DuckDB vs M3

Choosing the right database is a critical choice when building any software application. All databases have different strengths and weaknesses when it comes to performance, so deciding which database has the most benefits and the most minor downsides for your specific use case and data model is an important decision. Below you will find an overview of the key concepts, architecture, features, use cases, and pricing models of DuckDB and M3 so you can quickly see how they compare against each other.

The primary purpose of this article is to compare how DuckDB and M3 perform for workloads involving time series data, not for all possible use cases. Time series data typically presents a unique challenge in terms of database performance. This is due to the high volume of data being written and the query patterns to access that data. This article doesn’t intend to make the case for which database is better; it simply provides an overview of each database so you can make an informed decision.

DuckDB Overview

DuckDB is an in-process SQL OLAP (Online Analytical Processing) database management system. It is designed to be simple, fast, and feature-rich. DuckDB can be used for processing and analyzing tabular datasets, such as CSV or Parquet files. It provides a rich SQL dialect with support for transactions, persistence, extensive SQL queries, and direct querying of Parquet and CSV files. DuckDB is built with a vectorized engine that is optimized for analytics and supports parallel query processing. It is designed to be easy to install and use, with no external dependencies and support for multiple programming languages.

M3 Overview

M3 is a distributed time series database written entirely in Go. It is designed to collect a high volume of monitoring time series data, distribute storage in a horizontally scalable manner, and efficiently leverage hardware resources. M3 was initially developed by Uber as a scalable remote storage backend for Prometheus and Graphite and later open-sourced for broader use.

DuckDB for Time Series Data

DuckDB can be used effectively with time series data. It supports processing and analyzing tabular datasets, which can include time series data stored in CSV or Parquet files. With its optimized analytics engine and support for complex SQL queries, DuckDB can perform aggregations, joins, and other time series analysis operations efficiently. However, it’s important to note that DuckDB is not specifically designed for time series data management and may not have specialized features tailored for time series analysis like some dedicated time series databases.

M3 for Time Series Data

M3 is specifically designed for time-series data. It is a distributed and scalable time-series database optimized for handling large volumes of high-resolution data points, making it an ideal solution for storing, querying, and analyzing time-series data.

M3’s architecture focuses on providing fast and efficient querying capabilities, as well as high ingestion rates, which are essential for working with time-series data. Its horizontal scalability and high availability ensure that it can handle the demands of large-scale deployments and maintain performance as data volumes grow.

DuckDB Key Concepts

• In-process: DuckDB operates in-process, meaning it runs within the same process as the application using it, without the need for a separate server.
• OLAP: DuckDB is an OLAP database, which means it is optimized for analytical query processing.
• Vectorized engine: DuckDB utilizes a vectorized engine that operates on batches of data, improving query performance.
• Transactions: DuckDB supports transactional operations, ensuring the atomicity, consistency, isolation, and durability (ACID) properties of data operations.
• SQL dialect: DuckDB provides a rich SQL dialect with advanced features such as arbitrary and nested correlated subqueries, window functions, collations, and support for complex types like arrays and structs

M3 Key Concepts

• Time Series Compression: M3 has the ability to compress time series data, resulting in significant memory and disk savings. It uses two compression algorithms, M3TSZ and protobuf encoding, to achieve efficient data compression.
• Sharding: M3 uses virtual shards that are assigned to physical nodes. Timeseries keys are hashed to a fixed set of virtual shards, making horizontal scaling and node management seamless.
• Consistency Levels: M3 provides variable consistency levels for read and write operations, as well as cluster connection operations. Write consistency levels include One (success of a single node), Majority (success of the majority of nodes), and All (success of all nodes). Read consistency level is One, which corresponds to reading from a single nod

DuckDB Architecture

DuckDB follows an in-process architecture, running within the same process as the application. It is a relational table-oriented database management system that supports SQL queries for producing analytical results. DuckDB is built using C++11 and is designed to have no external dependencies. It can be compiled as a single file, making it easy to install and integrate into applications.

M3 Architecture

M3 is designed to be horizontally scalable and handle high data throughput. It uses fileset files as the primary unit of long-term storage, storing compressed streams of time series values. These files are flushed to disk after a block time window becomes unreachable. M3 has a commit log, equivalent to the commit log or write-ahead-log in other databases, which ensures data integrity. Client Peer streaming is responsible for fetching blocks from peers for bootstrapping purposes. M3 also implements caching policies to optimize efficient reads by determining which flushed blocks are kept in memory.