Growing pains at AI speed: How Hugging Face and Instacart scaled their in-memory

AI companies grow at rates legacy infrastructure wasn’t built for, and the in-memory data store often hits hard limits. Bottlenecks on threads, risky resharding, climbing tail latency, and memory spikes during backups all risk production outages.

In this 30-minute session, you'll hear how two teams hit that ceiling and got past it. Hugging Face had to split its Redis setup to keep performance up as the Hub raced to 3 million models, then brought in Dragonfly to simplify operations. Instacart replaced a 348-node ElastiCache cluster with 100 Dragonfly nodes and cut average and P99 latency by 40 to 50%. Both did this while growing rapidly and without impacting their customers.

You'll learn:

• The five ways in-memory data stores break under AI-scale growth, and the warning signs to watch for
• Why scaling up on fewer, larger nodes beats adding shards
• How Hugging Face and Instacart made the switch, and what changed afterward

Who should attend: AI/ML, platform, DevOps, and engineering leaders at fast-growing AI companies. Stay for live Q&A with the Dragonfly team.

Speakers: Jim Allen Wallace, Product Marketing Leader, Dragonfly. Owen Taylor, Principal Solution Architect, Dragonfly.