At QCon London 2026, Netflix engineers Prasanna Vijayanathan and Renzo Sanchez-Silva presented E2EGraph, a proposed knowledge graph for connecting user experience data with client applications, services, networks, and infrastructure. Its central idea is to give those systems a shared semantic model, so engineers can trace a user-visible regression across boundaries instead of piecing together disconnected metrics, traces, and logs. The presentation also described AutoSRE and automated remediation as planned directions—not as proven production capabilities.
Why observability breaks at system boundaries
A streaming experience depends on a chain of components: a client application, network connections, backend services, and supporting infrastructure. Each part can produce telemetry, but those signals may use different formats, names, and operational tools. When a user-visible regression appears, the practical question is whether it began in the client, the network, or a backend dependency—and finding the answer can require correlating evidence across those boundaries.
As an Amazon Associate I earn from qualifying purchases.
The QCon session framed this as a data integration and semantics problem, not simply a shortage of telemetry. Metrics, traces, and logs can each describe part of an event without sharing enough context to explain how that event affected a particular experience.
Free tools Windows power users keep installed
One-click scans. No signup required.
What E2EGraph is designed to connect
E2EGraph represents a user experience as a graph. The session abstract describes entities such as user sessions, client applications, microservices, and network components as nodes. Requests, user interactions, and dependencies connect them; those relationships can carry operational and experience-related attributes.
#1 Best Overall
- Used Book in Good Condition
| Graph element | Examples in the presentation | Role |
|---|---|---|
| Nodes | User sessions, client applications, microservices, network components | Represent the entities involved in delivering an experience. |
| Edges | Requests, user interactions, dependencies | Represent how entities interact or rely on one another. |
| Attributes | Latency, error rate, quality of experience (QoE), version, geography | Add operational or contextual detail to graph entities and relationships. |
| Inputs | Client telemetry, server logs, traces, infrastructure metrics, experiments, deployments | Supply evidence from different parts of the system. |
The intended benefit is a connected view of the path behind an experience, rather than a collection of signals that engineers must manually align. The abstract does not specify how the graph is implemented, how large it is, or how quickly it can be queried.
Why an ontology matters
A graph can connect data without ensuring that different systems mean the same thing by the labels they use. E2EGraph’s ontology is intended to provide that shared meaning: a formal specification of types, properties, and relationships that lets heterogeneous sources be interpreted consistently.
The session names concepts including a user session, API call, deployment event, experimentation, and QoE regression. In its March 18, 2026 event report, InfoQ described graph facts as subject-predicate-object triples. Its examples linked an API gateway to an application type and an owner, and an incident to an affected gateway. The same report said the presentation described 12 operational namespaces, including Slack, alerts, metrics, logs, incidents, E2E, and Harvest. These are conference-reported presentation details, not an independently verified production schema.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteIn practical terms, the ontology is the layer that can make an API call, a deployment, and a QoE regression recognizable as related operational concepts—not just records with fields that happen to resemble one another. That shared vocabulary is what would allow queries and analysis to cross domain boundaries more consistently.
How graph snapshots could help investigate regressions
The abstract describes creating graph snapshots around regression events, then comparing healthy and degraded states over time. That comparison could help an engineer follow the affected experience through connected components and inspect changes in context—for example, whether a client version, an experiment, or a dependency changed near the onset of a regression.
The session’s example operational query was: “Why is TV UI lolomo TTR regressing in the latest version?” The example illustrates the kind of question the system is meant to support; the public session material does not define those domain-specific terms or report a measured answer to that query.
A graph-backed investigation is useful only if its relationships and evidence remain inspectable. The presentation’s stated design suggests a way to organize that evidence, but the available public material does not document an evaluation method, diagnosis accuracy, or operational outcomes.
AutoSRE is described as a planned analysis layer
The proposed AutoSRE architecture sits above the graph. A coordinator agent would split an operational question into tasks, specialized agents would query graph-connected areas such as metrics, alerts, experiments, client platforms, events, and deployments, and the coordinator would synthesize the returned evidence into likely root causes.
For example, when asked why a TV interface regression is occurring, the system might investigate a client rollout, a misconfigured experiment, or a backend dependency regression. These are possible causes in the presentation’s example, not findings from a reported production diagnosis.
InfoQ also reported a Netflix incident that took four hours from initial alert to resolution. According to that March 18, 2026 report, nine teams and more than 30 engineers worked on that incident and three related incidents. The public material cited here does not include an independently linked Netflix incident account verifying those figures, so they should be understood as InfoQ’s reporting rather than as independently confirmed operating metrics.
What the roadmap would add—and what remains unestablished
The presentation outlined future ambitions rather than documented results. It described using patterns in failing subgraphs, propagation paths, and risky combinations of versions and experiments to predict problems. Possible mitigations included targeted rollback, traffic shifting, feature-flag changes, and capacity adjustment. InfoQ characterized the wider aims as automating root-cause analysis, enabling auto-remediation, and building self-healing infrastructure.
Recommended Free Tools
That distinction matters: proposing an agent workflow or mitigation is not evidence that it diagnoses incidents accurately or can safely act without human approval. The public session sources do not establish whether AutoSRE is deployed, how automated any response is, or what safeguards or approval steps would govern a production change.
QCon’s session page says slides are unavailable. The public sources described here do not provide the implementation architecture, graph size, query latency, accuracy, evaluation method, adoption level, or measured outcomes for E2EGraph. Those details therefore remain unresolved; the presentation supports understanding the design direction, not judging its production performance.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
