@Scale: Systems & Reliability 2026
| title | author | length |
|---|---|---|
| Opening Keynote | Surupa Biswas | 25m |
Meta reports over 90% of day-to-day infra changes now include AI-authored code and cites roughly $700B of hyperscaler CapEx announced, driving supply shortages… | ||
| Our Journey to Safely Unleash Agents at Meta Scale | David Pariag | 17m |
Meta provisions X.509 cert per agent to differentiate agent from human identity, uses an LLM risk scoring of RPCs and CLIs to tag dangerous operations, and… | ||
| The Agentic Infrastructure Gap: In-Distribution Languages Make It a Coding Problem | Joe Duffy | 23m |
Claim: infrastructure fails because cloud consoles, API keys and copy‑pasted YAML are out of distribution for LLMs, so the fix is to express infra as real code… | ||
| Live Q&A | Joe Romano, David Pariag, Sargun Dhillon, Joe Duffy | 19m |
Discussion centers on concrete guardrails and failures: agents delete production databases and scale clusters to zero, one incident 20 minutes to break, 1 day… | ||
| Building Privacy Aware Infrastructure in the AI-Native Era | Rituraj Kirti | 19m |
An AI-native asset classifier combines a hybrid deterministic rules + LLM fallback that cuts the set of nodes needing LLM review from a 100,000-node example… | ||
| Securing Production Debugging at Hyperscale | Shridivya Sharma, Luke Kelly | 15m |
Microsoft debugs a global K8s platform called Cosmic that runs 12,000 clusters, tens of millions of pods and nearly 6 million cores by issuing short-lived SSH… | ||
| Taming AI Infrastructure Failures with Agentic Debugging | Phillip Liu | 20m |
NCCL watchdog timeouts cause massive pain in distributed PyTorch training — at ~2,000 GPUs job success dips below 2% — and Meta finds >99% of timeouts are… | ||
| AI Storage Blueprint | Sidharth Bajaj, Venkatraghavan Srinivasan | 21m |
Claims concrete changes: collapsing distributed metadata into a single flat schema yields an order-of-one lookup instead of cross-region, multi-step… | ||
| Live Panel with | 30m | |
AI agents automate incident ops by ingesting onboarding troubleshooting docs and extracted query semantics to rebuild effective DB queries, and by checking… | ||
| [Almost] Fail & Tell: Stop The World | Phil Lopreiato, Rahul Iyengar | 20m |
Meta avoids a major site outage by containing impact to one region and restoring service within about 15–20 minutes. Recovery re-bootstraps the regional… | ||
| Extending Meta to the Public Cloud | Dinesh Govindasamy, Sargun Dhillon | 21m |
Meta moves GPU demand to public clouds in under three months, enabling first capacity in ~60 days using the offnet services boot-and-attest flow and… | ||
| Teaching AI to Fight Fires: Autonomous Reliability Agents at Meta Scale | Gaurav Mitra | 19m |
A reliability flywheel encodes expert workflows into an always‑on copilot that has investigated 1,000+ incidents at Meta scale and cuts detection‑to‑mitigation… | ||
| Experience Productionizing and Operating GB200 Cluster | Sachin Kumar Lakharia, Douglas Wightman, Ankur Srivastava | 21m |
Shows that multi-node NVLink requires scheduler changes: the team deploys a Slurm topology block plugin to treat each NVL72 domain as a rigid unit, then uses a… | ||
| Closing Keynote | Peter Hoose | 24m |
Claims that the industry spent ~$450B on AI in 2025 and ~$765B in 2026 back a year of concrete engineering: from prototypes of compute tents serving hundreds… | ||
| nothing matches | ||