Awaiting customer approval. This page isn't marketed anywhere.
Cerebras
Scaling AI Hardware CI Throughput With Batch Merge & Two-Step CI
Cerebras
Location
Sunnyvale, California, USA
Customer Since
October 2023
Team Size
350 engineers
Rushin Shah
Sr. Manager, DevInfra and Automation Architect
Cerebras Systems builds the world's fastest AI inference and training platform. Organizations across medical research, cryptography, energy, and agentic AI use Cerebras systems to build on-prem supercomputers, while enterprises globally adopt their AI cloud services. As the company accelerates into its next decade, engineering velocity and CI reliability have become strategic foundations for its success.
Long-running CI jobs blocked engineers unnecessarily
Limited hardware capacity required careful scheduling
Heavy merge pressure from a fast-moving monorepo
With 350 engineers committing daily to a large monorepo, their CI pipeline became one of the most demanding parts of their development workflow. Hardware tests are routinely run for 1.5 to 3 hours and can only run on a limited pool of specialized machines. Every merge had the potential to block others, CI pressure accumulated quickly, and the cost of re-running long pipelines created friction that slowed both iteration speed and release cadence. Hardware CI is unforgiving: when a test takes hours and resources are limited, you can't afford unpredictability.
Two-step CI completely transformed how we handle long hardware tests. Instead of burning hours on repeated runs, we validate once and merge with confidence. Our CI pipeline actually works with our hardware constraints, rather than fighting against them.
Rushin Shah
Sr. Manager, DevInfra and Automation Architect
Mergify became the backbone that stabilized this environment. By combining two-step CI with batch merging, Cerebras reduced the number of full hardware test cycles needed per day while preserving the reliability of the main branch. Fast, lightweight checks run first; only once a change is confirmed safe does Mergify construct a clean speculative branch for the long hardware tests. This ensures those expensive runs happen exactly once per merge, not repeatedly for every PR update. Batch merging further amplified this efficiency: multiple compatible pull requests are validated together in a single hardware CI cycle, dramatically improving throughput without requiring any additional machines.
Mergify finally gave us a merge queue that behaves predictably under heavy load. With limited hardware capacity and multi-hour tests, reliability is everything, and Mergify made stability a solved problem for us.
Rushin Shah
Sr. Manager, DevInfra and Automation Architect
The result is a merge process that finally matches Cerebras' scale and constraints. Mergify's Merge Queue controls load precisely, prevents main-branch breakages, and turns what used to be a fragile, delay-prone pipeline into a predictable, high-throughput system. Engineers no longer wait hours for merges to unblock, nor do they manually coordinate around hardware availability; the queue orchestrates everything. Cerebras now ships changes with confidence, without adding hardware capacity or compromising quality.
For a company pushing the boundaries of AI hardware, Mergify has become an essential part of keeping development velocity as ambitious as the technology itself.
