Merge Strategies at Scale: Developer Happiness Through Better Code-submission Workflows
summary
Argues that simple merge queue validation fixes some breakage but scales poorly — 100 PRs/day with sequential revalidation can require ~50 hours of CI, whereas batching (batch size 4) cuts CI runs roughly to 25 and best-case merges to ~12.5 hours. Proposes optimistic parallel queues that run many validations in parallel (best-case <1 hour, ~6 hours with 10% failures) at the cost of more CI (example ~150 runs). Adds predictive scoring (linear regression on LOC, files changed, tests, deps; cutoff ~0.5) to skip unlikely chains. Describes implemented tools: aviator stack sync, av stack merge, flaky-test reruns and atomic multi-repo PR merges; notes some advanced reorder/fail-fast optimizations remain unimplemented.
AI-written summary · about
chapters (11)
- 0:00Introduction and background
- 3:05Why merging is hard at scale
- 7:21Simple merge queue concept
- 9:27Batching for efficiency
- 11:12Parallel universes for merging
- 14:22Optimistic batching flow
- 15:26Multiple alternate realities
- 17:56Multi-queue optimizations
- 21:05Merge strategies in poly repos
- 23:52Aviator’s workflow automation
- 25:18Q&A session
related talks
14mMerge Queue
42m7 Metrics to Accelerate Delivery without Damaging Culture
29mBreaking the PR Review Bottleneck
56mAsync Code Reviews Are Choking Your Throughput
42m30 to 70 PRs a day: How we managed to not wreck our systems
10mBeyond “try harder”: Effective strategies to tackle bugs
20mFireside Chat
43mEffective CI/CD for Large Systems