databricks Databricks Blog ·

Databricks Re-architects Serverless Network Config Delivery for 97.5% Latency Cut

bloginfradatabricksarchitect
feature

Databricks has re-architected its serverless network configuration delivery, moving from synchronous upstream calls to an event-driven pre-computation pipeline. This change significantly improved performance and reliability for its serverless compute platform, which launches tens of millions of VMs daily. The new architecture cut RPC p99 latency by 97.5% (from 5,000ms to 125ms) and boosted service availability to 99.99%. It also reduced upstream call volume by 86%, benefiting customers using SQL warehouses, notebooks, and ML serving endpoints.

Features (1)
  • Re-architected Serverless Network Configuration Delivery

    Databricks re-architected its serverless network configuration delivery, moving from synchronous upstream calls to an event-driven pipeline that pre-computes configurations and serves them from a snapshot store. This change reduced RPC p99 latency by 97.5% (5,000ms to 125ms), boosted availability to 99.99%, and cut upstream call volume by 86% across billions of daily requests.

Read the original announcement →

https://www.databricks.com/blog/databricks-network-configuration-delivery-tens-millions-serverless-vms

Related releases