Scaling Git at OpenAI

OpenAI's new Git system handles 1 billion users with faster pushes and reliable merges.

16 cards · 2m · tap to begin

From · · 2m

Scaling Git at OpenAI

OpenAI's new Git system handles 1 billion users with faster pushes and reliable merges.

In brief

OpenAI's new Git system handles 1 billion users with faster pushes and reliable merges. OpenAI launched Pando, a Git system that scales to handle 1 billion users with 5x faster pushes and 99% cache efficiency. Originally reported by OpenAI.

Why OpenAI needed new Git system

OpenAI's monorepo faced slow fetches and failed merges as Codex increased parallel development demands.

Pando's dual goals

Pando scales writes across branches and maximizes read throughput for 1 billion users.

Pando's causal consistency model

Pando uses causal consistency to allow parallel writes without full repository consistency.

Pando vs. existing Git systems

Pando outperforms three-phase commit and write-ahead log systems in high-concurrency environments.

Pando's packfile reuse

Pando reuses packfiles across clients to reduce network bandwidth and server load.

Pando's metadata store

Pando stores ref versions in a horizontally scalable metadata service for efficient packfile selection.

Pando's compaction system

Pando uses geometric compaction to reduce fetch latency and improve packfile efficiency.

Pando's deployment timeline

Pando moved from development to production in 10 weeks for OpenAI's monorepo.

Pando's performance metrics

Pando cut median push latency by 5x and delivers nearly a petabyte of git data per day.

Pando's cache efficiency

Pando achieves 99% cache hit rates with regional caches for low-latency git operations.

Pando's contribution to Git

OpenAI contributed improvements to git's packfile-uri and explicit haves features.

Pando's future plans

OpenAI will share Pando technology with other organizations for agent-scale development.

Pando's impact on Codex

Codex helped developers migrate to Pando with skills for checkout and branch management.

Pando's scalability

Pando handles 57 terabytes of git packs in one hour with minimal server resources.

Pando's real-world testing

Pando reduced median push latency by 80% and improved merge reliability in OpenAI's monorepo.

Pando's next steps

OpenAI will share Pando's technology with other organizations for agent-scale development.

The takeaway

OpenAI launched Pando, a Git system that scales to handle 1 billion users with 5x faster pushes and 99% cache efficiency.

Read the original on OpenAI

React

Sign in to react and comment.

Comments (0)

Life is short. Keep it sweet. Respect others' opinions and be kind!

    Recommended next

    More decks on openai and related topics.