Apache Druid vs Prometheus

Apache Druid Overview

Apache Druid is an open-source, real-time analytics database designed for high-performance querying and data ingestion. Originally developed by Metamarkets in 2011 and later donated to the Apache Software Foundation in 2018, Druid has gained popularity for its ability to handle large volumes of data with low latency. With a unique architecture that combines elements of time series databases, search systems, and columnar storage, Druid is particularly well-suited for use cases involving event-driven data and interactive analytics.

Prometheus Overview

Prometheus is an open-source monitoring and alerting toolkit initially developed at SoundCloud in 2012. It has since become a widely adopted monitoring solution and a part of the Cloud Native Computing Foundation (CNCF) project. Prometheus focuses on providing real-time insights and alerts for containerized and microservices-based environments. Its primary use case is monitoring infrastructure and applications, with an emphasis on reliability and scalability.

Apache Druid for Time Series Data

Apache Druid is designed for real time analytics and can be a good fit for working with time series data that needs to be analyzed quickly after being written. Druid also offers integrations for storing historical data in cheaper object storage so historical time series data can also be analyzed using Druid.

Prometheus for Time Series Data

Prometheus is specifically designed for time series data, as its primary focus is on monitoring and alerting based on the state of infrastructure and applications. It uses a pull-based model, where the Prometheus server scrapes metrics from the target systems at regular intervals. This model is suitable for monitoring dynamic environments, as it allows for automatic discovery and monitoring of new instances. However, Prometheus is not intended as a general-purpose time series database and might not be the best choice for high cardinality or long-term data storage.

Apache Druid Key Concepts

• Data Ingestion: The process of importing data into Druid from various sources, such as streaming or batch data sources.
• Segments: The smallest unit of data storage in Druid, segments are immutable, partitioned, and compressed.
• Data Rollup: The process of aggregating raw data during ingestion to reduce storage requirements and improve query performance.
• Nodes: Druid’s architecture consists of different types of nodes, including Historical, Broker, Coordinator, and MiddleManager/Overlord, each with specific responsibilities.
• Indexing Service: Druid’s indexing service manages the process of ingesting data, creating segments, and publishing them to deep storage.

Prometheus Key Concepts

• Metric: A numeric representation of a particular aspect of a system, such as CPU usage or memory consumption.
• Time Series: A collection of data points for a metric, indexed by timestamp.
• Label: A key-value pair that provides metadata and context for a metric, enabling more granular querying and aggregation.
• PromQL: Prometheus uses its own query language called PromQL (Prometheus Query Language) for querying time series data and generating alerts.

Apache Druid Architecture

Apache Druid is a powerful distributed data store designed for real-time analytics on large datasets. Within its architecture, several core components play pivotal roles in ensuring its efficiency and scalability. Here is an overview of the core components that power Apache Druid.

• Historical Nodes are fundamental to Druid’s data-serving capabilities. Their primary responsibility is to serve stored data to queries. To achieve this, they load segments from deep storage, retain them in memory, and then cater to the queries on these segments. When considering deployment and management, these nodes are typically stationed on machines endowed with significant memory and CPU resources. Their scalability is evident as they can be expanded horizontally simply by incorporating more nodes.
• Broker Nodes act as the gatekeepers for incoming queries. Their main function is to channel these queries to the appropriate historical nodes or real-time nodes. Intriguingly, they are stateless, which means they can be scaled out to accommodate an increase in query concurrency.
• Coordinator Nodes have a managerial role, overseeing the data distribution across historical nodes. Their decisions on which segments to load or drop are based on specific configurable rules. In terms of deployment, a Druid setup usually requires just one active coordinator node, with a backup node on standby for failover scenarios.
• Overlord Nodes dictate the assignment of ingestion tasks, directing them to either middle manager or indexer nodes. Their deployment mirrors that of the coordinator nodes, with typically one active overlord and a backup for redundancy.
• MiddleManager and Indexer Nodes are the workhorses of data ingestion in Druid. While MiddleManagers initiate short-lived tasks for data ingestion, indexers are designed for long-lived tasks. Given their intensive operations, these nodes demand high CPU and memory resources. Their scalability is flexible, allowing horizontal expansion based on the volume of data ingestion.
• Deep Storage is a component that serves as Druid’s persistent storage unit. Druid integrates with various blob storage solutions like HDFS, S3, and Google Cloud Storage.
• Metadata Storage is the repository for crucial metadata about segments, tasks, and configurations. Druid is compatible with popular databases for this purpose, including MySQL, PostgreSQL, and Derby.

Prometheus Architecture

Prometheus is a single-server, standalone monitoring system that uses a pull-based approach to collect metrics from target systems. It stores time series data in a custom, highly compressed, on-disk format, optimized for fast querying and low resource usage. The architecture of Prometheus is modular and extensible, with components like exporters, service discovery mechanisms, and integrations with other monitoring systems. As a non-distributed system, it lacks built-in clustering or horizontal scalability, but it supports federation, allowing multiple Prometheus servers to share and aggregate data.