Read AI: 5 billion tuples, 20ms p99 latency

Read AI is the AI meeting notetaker and assistant trusted by more than 100,000 organizations and 75% of the Fortune 500, adding more than one million new customers every month. OpenFGA backs the authorization layer that lets Read AI safely share intelligence across meetings, messages, email, and documents.

At a glance

Industry	AI productivity / meeting intelligence
In production since	April 28, 2023
Peak load	5,200 RPS
Latency	20ms p99 / 1.8ms average
Tuple count	5,323,283,829 (and growing)
Version	v1.8.16
Storage	PostgreSQL

Why OpenFGA

Read AI ran a proprietary, organically built authorization system that hit performance and scalability ceilings as the platform grew. The team evaluated alternatives such as Authzed before choosing OpenFGA, citing:

• Zanzibar foundations that aligned with the sharing semantics the product needed.
• Documentation clarity, especially the practical examples and modeling guides.
• The ability to self-host with predictable cost.
• Approachable, responsive maintainers.

Production at scale

The self-hosted OpenFGA service handles peak load of 5,200 requests per second with a 20ms p99 latency and 1.8ms average latency. The data store holds more than 5.3 billion tuples and grows daily.

OpenFGA upgrades are folded into a monthly cadence. The OpenFGA release pace is faster than Read AI's, but upgrades have been smooth with no significant backward-compatibility issues.

Outcomes

• Confidence in secure data authorization across the entire product surface.
• Adoption of ReBAC best practices improved internal design decisions.
• Compute and hosting costs dropped versus the prior solution.
• OpenFGA has not been the bottleneck even at peak.

Source

This case study is based on the public CNCF TOC adopter interview with Andrew Powers, Software Engineering Manager at Read AI, available in the cncf/toc repository.