SREcon25 Europe/Middle East/Africa
2025-10-07..2025-10-09
Dublin, Ireland
who it's for
site reliability engineers, platform engineers, infrastructure engineers, ml engineers, cloud engineers, observability engineers, engineering managers
topics
MLOps platforms and runtimes; GPU cluster architecture and costs; LLM accelerator utilization tuning; AI platform observability and SLOs; Terraform caching with OverlayFS; auto-instrumentation with eBPF for GPUs; TCP congestion control tuning (Cubic, BBR); zero-regression network CI/CD; high-scale build and cache systems; active-active failover and disaster exercises
| title | author | length |
|---|---|---|
| MLOps 2025: A Journey into the Past and the Future | Alejandro Saucedo | 42m |
History traces platforms like Michelangelo, SageMaker, TFX and runtimes like Flink, Ray, Metaflow while orchestration and experiment tools such as Kubeflow… | ||
| Challenges of Making Large AI Clusters Reliable | John P. Looney, Panos Christeas | 46m |
Crusoe.ai runs vertically integrated AI clusters and describes concrete limits: a single HGX server costs about $700,000 machine with $45k GPUs and a $118k… | ||
| Why Risk Management Requires Taking Risks: A Practical Guide... | Aravind S Kumar, Scott Johnson | 35m |
Eighteen months of production work on live game streaming (150+ services, 25k+ servers, ~100k GPUs) produces concrete patterns: deploy a RAG server for… | ||
| Securing AI at Scale: Practical Defenses against Prompt... | Derek Chamorro | 39m |
Cloud-scale defenses include deployable middleware that runs semantic filtering (embedding-based intent checks) at API gateways, sandboxed runtimes for… | ||
| ACK to the Future: TCP in 2025 | Philip Rowlands | 23m |
Covers concrete tuning: shows Cubic collapsing under 0.05% random loss while BBR tolerates loss and probes RTTs, notes BBR v3 out of tree versus mainline… | ||
| IPv6 in 2025: Quo Vadis? | Alexandros Kosiaris | 36m |
IPv6 adoption sits around 40% globally in late 2024 with a linear projection to 100% by 2045, but real-world signals complicate that math: Google shows 49.84%… | ||
| The Bitter and the Sweet of Running a Planet-Scale Build &... | Tomasz Koczorowski | 30m |
Runs a centralized Build & CI stack for 100,000+ engineers, sustaining 98% cache hit rate and handling hundreds of petabytes/day via Rabin chunk deduplication… | ||
| Honey, I Shrunk the Data Center! | Przemysław Lalak | 22m |
Allegro runs a five‑year program called Oziris that proves losing a whole data center is survivable: teams perform controlled shutdowns, since 2023 they… | ||
| Dashboards & Dragons: Reliability Magic for AI Platforms | Alexa Nicole Griffith, Sal Furino | 31m |
Bloomberg describes a multi-cluster GenAI platform instrumented with SLOs that target token streaming: an example SLO is time to first token (TTFT) ≤ 500 ms at… | ||
| Resilience for AI Workloads at Scale: The Fast and the Finicky! | Lerna Ekmekcioglu | 25m |
Claims that NIC and link flaps (optical transceivers overheating/dust) make RDMA fabrics crash large LLM jobs and waste checkpoint time; demos use a software… | ||
| Fifty Shades of Caching and How LLMs Paint It Blαck | Effie Mouzeli | 31m |
Industry-grade cache stack uses HAProxy, Varnish (memory), Apache Traffic Server (disk), PHP opcache/APCU, memcached behind mcrouter, and CDN edges to reduce… | ||
| The Computer Wants to Lose Your Data | Chris Sinjakli | 34m |
Databases must cope with torn 16KB (MySQL) or 8KB (Postgres) pages versus 4KB atomic sectors, so MySQL uses a doublewrite buffer that guarantees one intact… | ||
| Gaining Insights from a Black Box System | Thiara Ortiz | 29m |
Three concise case studies show how layered observability actually buys time: a popularity-prediction black box produces content shed alerts that surface… | ||
| Utilization Is the Key to Efficiency: What It Takes to Run... | Aleksei Semiglazov | 35m |
Cloudflare runs almost 100 models on GPUs spread across 200+ cities and tackles low utilization with three core components: an edge coordinator, permit-based… | ||
| Experimenting with AI-Driven Systems | Jay Lees, Javier Martin Montull | 31m |
Ad systems run roughly ~1,000 concurrent experiments by thousands of engineers, so production is constantly changing; a new safety pipeline adds offline… | ||
| Catch Me If You Can: Hunting Misconfigurations Before They... | Jagadeesh Devaraj, Marcel Punselie | 27m |
ING builds a scalable "drift radar system" that detects misconfigurations across CI/CD and runtime, ingesting sensor events via Kafka and evaluating them with… | ||
| What I Wish I Knew Before Choosing Spot Instances | Lasse Canth Hels | 27m |
Runs Loki, Grafana, Tempo and Mimir on Kubernetes as 8,000 pods across 330 nodes and saves 60–90% on compute by using spot instances, but evictions are… | ||
| Making Reproducible Builds Faster with Docker Bake | Liz Fong Jones | 21m |
Honeycomb replaces multi-step workspace copying with Docker Bake plus BuildKit to keep outputs on one worker and chain targets, eliminating writing ~3 GB per… | ||
| Speeding Up Terraform Caching with OverlayFS | Ricard Bejarano | 15m |
Claim: switching Terraform provider caching to OverlayFS plus a central readonly store and per-process upper layers makes concurrent inits safe and cuts plan… | ||
| Service Messy: Pagerduty's Service Mesh Migration | Liz Frost | 36m |
Pagerduty migrates roughly 500 services to a service mesh using an away team model that makes initial code changes and asks teams to review; after a year… | ||
| HyperRouter: Lessons Learnt from Building an L4 Load... | Linhua Tang, Jayaganesh Kalyanasundaram | 34m |
Describes building a Level 4 software load balancer that uses DPDK over eBPF/XDP for a user‑space kernel bypass data plane to hit about 10M pps per CPU core… | ||
| Leader Election: Pitfalls and Alternatives | Andrew Medworth | 29m |
Leader election requires explicit fencing not just consensus: a lost leader can keep acting if downstream systems accept stale writes. Google Cloud Router’s… | ||
| Lessons from 25 Years in Infra: What’s Changed, What Hasn’t,... | Michal Gryko | 23m |
25 years of infrastructure evolution condenses into concrete lessons: hand-managed servers with ~10GB disks and 64MB RAM give way to CI/CD and config tools… | ||
| Level Up Your Edge: Do You Know Where Your Traffic Comes From? | Radha Kumari | 37m |
Slack implements an Authentication Service at the L7 edge load balancers using L7 edge external auth via Envoy x-oz filter that sends headers and portions of… | ||
| Auto-Instrumentation for GPU Performance using eBPF | Nikola Grcevski | 21m |
eBPF hooks into CUDA APIs to record kernel launches, grid sizes and host↔device memory copies and export them as Prometheus metrics tied back to Kubernetes… | ||
| The Chicken or the Egg: Creating a Stable Preprod... | Christine Pukropski | 39m |
Built a full, isolated production replica called TestCloud using a two-phase double bootstrap: two bare metal management hosts bootstrapped with Ansible and… | ||
| ARM Migration Made Practical | Nati Cohen | 34m |
Explains concrete migration steps: scan repos with Porting Advisor for Graviton (and watch Migration Ease), run realistic dollar‑per‑request performance tests… | ||
| From Vibes to Outages: Riding the AI Code Wave | Sylvain Kalache | 32m |
AI-generated code already makes up a large share of commits (Cursor claims one billion lines accepted a day; Stack Overflow reports 84% adoption), and teams… | ||
| How We Used Statistics to Find Toil among 36,000 Changes | Dylan Ratcliffe | 19m |
Dataset of ~36,000 Terraform plans (saved as Terraform plans JSON) gets analyzed to flag routine changes, claiming 12–21% of PRs could bypass review. The… | ||
| Training New Incident Commanders: Pokemon Style! | Laura de Vesine | 17m |
Company builds a nontechnical, Pokemon-flavored incident command exercise: a 45-minute lecture plus a 45-minute practice using a made-up Charizard scenario… | ||
| Run, Walk, Crawl, or How We Failed Our Way to SLO Readiness | Rob Durst | 14m |
A hypergrowth startup repeatedly fails at SLO adoption because early SLOs use platform APM metrics that aren’t tied to user experience, there’s no standard or… | ||
| Taming the Cost of Telemetry: How Riot Games Reined In... | Maxfield F Stewart | 32m |
Riot ingests ~3.9 petabytes/month into telemetry and finds ~40–60% unused; migration to DataDog agents + Vector collectors plus team rules (metric graduation… | ||
| Is My Thing Reliable? Adventures in Driving Consistency of... | Dominic Hutton | 35m |
An org builds an operational readiness checklist as a spreadsheet, runs manual reviews that take ~24 months to cover everything, then automates five checklist… | ||
| Tales from the Platform Adoption Mines | Dave O'Connor | 29m |
Platform teams focus on building but miss adoption: start with shiny pebbles (discoverable docs, examples and onboarding that set norms), formalise decisions… | ||
| From 4 Hours to 8 Minutes with AI Agents That Transform SRE... | Peter Jausovec | 44m |
Claims a workflow that cuts a typical SRE investigation from four hours to eight minutes by chaining specialized agents. It builds three concrete agents —… | ||
| Modernizing Incident Response with LLMs, RAG, and the MCP | Theofilos Papapanagiotou | 36m |
A large SRE team replaces fragmented scripts with an LLM-driven workflow that uses the Model Context Protocol (MCP) to inject real-time logs, configs and… | ||
| The Un-Incident: Extracting Value from the Gray Area of... | Andreas Deuschl | 24m |
Claims that 30–60% untracked potential incidents hide learnings and defines four concrete gray‑area patterns: no‑CI (non‑critical), near miss, not‑our‑fault… | ||
| Learning Inside the Box: The Scope of Incident Reviews | Rachel Silber | 17m |
Incident reviews focus on concrete scope: start timelines before the declaration, include planning and hallway decisions, and map which systems and orgs… | ||
| Embedding Platform SREs: A Hybrid Model for Driving Adoption | Jorge Lainfiesta | 15m |
Argues that a hybrid Platform SRE model—build self‑service reliability into the platform for new services and use time‑boxed embedded SREs to lead… | ||
| Lessons from an Asset Manager’s First Embedded SRE | Callum Donald | 28m |
A 12‑month embedded SRE trial inside a trading team selects six critical services, uses Grafana and queue‑depth metrics, catalogs alerts in a wiki table, cuts… | ||
| Practical DORA (Digital Operational Resilience Act) for SREs | Laura Woo | 24m |
EU DORA takes effect Jan 2025 and covers financial entities and their critical third‑party providers; SREs must identify critical/important functions, run… | ||
| CPU Utilization: The Hidden Cost of Running Hot | Andreas Strikos | 35m |
Claims that CPU utilization predicts latency and capacity with a simple practical formula, validated on production-like load using a dedicated performance… | ||
| Lessons from a Year with Backstage: What Worked, What Didn’t... | Nick Mckenzie | 27m |
Accidental adoption of Backstage scales to 40 teams and 350+ developers by shipping an end-to-end slice in a week then a four-week build and two-month… | ||
| Quality Gates in Production: How We Turn OpenTelemetry... | Marcel Birkner | 21m |
Describes an automated pipeline that prevents unsafe rollouts by gating production deployments on real production-like checks: OpenTelemetry traces logs… | ||
| Maximizing Utilization for LLM Accelerators | John Lunney | 17m |
Measures utilization at the 95th-percentile utilization across a week and breaks losses into five concrete categories: a model-dependent ceiling… | ||
| Finding the Needle in a Haystack: Tenant-Level Network... | Giscard Fernandes Faria | 29m |
Cloud-scale tool traces tenant flows end-to-end by deploying agents on bare metal and virtual switches, mirroring headers and collecting data in a central… | ||
| Failover Without Fear: A Practical Guide to Active-Active... | Sina Moghaddas | 27m |
Mollie migrates a single-region platform to multi-region on Google Cloud, rewriting application code to be region-aware, moving from the classic to an advanced… | ||
| Defending the Global Financial Backbone: Euroclear's... | Haytham Elkhoja, Erik Van Nooten | 45m |
Euroclear claims custody of €40 trillion assets and processes roughly 300 million transactions a year, running the equivalent of Belgium’s GDP through systems… | ||
| The SRE’s Crystal Ball: Predicting System Performance with... | Aravindh Sampathkumar | 38m |
Uses Little's law and the Universal Scalability Law to move from firefighting to capacity predictions, with concrete numbers: 180 rps, 50 ms service time, 4… | ||
| Profiling Your Code: 5 Tips to Significantly Boost Performance | Sonam Gupta, Kunal Nawale | 19m |
Adds a Go pprof endpoint to collect CPU, heap, block and contention profiles from production workloads, runs profiles on real traffic, and iterates: find the… | ||
| Performance Consistency in the Cloud | Nati Cohen, Sercan Karaoğlu | 21m |
Cloud performance jitter affects P99/P99.9 more than averages and can cost millions—one case cuts tail latency by 100 microseconds and recovers revenue. Root… | ||
| STPA for Software Systems–Illuminate the Unknown Unknowns | Theo Klein, Garrett Holthaus, Ruben Barroso | 01h 23m |
Claim: STPA (systems theoretic process analysis) anticipates design flaws before production by modeling systems as control-feedback loops. Applied to a Google… | ||
| Reliability in Finance: Delivering Quality Market Data at Scale | Alan Campbell | 32m |
Describes a production pipeline that adds checkpointed visibility, automated milestones and lineage so files are validated on arrival, processed through… | ||
| From Toil to Empowerment: Building Self-Service Ingress with... | Jeswin Koshy Ninan | 16m |
Claims a GitOps pipeline that turns fragmented, manual ingress (load balancers, TLS, DNS, edge rules) into zero-touch ingress provisioning across many clusters… | ||
| Systematising Production Operations: Reddit Control Tower | Kelly Dodd | 15m |
Built a self-service platform called Reddit Control Tower that turns manual prod steps into auditable admin APIs; components are CLI + admission server +… | ||
| Cross-Platform Data Lineage with OpenLineage, A Foundational... | Maciej Obuchowski | 30m |
Shows a concrete JSON‑schema standard with Java and Python libs and active integrations for Airflow, Spark, dbt, Flink, Trino and Snowflake; emits run events… | ||
| Zero-Regression Network CI/CD for Finance-Grade Reliability | David Ferrandez, Damian Krogul | 37m |
Claims elimination of post-deployment network issues by running a full virtual replica of production with the same playbooks, daily CI/CD runs, and failure… | ||
| What Do We Do Now, Now That We’re Happy? | Niall Murphy | 38m |
Claims that AI task capability is doubling every seven months and that agents recently solved six of seven IMO problems frame a push to apply models to SRE… | ||
| Lightning Talk - Parquet: The Pillar of Modern Telemetry... | Nitish Tiwari | 02m |
Parquet is a binary columnar storage format that groups a few hundred thousand rows into row groups, stores each field physically together as Parquet columnar… | ||
| Lightning Talk - AI is Like Chocolate | Liz Fong Jones | 05m |
Claim: AI is useful but not universal — it works in many pipelines but breaks when shoehorned everywhere. Recommendations: always disclose training data use… | ||
| The Executive Signal: How to Cut Through... | Eric D. Schabell | 05m |
Claims a short, repeatable signal cuts through executive noise: expect five minutes of attention, state the core four points immediately and repeat them to… | ||
| Computers Were a Mistake, We Should Burn... | Laura de Vesine | 05m |
Computers are full of concrete, reproducible failures and quirky fixes: the HTTP header typo Referer is unchangeable; the left-pad removal in March 2016 is a… | ||
| nothing matches | ||