# Vector Index Capacity Planning

When preparing to create a MyScale cluster, selecting the right configuration is crucial, considering your vector capacity and performance needs.

Configuring MyScale clusters is straightforward, involving two key parameters: pod size and replica count. These parameters are easily adjustable to adapt to evolving requirements.

Cluster Configuration

# Factors to Consider

To determine the best MyScale cluster setup, consider:

  • Type of vector index
  • Number of vectors
  • Dimensions of vectors
  • Queries per second (QPS) requirement

We advocate using MyScale's MSTG algorithm for vector indexing, known for exceptional performance and vector density. Our capacity and performance guidelines are based on MSTG (see our blog post (opens new window) for detailed benchmarks).

Other metadata have minimal impact on vector capacity, but QPS could be influenced if queries include additional metadata. To optimize QPS in such scenarios, MyScale provides efficient filter search (opens new window).

# Pod Sizes

MyScale combines algorithmic innovations, like MSTG, with system engineering to achieve high vector density and performance. We offer a single pod type, with size options like x4 providing quadruple the capacity and CPU performance of an x1 pod.

Pod sizes available are x1, x2, x4, x8, x16, and x32. Use our Price Estimator on the pricing page (opens new window) to find the best pod size for your needs.

Price Estimator

Following are some estimated capacity and performance benchmarks for the x1 pod.

# Capacity

Table 1: Maximum Vector Capacity per Pod by Dimension

Pod Size Dimensions Max Capacity
x1 512 7,500,000
768 5,000,000
1024 3,750,000
1536 2,500,000

Opt for a larger pod size for increased vector capacity.


Table 2: QPS Estimates for 5M Vectors at 768 Dimensions by top_k*

Pod Size top_k 10 top_k 100
x1 152 117

*Based on MSTG index with alpha=3 (default setting). See full benchmarks (opens new window) for more.

Boost QPS by adding replicas.

# Replicas

Replicas enhance cluster availability and QPS. A cluster's total QPS roughly equals the QPS of a single pod multiplied by the number of replicas. Note that replicas do not increase vector capacity.

# Examples

# Example 1: ChatData Application

In our ChatData (opens new window) app, we manage an arXiv papers dataset with approximately 2.25 million vectors at 768 dimensions. A single x1 pod, supporting up to 5 million vectors at this dimension, suffices for this dataset.

Consider a large scale image search application requiring 80 million vectors at 512 dimensions. Referring to the Capacity section, an x16 pod is needed. For higher availability and QPS, increase the replica count to 2 or 3.

Why `x16` Pod?

An x1 pod supports 7.5 million vectors at 512 dimensions. Given the requirement of 80 million vectors, x16 is the suitable choice among our pod sizes of x1, x2, x4, x8, x16, and x32.

# Visualized Usage

The cluster monitoring page features a chart displaying the memory usage by the loaded vector index relative to total available memory. This tool helps users track vector usage and plan pod size upgrades before reaching capacity limits.

Loaded vector index memory percent

We employ the LRU algorithm for in-memory vector index management. Exceeding a pod's vector limit results in some indexes being unloaded from memory. Re-loading these indexes for subsequent searches may reduce QPS or increase query latency. In such cases, consult the aforementioned chart and consider increasing your pod size accordingly.

Last Updated: Mon Jun 24 2024 10:08:20 GMT+0000