Vector Database benchmark, streaming case and more

VDBBench is not just an offering of benchmark results for mainstream vector databases and cloud services, it's your go-to tool for the ultimate performance and cost-effectiveness comparison. Designed with ease-of-use in mind, VDBBench is devised to help users, even non-professionals, reproduce results or test new systems, making the hunt for the optimal choice amongst a plethora of cloud services and open-source vector databases a breeze.

Understanding the importance of user experience, we provide an intuitive visual interface. This not only empowers users to initiate benchmarks at ease, but also to view comparative result reports, thereby reproducing benchmark results effortlessly. To add more relevance and practicality, we provide cost-effectiveness reports particularly for cloud services. This allows for a more realistic and applicable benchmarking process.

Closely mimicking real-world production environments, we've set up diverse testing scenarios including insertion, searching, and filtered searching. To provide you with credible and reliable data, we've included public datasets from actual production scenarios, such as SIFT, GIST, Cohere, and a dataset generated by OpenAI from an opensource raw dataset. It's fascinating to discover how a relatively unknown open-source database might excel in certain circumstances!

Prepare to delve into the world of VDBBench, and let it guide you in uncovering your perfect vector database match.

VDBBench is sponsered by Zilliz，the leading opensource vectorDB company behind Milvus. Choose smarter with VDBBench - start your free test on zilliz cloud today!

Leaderboard: https://zilliz.com/benchmark

Install vectordb-bench with only PyMilvus

Install all database clients

Install the specific database client

All the database client supported

OR:

To list the clients that are runnable via the commandline option, execute: vectordbbench --help

To list the options for each command, execute vectordbbench [command] --help

To list the options for awsopensearch, execute vectordbbench awsopensearch --help

Execute tests for the index types: HNSW, HNSW_SQ, or HNSW_BQ.

To list the options for oceanbase, execute vectordbbench oceanbasehnsw --help, The following are some OceanBase-specific command-line options.

Execute tests for the index types: IVF_FLAT, IVF_SQ8, or IVF_PQ.

To list the options for oceanbase, execute vectordbbench oceanbaseivf --help, The following are some OceanBase-specific command-line options.

The vectordbbench command can optionally read some or all the options from a yaml formatted configuration file.

By default, configuration files are expected to be in vectordb_bench/config-files/, this can be overridden by setting
the environment variable CONFIG_LOCAL_DIR or by passing the full path to the file.

The required format is:

Example:

Notes:

• Options passed on the command line will override the configuration file*
• Parameter names use an _ not -

The vectordbbench command can read a batch configuration file to run all the test cases in the yaml formatted configuration file.

By default, configuration files are expected to be in vectordb_bench/config-files/, this can be overridden by setting
the environment variable CONFIG_LOCAL_DIR or by passing the full path to the file.

The required format is:

Example:

Notes:

• Options can only be passed through configuration files
• Parameter names use an _ not -

How to use?

To facilitate the presentation of test results and provide a comprehensive performance analysis report, we offer a leaderboard page. It allows us to choose from QPS, QP$, and latency metrics, and provides a comprehensive assessment of a system's performance based on the test results of various cases and a set of scoring mechanisms (to be introduced later). On this leaderboard, we can select the systems and models to be compared, and filter out cases we do not want to consider. Comprehensive scores are always ranked from best to worst, and the specific test results of each query will be presented in the list below.

1. For each case, select a base value and score each system based on relative values.

   • For QPS and QP$, we use the highest value as the reference, denoted as base_QPS or base_QP$, and the score of each system is (QPS/base_QPS) * 100 or (QP$/base_QP$) * 100.
   • For Latency, we use the lowest value as the reference, that is, base_Latency, and the score of each system is (base_Latency + 10ms)/(Latency + 10ms) * 100.

   We want to give equal weight to different cases, and not let a case with high absolute result values become the sole reason for the overall scoring. Therefore, when scoring different systems in each case, we need to use relative values.

   Also, for Latency, we add 10ms to the numerator and denominator to ensure that if every system performs particularly well in a case, its advantage will not be infinitely magnified when latency tends to 0.

2. For systems that fail or timeout in a particular case, we will give them a score based on a value worse than the worst result by a factor of two. For example, in QPS or QP$, it would be half the lowest value. For Latency, it would be twice the maximum value.

3. For each system, we will take the geometric mean of its scores in all cases as its comprehensive score for a particular metric.

OR:

OR:

If you are using dev container, create the following dataset directory first:

After reopen the repository in container, run python -m vectordb_bench in the container's bash.

To fix the coding styles automatically

This is the main page of VDBBench, which displays the standard benchmark results we provide. Additionally, results of all tests performed by users themselves will also be shown here. We also offer the ability to select and compare results from multiple tests simultaneously.

The standard benchmark results displayed here include all 15 cases that we currently support for 6 of our clients (Milvus, Zilliz Cloud, Elastic Search, Qdrant Cloud, Weaviate Cloud and PgVector). However, as some systems may not be able to complete all the tests successfully due to issues like Out of Memory (OOM) or timeouts, not all clients are included in every case.

All standard benchmark results are generated by a client running on an 8 core, 32 GB host, which is located in the same region as the server being tested. The client host is equipped with an Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz processor. Also all the servers for the open-source systems tested in our benchmarks run on hosts with the same type of processor.

1. Initially, you select the systems to be tested - multiple selections are allowed. Once selected, corresponding forms will pop up to gather necessary information for using the chosen databases. The db_label is used to differentiate different instances of the same system. We recommend filling in the host size or instance type here (as we do in our standard results).
2. The next step is to select the test cases you want to perform. You can select multiple cases at once, and a form to collect corresponding parameters will appear.
3. Finally, you'll need to provide a task label to distinguish different test results. Using the same label for different tests will result in the previous results being overwritten. Now we can only run one task at the same time.

Our client module is designed with flexibility and extensibility in mind, aiming to integrate APIs from different systems seamlessly. As of now, it supports Milvus, Zilliz Cloud, Elastic Search, Pinecone, Qdrant Cloud, Weaviate Cloud, PgVector, Redis, Chroma, etc. Stay tuned for more options, as we are consistently working on extending our reach to other systems.

We've developed lots of comprehensive benchmark cases to test vector databases' various capabilities, each designed to give you a different piece of the puzzle. These cases are categorized into four main types:

• Large Dim: Tests the database's loading capacity by inserting large-dimension vectors (GIST 100K vectors, 960 dimensions) until fully loaded. The final number of inserted vectors is reported.
• Small Dim: Similar to the Large Dim case but uses small-dimension vectors (SIFT 500K vectors, 128 dimensions).

• XLarge Dataset: Measures search performance with a massive dataset (LAION 100M vectors, 768 dimensions) at varying parallel levels. The results include index building time, recall, latency, and maximum QPS.
• Large Dataset: Similar to the XLarge Dataset case, but uses a slightly smaller dataset (10M-1024dim, 10M-768dim, 5M-1536dim).
• Medium Dataset: A case using a medium dataset (1M-1024dim, 1M-768dim, 500K-1536dim).
• Small Dataset: For development (100K-768dim, 50K-1536dim).

• Int-Filter Cases: Evaluates search performance with int-based filter expression (e.g. "id >= 2,000").
• Label-Filter Cases: Evaluates search performance with label-based filter expressions (e.g., "color == 'red'"). The test includes randomly generated labels to simulate real-world filtering scenarios.

• Insertion-Under-Load Case: Evaluates search performance while maintaining a constant insertion workload. VDBBench applies a steady stream of insert requests at a fixed rate to simulate real-world scenarios where search operations must perform reliably under continuous data ingestion.

Each case provides an in-depth examination of a vector database's abilities, providing you a comprehensive view of the database's performance.

Through the /custom page, users can customize their own performance case using local datasets. After saving, the corresponding case can be selected from the /run_test page to perform the test.

We have strict requirements for the data set format, please follow them.

• Folder Path - The path to the folder containing all the files. Please ensure that all files in the folder are in the Parquet format.

  • Vectors data files: The file must be named train.parquet and should have two columns: id as an incrementing int and emb as an array of float32.
  • Query test vectors: The file must be named test.parquet and should have two columns: id as an incrementing int and emb as an array of float32.
    • We recommend limiting the number of test query vectors, like 1,000. When conducting concurrent query tests, Vdbbench creates a large number of processes. To minimize additional communication overhead during testing, we prepare a complete set of test queries for each process, allowing them to run independently. However, this means that as the number of concurrent processes increases, the number of copied query vectors also increases significantly, which can place substantial pressure on memory resources.
  • Ground truth file: The file must be named neighbors.parquet and should have two columns: id corresponding to query vectors and neighbors_id as an array of int.

• Train File Count - If the vector file is too large, you can consider splitting it into multiple files. The naming format for the split files should be train-[index]-of-[file_count].parquet. For example, train-01-of-10.parquet represents the second file (0-indexed) among 10 split files.

• Use Shuffled Data - If you check this option, the vector data files need to be modified. VDBBench will load the data labeled with shuffle. For example, use shuffle_train.parquet instead of train.parquet and shuffle_train-04-of-10.parquet instead of train-04-of-10.parquet. The id column in the shuffled data can be in any order.

Our goals of this benchmark are:

One of the primary goals of VDBBench is to enable users to reproduce benchmark results swiftly and easily, or to test their customized scenarios. We believe that lowering the barriers to entry for conducting these tests will enhance the community's understanding and improvement of vector databases. We aim to create an environment where any user, regardless of their technical expertise, can quickly set up and run benchmarks, and view and analyze results in an intuitive manner.

VDBBench aims to provide a more comprehensive, multi-faceted testing environment that accurately represents the complexity of vector databases. By moving beyond a simple speed test for algorithms, we hope to contribute to a better understanding of vector databases in real-world scenarios. By incorporating as many complex scenarios as possible, including a variety of test cases and datasets, we aim to reflect realistic conditions and offer tangible significance to our community. Our goal is to deliver benchmarking results that can drive tangible improvements in the development and usage of vector databases.

1. Fork the repository and create a new branch for your changes.
2. Adhere to coding conventions and formatting guidelines.
3. Use clear commit messages to document the purpose of your changes.

Step 1: Creating New Client Files

1. Navigate to the vectordb_bench/backend/clients directory.
2. Create a new folder for your client, for example, "new_client".
3. Inside the "new_client" folder, create two files: new_client.py and config.py.

Step 2: Implement new_client.py and config.py

1. Open new_client.py and define the NewClient class, which should inherit from the clients/api.py file's VectorDB abstract class. The VectorDB class serves as the API for benchmarking, and all DB clients must implement this abstract class. Example implementation in new_client.py: new_client.py

2. Open config.py and implement the DBConfig and optional DBCaseConfig classes.
3. The DBConfig class should be an abstract class that provides information necessary to establish connections with the database. It is recommended to use the pydantic.SecretStr data type to handle sensitive data such as tokens, URIs, or passwords.
4. The DBCaseConfig class is optional and allows for providing case-specific configurations for the database. If not provided, it defaults to EmptyDBCaseConfig. Example implementation in config.py:

Step 3: Importing the DB Client and Updating Initialization

In this final step, you will import your DB client into clients/init.py and update the initialization process.

1. Open clients/init.py and import your NewClient from new_client.py.
2. Add your NewClient to the DB enum.
3. Update the db2client dictionary by adding an entry for your NewClient. Example implementation in clients/init.py:

Step 4: Implement new_client/cli.py and vectordb_bench/cli/vectordbbench.py

In this (optional, but encouraged) step you will enable the test to be run from the command line.

1. Navigate to the vectordb_bench/backend/clients/"client" directory.
2. Inside the "client" folder, create a cli.py file. Using zilliz as an example cli.py:

3. Update cli by adding:
   1. Add database specific options as an Annotated TypedDict, see ZillizTypedDict above.
   2. Add index configuration specific options as an Annotated TypedDict. (example: vectordb_bench/backend/clients/pgvector/cli.py)
      1. May not be needed if there is only one index config.
      2. Repeat for each index configuration, nesting them if possible.
   3. Add a index config specific function for each index type, see Zilliz above. The function name, in lowercase, will be the command name passed to the vectordbbench command.
   4. Update db_config and db_case_config to match client requirements
   5. Continue to add new functions for each index config.
   6. Import the client cli module and command to vectordb_bench/cli/vectordbbench.py (for databases with multiple commands (index configs), this only needs to be done for one command)
   7. Import the get_custom_case_config function from vectordb_bench/cli/cli.py and use it to add a new key custom_case to the parameters variable within the command.

cli modules with multiple index configs:

• pgvector: vectordb_bench/backend/clients/pgvector/cli.py
• milvus: vectordb_bench/backend/clients/milvus/cli.py

That's it! You have successfully added a new DB client to the vectordb_bench project.

The system under test can be installed in any form to achieve optimal performance. This includes but is not limited to binary deployment, Docker, and cloud services.

For the system under test, we use the default server-side configuration to maintain the authenticity and representativeness of our results. For the Client, we welcome any parameter tuning to obtain better results.

Many databases may not be able to complete all test cases due to issues such as Out of Memory (OOM), crashes, or timeouts. In these scenarios, we will clearly state these occurrences in the test results.

We strive for accuracy in learning and supporting various vector databases, yet there might be oversights or misapplications. For any such occurrences, feel free to raise an issue or make amendments on our GitHub page.

In our pursuit to ensure that our benchmark reflects the reality of a production environment while guaranteeing the practicality of the system, we have implemented a timeout plan based on our experiences for various tests.

1. Capacity Case:

• For Capacity Case, we have assigned an overall timeout.

2. Other Cases:

For other cases, we have set two timeouts:

• Data Loading Timeout: This timeout is designed to filter out systems that are too slow in inserting data, thus ensuring that we are only considering systems that is able to cope with the demands of a real-world production environment within a reasonable time frame.

• Optimization Preparation Timeout: This timeout is established to avoid excessive optimization strategies that might work for benchmarks but fail to deliver in real production environments. By doing this, we ensure that the systems we consider are not only suitable for testing environments but also applicable and efficient in production scenarios.

This multi-tiered timeout approach allows our benchmark to be more representative of actual production environments and assists us in identifying systems that can truly perform in real-world scenarios.

Note: Some datapoints in the standard benchmark results that voilate this timeout will be kept for now for reference. We will remove them in the future.