Skip to main content
XenonStack

Enterprise Data Management

XenonStack Insights AresDB: Uber's GPU Real-Time Analytics Engine

| Published | Updated

AresDB architecture: fact and dimension tables, live store, archive store and AQL queries on GPU

Key takeaway

  • AresDB is Uber's GPU-powered real-time analytics engine. See why it was built, how fact and dimension tables work, the AQL query format and its limits.

AresDB is Uber's open-source, GPU-powered real-time analytics storage and query engine, built to answer queries on fresh event data in parallel on graphics processors. According to Uber's engineering blog, it was released in November 2018.

This guide is for data engineers and architects comparing real-time analytics engines. It explains why Uber built AresDB, how its storage and tables work, what its documented limits are, and how AI agents change the way such an engine is used. It describes the project as published at release, so check the repository for its current status before choosing it.

Why did Uber build AresDB?

Uber's post says it needed real-time analytics for dashboards, automated decisions such as pricing and fraud detection, and ad hoc troubleshooting. Existing options each lacked something it needed. Uber says Apache Pinot lacked key-based deduplication, upserts and joins, and Elasticsearch was not optimised for time-range storage and did not support joins. It also reports that other GPU databases such as Kinetica and OmniSci were missing capabilities such as deduplication, schema alteration or geospatial joins, or were academic prototypes not built for production.

General-purpose cloud analytics services solve a related problem at a different scale; see AWS data lake and analytics solutions for that route. AresDB targets the narrower case of fresh, time-based event data queried at speed.

How is AresDB architected?

The Uber blog describes the main design choices. Most data sits in host memory. Data is ingested through CPUs, recovered from disk, and copied to GPU memory at query time.

Two-column diagram. Ingest and store: upsert batch, live store, backfill queue, archive store. Query and limits: AQL query, GPU execution, join limits, memory budget.
  • Live store. Holds recent, uncompressed, unsorted columnar batches, with a primary key index that handles deduplication and updates. This is where real-time data lands first.
  • Archive store. Holds sorted, compressed columnar data, batched by UTC day. Compression uses run-length encoding on columns in the user-defined sort order.
  • Ingestion path. Upsert batches are written to a redo log first. Late records go to a backfill queue, and a scheduled archiving process moves mature records into the archive store.
  • Backfill. Asynchronous and idempotent, used for very late data, manual fixes and new columns.

What are fact tables and dimension tables in AresDB?

AresDB has no database or schema scope: all tables in a cluster share one namespace. It uses two table types.

Table typeWhat it storesTypical exampleTime column
Fact tableOngoing event streams, usually queried by event timeTripsYes, an event time
Dimension tableBounded, current properties of entitiesCitiesNot needed

How do you query AresDB?

Queries use Ares Query Language (AQL), a structured JSON, YAML or Go-based format and not SQL. A query is an object with fields for dimensions, measures, row filters and time filters, with built-in time bucketing and time zone support. Uber says this makes queries easier to compose in code and avoids SQL injection risk. At release, the post says SQL support was planned but not yet available.

What are the documented limits of AresDB?

Uber's post lists several limits at release. Strings are converted to enumerated values, so concatenation, substring and regular expressions are not supported. Joins are limited to hash joins from fact tables to dimension tables. Window functions were not yet available, and input data is not cached in GPU memory across queries. Memory is capped by a configured budget, and when the backfill queue fills, clients are blocked until it clears. The distributed design, with replication and sharding, was still in progress.

The AresDB repository README describes it as a GPU-powered real-time analytics storage and query engine and states that it needs an NVIDIA driver and the CUDA toolkit, with CPU query mode available through a build option. The versions it lists date from the 2018 release era, so verify current requirements and maintenance activity before adopting it.

Where does AresDB sit among related tools?

Our guides to M3DB and Uber Marmaray cover related projects from the same engineering ecosystem; see also Marmaray features and best practices. For background on GPU computing, read GPU computing for data. For a concurrency pattern you may meet in supporting services, see coroutines benefits and use cases.

How do AI agents change real-time analytics engines?

A person reads a dashboard once. An agent may query the same engine many times while it decides what to do, which raises the stakes on freshness, deduplication and query limits. A structured query format such as AQL suits agents because queries are composed as objects, but every agent query still needs permissions and a record of which data informed which decision. ElixirData is XenonStack's Context OS for governing the data and decision context that agents draw on.

What should you do first?

  1. Write down the latency, freshness and query shapes your use case needs.
  2. Check them against the limits above: joins, strings, window functions and memory.
  3. Confirm the project's current maintenance status and GPU requirements.
  4. Test with one fact table and one dimension table on representative data.
  5. Decide who may query it, human or agent, and how access is logged.

To map one decision your agents make from real-time data, and the data and policy it depends on, book a Context Workshop. See what happens in the workshop. For a data strategy discussion, you can also talk to our big data specialists.

Frequently asked questions

What is AresDB?

AresDB is Uber's open-source GPU-powered real-time analytics storage and query engine, released in November 2018.

What is the difference between a fact table and a dimension table in AresDB?

A fact table stores ongoing event streams and is usually queried by event time. A dimension table stores bounded, current properties of entities and needs no time column.

What query language does AresDB use?

AresDB uses Ares Query Language (AQL), a structured JSON, YAML or Go-based format. At release, SQL support was planned but not available.

What are the limits of AresDB?

Uber's post lists limits including no string concatenation or regular expressions, hash joins only from fact tables to dimension tables, and a configured memory budget.

Does AresDB require a GPU?

The repository README says it needs an NVIDIA driver and the CUDA toolkit, and that queries can run in CPU mode through a build option.

Ready to put AI agents to work? See an ElixirClaw demo: the agentic OS to build, orchestrate and govern AI agents across your enterprise systems.

Table of Contents

Get the latest articles in your inbox

Subscribe Now

Interested in solving your challenges with the XenonStack team?
Get Started

Tell us about you and your priorities so we can prepare the right conversation. It takes about 2 minutes.

Please Select your Industry
Banking
Fintech
Payment Providers
Wealth Management
Discrete Manufacturing
Semiconductor
Machinery Manufacturing / Automation
Appliances / Electrical / Electronics
Elevator Manufacturing
Defense & Space Manufacturing
Computers & Electronics / Industrial Machinery
Motor Vehicle Manufacturing
Food and Beverages
Distillery & Wines
Beverages
Shipping
Logistics
Mobility (EV / Public Transport)
Energy & Utilities
Hospitality
Digital Gaming Platforms
SportsTech with AI
Public Safety - Explosives
Public Safety - Firefighting
Public Safety - Surveillance
Public Safety - Others
Media Platforms
City Operations
Airlines & Aviation
Defense Warfare & Drones
Robotics Engineering
Drones Manufacturing
AI Labs for Colleges
AI MSP / Quantum / AGI Institutes
Retail Apparel and Fashion

Continue