset sizes grow, the memory requirements for Qdrant can increase substantially, potentially leading to higher infrastructure costs.

Zilliz Cloud: Engineered for Large-Scale Deployments

In contrast, Zilliz Cloud has focused on addressing the challenges of extremely large-scale vector search:

Cardinal Search Engine: Zilliz has developed a proprietary search engine called Cardinal, which can leverage a mix of the graph and IVF index, which are in-memory and disk-based indexing. This hybrid approach allows Zilliz to provide more cost-efficient search capabilities for large datasets.
Disk-Based Indexing Advantages: Using disk-based indexes significantly reduces memory requirements and scales better with increasing dataset sizes. This makes Zilliz more suitable for very large workloads while maintaining higher query throughput compared to memory-mapped HNSW solution adopted by Qdrant, especially when the working set exceeds available RAM.

Managed Services and Operational Considerations

While both platforms offer cloud-based solutions, there are significant differences in their approach to managed services:

Qdrant's Self-Managed Focus: Qdrant is primarily designed as an open-source vector database product. While it offers a cloud product, it requires more hands-on management from developers. As noted in Qdrant's comparison with other services, it may not be the best option for customers seeking a fully managed SaaS solution that abstracts infrastructure complexity away.
Scaling Complexities with Qdrant: Qdrant's horizontal scaling requires pre-planning, such as setting up shards when creating a collection. Resharding is not supported, which can limit flexibility as data grows. Scaling up often requires significant engineering effort.
Zilliz's Managed Approach: Zilliz Cloud provides a real managed service, offering features like one-click horizontal scaling without extensive pre-planning or engineering efforts and minimal ops effort, such as version upgrades, among others.

Enterprise Readiness and Security Features

When it comes to enterprise-grade features and security capabilities, Qdrant and Zilliz Cloud take different approaches. Qdrant, as an open-source solution, offers a more basic set of enterprise features, focusing primarily on core functionality and leaving more advanced features to be implemented by developers or through third-party integrations. In contrast, Zilliz Cloud provides a more comprehensive suite of enterprise-ready features out of the box. This includes advanced security measures, extensive monitoring and alerting capabilities, and a broader range of compliance certifications. While both platforms can be used in enterprise environments, organizations with stringent security, compliance, and monitoring requirements may find Zilliz Cloud's offerings more aligned with their needs without requiring additional development or integration work. We’ll provide a detailed feature comparison later in this blog.

We will venture beyond these distinctions within this Qdrant Cloud vs. Zilliz Cloud comparison, digging deep into several other comparison areas. We'll dive into benchmarks to offer a performance perspective and perform an in-depth Qdrant Cloud feature analysis.

Benchmarking Qdrant Cloud and Zilliz Cloud

At Zilliz, developers and architects alike ask us, "How does Zilliz measure up to Qdrant Cloud for vector embedding workloads?" This curiosity might spring from various motivations, all stemming from building semantic similarity search for capabilities like product recommenders or reverse image search, and lately, when building LLM applications with Retrieval Augmented Generation (RAG).

Over the past few weeks, we undertook a comprehensive analysis, delving into the performance and attributes of Zilliz Cloud vs. Qdrant Cloud regarding vector embedding workloads. With our open-source benchmark tool, VectorDBBench, our exploration focused intently on critical aspects such as queries per second (QPS), queries per dollar (QP$), and latency.

Large-size datasets tested (≥5M vectors)

Dataset 1: 10,000,000 vectors with 768 dimensions
Dataset 2: 5,000,000 vectors with 1,536 dimensions

Products tested (with similar capabilities)

Zilliz Cloud (8cu-perf): Zilliz Cloud with eight performance-optimized compute units
Zilliz Cloud (2cu-cap): Zilliz Cloud with two capacity-optimized compute units
Qdrant Cloud (4c16g-5node): Qdrant Cloud with four CPUs and 16G of RAM using five machines

For more information about Zilliz Cloud’s compute units, see Zilliz blog introducing Zilliz Cloud CU type and size.

Results: QPS

When processing 10 million vectors with 768 dimensions, Zilliz Cloud options outperform Qdrant Cloud by over 7x and 1x, respectively, in QPS.

When processing 5 million vectors with 1536 dimensions, Zilliz Cloud(8cu-perf) outperforms Qdrant Cloud by almost 8x in QPS, while Zilliz Cloud (2cu-cap) shows lower performance than Qdrant Cloud.

Results: QP$

When processing 10 million vectors with 768 dimensions, Zilliz Cloud options outperform Qdrant Cloud by over 8x and 5x respectively in QP$.

When processing 5 million vectors with 1536 dimensions, Zilliz Cloud options outperform Qdrant Cloud by 7x and 3x respectively in QP$.

Results: Latency

When processing 10 million vectors with 768 dimensions, Zilliz Cloud options outperform Qdrant Cloud by over 12x and 3x respectively in latency.

When processing 5 million vectors with 1536 dimensions, Zilliz Cloud options outperform Qdrant Cloud by over 8x and 1x respectively in latency.

Medium-size datasets tested (< 5M vectors)

Dataset 3: 1,000,000 vectors with 768 dimensions

Dataset 4: 500,000 vectors with 1,536 dimensions