SREcon23 Europe/Middle East/Africa
2023-10-10..2023-10-12
Dublin, Ireland
who it's for
site reliability engineers, platform engineers, observability engineers, incident responders, SRE managers, cloud infrastructure engineers, DevOps engineers
topics
eBPF tracing and enforcement; Pixie cluster aggregation; Cilium Hubble flow visibility; Kubernetes service networking; HTTP/3 and QUIC debugging; Prometheus native histograms; OpenTelemetry collectors vs Prometheus; continuous profiling in cloud-native; LLM training and data lineage; SLO design and adoption; DNS internals and troubleshooting; TLS cryptography essentials
| title | author | length |
|---|---|---|
| eBPF Superpowers for SRE | Liz Rice | 43m |
eBPF lets operators load verified programs into the Linux kernel to observe or change events at syscall and networking hooks; the verifier prevents typical… | ||
| Building a 5-Exaflop Supercomputer for Meta-AI Research and... | Kalyan Saladi, Chris Bray | 35m |
A single cluster delivers 5 exaflops of aggregate compute with 16,000 GPUs across three halls, a non‑oversubscribed InfiniBand fabric, roughly 500 PB of… | ||
| From Sysadmins to (almost) Flying Unicorns | Guillaume Hérail, Gilberto Müller | 29m |
Sony’s PlayStation cloud-streaming team reduces deployment time from months to minutes by adopting Kubernetes and creating cross-functional CFTs, implements… | ||
| Implementing SRE in a Telco with Reliability Enhancing... | Florian Kammermann, Romain Bonjour | 37m |
A telco that suffered major outages in 2021 (including a 6 hour emergency‑call failure and a recent 40 minute internet outage) builds nine Reliability… | ||
| Symptom-based Alerting for Machine Learning - What I Learned... | Lina Weichbrodt | 25m |
More than 30 ML use cases and roughly 30 production models show that silent failures (for example an input unit change from seconds to milliseconds) break… | ||
| Reliable Data for Large ML Models: Principles and Practices | Mary McGlohon | 31m |
LLM training at pretraining scale multiplies failure modes: corrupted or unlabeled corpus can waste massive GPU weeks and poison every downstream fine-tune… | ||
| New Grads Becoming New SREs: Catalyzing a “Circle of Life”... | Jennifer Petoff, Catalina Rete | 35m |
Program builds a campus-to-SRE pipeline using open-house events, algorithms workshops, CV coaching, sponsored coding competitions and alumni volunteers; it… | ||
| Scale Your Future: An Immersive Engineering Programme | Radha Kumari, Margarita Glushkova, Berkeli Halmyradov | 30m |
Slack partners with Code Your Future to run CF Plus, a 1‑month self study then 3‑month intensive program and 9‑month internship that teaches SRE skills with… | ||
| Over, Under, Around, and Through: A Detailed Comparison... | Lucas Pardue | 38m |
Claims that QUIC is "just TCP" or "just HTTP over UDP" are false; QUIC is a transport with multiplexed streams, connection IDs and an optional datagram… | ||
| Deploying and Debugging HTTP/3 | Robin Marx | 43m |
Explains concrete failure modes: TLS session-resumption tokens cause a contagion that forces Google to disable QUIC after 10% of fleet fails; connection… | ||
| The Engineer/Manager Pendulum Goes Mainstream | Charity Majors | 33m |
Argues that switching between engineering and management is now common and urges companies to build policies to support it, with concrete rules: engineer first… | ||
| Do Not Thrash the Node.js Event Loop | Matteo Collina | 30m |
Node.js apps self‑DoS when synchronous work blocks the event loop, forcing huge overprovisioning. Concrete fixes: measure with the eventLoopUtilization API (or… | ||
| Scaling Chef Emotionally | Brett Pemberton | 33m |
Slack runs Chef on a single stateful backend then migrates to AWS managed services (Aurora, managed Elasticsearch, NLB) and stateless EC2 frontends to stop… | ||
| SRE for [cyber]security | Nicolas Fischbach | 35m |
Describes running low-level and customer-facing security at hyperscale, claiming mitigation of a DDoS at 47 million requests per second and another event near… | ||
| Cloud, Kubernetes, and Service Networking - Taming the Turtles | Matt Turner | 40m |
Explains how cloud VPCs present a flat Layer‑2‑like space while Kubernetes enforces a separate pod IP space, how CNI plugins (Calico, Cilium, AWS ENI)… | ||
| Designing Matrix: A Global Decentralised End-to-End Encrypted.. | Matthew Hodgson | 42m |
Matrix describes an open standard and ecosystem that now spans 106M users across ~100K servers, claims active replication of full room history like git, and… | ||
| Tracing the Journey into Distributed Tracing | Pedro Alves | 28m |
Large-scale org rollout of distributed tracing targets concrete gaps: causality, discoverability and alert noise. Team runs a proof-of-concept, chooses a… | ||
| The World Blew Up but We’re All Okay: How We Managed a... | Laura de Vesine, Laurent Bernaille | 40m |
Datadog traces a 10+ hour global outage to a systemd change that makes systemd Networkd restart flushing routing rules, causing pod networking to drop during a… | ||
| When One Line Took Thousands of Websites Offline | Francisco Borges Aurindo Barros, Jack Henschel | 27m |
A single line change in the GitLab CI deployment disables a CRD version and triggers an auth operator pointed at the staging API to delete resources, wiping… | ||
| HTTP Headers that Make Your Website Go Faster | Thijs Feryn | 38m |
Shows how to use HTTP headers with Varnish Cache to cut origin load: claims 1.3 Tbps on one machine and 1.17 Gbps per watt as performance/efficiency proof… | ||
| Cache Me If You Can: How Grafana Labs Scaled Up Their... | Danny Kopping | 26m |
Grafana Loki adds large local SSD-backed memcached (mcached with extstore) clusters—33 nodes with four 375GB local SSDs each—to reach ≈50 TB memcached cache… | ||
| Embracing the Multi-Party Dilemma: Incident Response Across... | Sarah Butt, Alex Elman | 37m |
Over a dozen deep case analyses from 2021 use thematic incident analysis (Adaptive Capacity Labs method) to show a repeatable pattern: information and impact… | ||
| The Incident Is The Way: Using Your Incidents to Win... | Niall McCarthy | 13m |
Incidents are reframed as concrete investment moments by four practices: map system capabilities (group 24 microservices into user-facing capabilities like… | ||
| That Time I Accidentally DDoS'd My Company | Mike Bongardino | 14m |
A homegrown Chrome extension polling health-check URL runs a GET every three minutes but the Chrome API actually schedules independent, persistent checks on… | ||
| Artificial Intelligence: How Much Will It Cost You? | Todd Underwood | 38m |
Generative models require huge up-front capital: accelerator hardware costs dominate (H100-class cards cost tens to hundreds of thousands each) and training a… | ||
| Just the Cryptography You Need to Know for TLS | Lerna Ekmekcioglu | 14m |
Explains TLS 1.3 mechanics: client sends a public key share in ClientHello so the handshake is one RTT, they perform ECDHE using X25519 / 256-bit to get… | ||
| You Depend on DNS, This Is How It Works and You Won't... | Philip Rowlands | 15m |
Live TCPdump demo shows a single hostname resolution produces six lookups (A/AAAA plus search path) and that recursion plus caching is why DNS scales. Cache… | ||
| 9 Things You Should Do When Starting to Use SLOs | Sal Furino | 32m |
Nine practical steps: pick user-facing SLIs, avoid pre-aggregation, and measure success rates not just errors; use raw event-level data so you can spot… | ||
| Silent Spring: What if the GDPR Was Real? | John P. Looney | 39m |
Predicts EU law enforcement suddenly applied to tech: concrete fixes proposed include store encryption keys in EU and rehome EU user data to EU data centers… | ||
| From Exceptional Maintenance to Automated Routine Operation:... | Giuseppe Lavagetto | 37m |
A single core datacenter migration becomes routine by replacing 15 coordinated code commits and 45 minutes of read‑only downtime with an etcd state store plus… | ||
| Should I Use OTel (collectors), or Is Prometheus Good Enough? | Krisztian Fekete | 37m |
Replaces homegrown scraping agents with an OTel collector cluster that runs alongside workloads and a dedicated management cluster, enabling exporting to… | ||
| Implementing Open-source Observability within Maersk | Charlie Nederhoed, Martin Jaeger | 17m |
Maersk standardizes observability with a 100% self-service platform built on the LGTM stack (Grafana, Mimir, Loki, Tempo) plus in-house tooling and a… | ||
| Journey from Fluent Bit, Fluentd and Prometheus to Open... | Marcin Stozek | 15m |
Replaces Fluentd, Fluent Bit and Prometheus with OpenTelemetry Collector and a vendor-built OTEL distribution, unifying logs, metrics and traces. Fluentd… | ||
| Level 7 Egress Control in Kubernetes: Current Solutions,... | Joshua Fox | 36m |
Problem: allowing egress to a specific FQDN without opening IP ranges. Demonstrates workable options and limits: GKE FQDN-aware NetworkPolicy (data plane V2)… | ||
| Leveraging Unikernels and Kubernetes to (Transparently)... | Felipe Huici | 15m |
Unikraft builds specialized unikernels that run unmodified apps with dramatically smaller footprints and faster starts: 7–14 ms cold boot and 1.2–1.9 MB images… | ||
| Monoceros: Faster and Predictable Services through In-pod... | Sotiris Nanopoulos, Ben Kochie | 15m |
Built a Go multiprocess controller called Monoceros that replaces Ruby Einhorn, initially using SO_REUSEPORT then running a proxied data path with Envoy… | ||
| Continuous Profiling in the Cloud-Native era | Matthias Loibl | 36m |
Parca collects pprof profiles from Go and other runtimes and also uses an eBPF-based Parca agent that samples stacks about 19 samples/sec in-kernel and reads… | ||
| How to Use Prometheus's Native Histograms | Björn Rabenstein | 36m |
Prometheus adds a new sample type: timestamped histograms stored as one series with dynamic exponential buckets that disappear when empty, enabling affordable… | ||
| Overcoming Challenges in Serving Large Language Model | Theofilos Papapanagiotou | 31m |
Shows how to host GPT-style models on Kubernetes with multi-GPU nodes, demonstrating a live KServe demo of a 7B Bloom model and concrete trade-offs: a 6 GB… | ||
| The Value of Reliability | Niall Murphy | 33m |
Argues reliability value isn’t just uptime seconds but measurable business impact: proposes stack labeling (attribute revenue to call-stack-like components)… | ||
| A Dual Approach to Accountability Engineering | Anthony Sandoval | 22m |
Presents a concrete decision model called accountability engineering that bifurcates projects into direct (high control, visible wins) and diffuse (long‑game… | ||
| Succeeding as the Lone SRE in a Small Team | Danny Kirchmeier | 13m |
Argues that small SREs (5–50 engineers) succeed by practicing extreme ownership: taking responsibility for failures, adding guard clauses (the broken script… | ||
| Deconstructing an Abstraction to Reconstruct an Outage | Chris Sinjakli | 33m |
At 15:09 a RAID controller ejects disks and the primary goes down; engineers reproduce a three-node Postgres cluster in Docker Compose, inject faults with… | ||
| When Clouds Stop Raining Discounts: Surviving the Drought | Max Blaze | 17m |
Migrated major services to Graviton and initially cut costs by 45–50%, then AWS Spot discounts collapse—from roughly 60% to near 0% between March and April… | ||
| Should an SRE Care About FinOps? Using Observability to... | Rodrigo Serra | 13m |
Company runs FinOps POV on AWS with Kubernetes using Kubecost and other open source tools and reports 35% cost savings across clusters by applying specific… | ||
| Looking at SRE Needs and Trends over Two Decades with a... | Salim Virji, Murali Suriar | 35m |
Chubby starts as a leader election service for GFS and becomes a general-purpose small-key storage used for service discovery, configs and ACLs. Implementation… | ||
| How to Make Your Automation a Better Team Player | Laura Nolan | 35m |
Argues that automation changes the whole socio‑technical system and often causes automation surprise—examples include Argo CD irreversibly deleting a site and… | ||
| Dark Matter and Deep State: The Unseen Majority of Everything | Heidi Waterhouse | 38m |
Argues that the infrastructure people never notice are the consequential majority: mundane rules that make the conference stairs safe, the Apollo 13… | ||
| nothing matches | ||