A request at the edge
Inspect the country, serving location and protocol Cloudflare reports for this request.
Explore experiment ↗Miles Whittaker / Engineering labs
Systems you can
inspect.
Small experiments in delivery, signals and failure. Follow a path, inspect the evidence, and see where its limits are.
Worker experiment · request on demandSynthetic model · fixture dataPlanned · evidence to come
Measured architecture
This site, from the inside.
Walk the source. Follow a request. See how static pages and edge experiments fit together.
20 modules4 real flowsMeasured source
Explore the architecture ↗Pick an experiment.
Jump to Worker labs ↓10 labs
Watch a response arrive
Receive six bounded frames. Compare arrival times, then cancel the next run.
Explore experiment ↗Checks with a history
Read scheduled asset checks with timestamps, failures and missing observations.
Explore experiment ↗Two windows, one sequence
Send a pulse and refresh another window to inspect shared, serialized state.
Explore experiment ↗Internet protocol mix
Compare country-level HTTP versions using dated Cloudflare Radar observations.
Explore experiment ↗An incident, a second opinion
Ask a bounded model about a fixed scenario. Keep facts separate from suggestions.
Explore experiment ↗Follow a platform change
Trace configuration, a request or a failure through a generic platform model.
Explore experiment ↗One release, three signals
Compare a CPU profile, trace waterfall and release timeline for the same fixture.
Explore experiment ↗GCP delivery, end to end
A planned clean-room GKE platform: delivery, promotion, rollback and observable operation.
Explore the plan ↗AWS, reconciled from Git
A planned ephemeral control plane: prove readiness, export evidence and verify teardown.
Explore the plan ↗Cloudflare / Workers
Inspect the edge.
Each experiment makes a bounded request. Missing configuration and missing observations are shown explicitly; results are never invented.
01 / Workers · request.cf
A request at the edge
- Country
- —
- Serving edge
- —
- Protocol
- —
- TLS
- —
No request yet. Your IP address is not returned or stored.
This is an edge observation, not a traceroute or a precise location.
02 / Workers · ReadableStream
Watch a response arrive
123456
Six fixed frames. Arrival times are measured in your browser.
Buffering can change arrival timing. This is not a benchmark.
03 / Cron Triggers · D1
Checks with a history
No observations loaded.
Inspect readings
Five-minute asset-binding checks. Missing slots remain gaps. Not an external availability monitor or an SLO.
04 / Durable Objects · shared state
Two windows, one sequence
Open a second window, send a pulse here, then read the state there. No automatic polling.
A numbered counter, with no messages or visitor identity. The shared room accepts at most 200 pulses per minute.
05 / Cloudflare Radar · KV
Internet protocol mix
No Radar data loaded.
Cloudflare Radar ↗ · CC BY-NC 4.0
Seven-day HTTP-version shares within a country, transformed into bars. Not absolute traffic, packet paths or evidence of outages.
06 / Turnstile · Workers AI · D1
A second opinion
Fixed facts: latency rises after a release; readiness is green; the database span is slow. The cause is unconfirmed.
Runs only after verification and within a shared daily budget. The model cannot take actions.
A model suggestion is not a finding. Verify it against traces, profiles and a reproducible workload.
Observability / Synthetic model
Platform Atlas.
A deliberately generic, sanitised model of how platform changes and signals connect. Built as a teaching and onboarding aid; it does not depict my employer’s architecture or live systems.
01Entryintent → route
02Control planechange → release
03Servicesuseful work
04Signalobserve → learn
Intent → render → validate → release → observe
Evidence lab / 05
One release.
Three signals.
A small, public-safe fixture showing how the same deployment can be investigated through profiles, traces and a change timeline.
Continuous profiling / Go
Where did the CPU move?
baseliner2026.09.12
candidater2026.09.13
Select a frame to inspect its sample share. Relative shares do not establish higher absolute CPU use.
previous deploycurrent deploysample: 60 s CPU profile
OpenTelemetry / Tempo
Which span explains the wait?
GET /checkout
184 ms
auth middleware
18 ms
catalog service
74 ms
postgres query
112 ms
render response
21 ms
0 ms50100150200 ms
span durationinvestigate firsttrace id: 7f3a…c91e
Timings and limits
Start / duration in ms: request 0 / 184; auth 0 / 18; catalogue 18 / 74; database 51 / 112; rendering 163 / 21. Spans can overlap; their durations are not added together.
Release evidence / Git
What changed between runs?
r2026.09.12 baseline
Profile capturedsteady-state traffic · 60 s sampler2026.09.13 candidate
Template path changedprofile diff → trace check → investigate01build02profile03trace04decision
candidate releasecorrelation is a lead, not prooffixture data only
All evidence here is synthetic and illustrative — not taken from a live service or employer telemetry. This demo shows an investigation workflow, not a production incident.
Planned platform labs
Two clouds. Two control planes.
The same instrumented Go workload, through different delivery models. Both implementations are planned; no runs, costs or measured outcomes are claimed.
GCP / Planned
Delivery with deliberate promotion
An independently authored GKE reference implementation for repeatable delivery, IAM, promotion and rollback.
Git → CI → BuildKit / Kaniko↓Artifact Registry + Terraform↓Cloud Deploy → GKE → Go workload↓OTel → Collector / Alloy → evidence
What evidence is still needed?
- Reproducible provisioning and delivery with reviewed IAM boundaries.
- Promotion, rollback and controlled-failure evidence.
- Metrics, logs, traces and profiles tied to the same immutable revision.
- ADRs explaining the delivery model and its operational trade-offs.
AWS / KNR / Planned
Reconciliation, with a verified exit
An ephemeral Kubernetes-native AWS control plane. Inspired by polarsquad/knr-ops; to be implemented independently.
Git → disposable kind + Flux↓Kubernetes API → CAPI / CAPA + ACK↓Temporary EKS → Go workload + OTel↓Export evidence → destroy → audit AWS
What evidence is still needed?
- Git-to-ready timing with Flux and controller conditions.
- One controlled runtime failure and one reconciliation or teardown failure.
- An independent AWS audit of residual billable resources.
- Cost per run, environment lifetime and a versioned evidence manifest.
- OTel/Tempo/logs and a Pyroscope diff across two revisions, with profiling overhead documented.
Dormant is the expected state. Historical evidence remains available. A CAPI management-cluster pivot is a separate advanced experiment.
Presentation informed by Dan Hughes’s portfolio ↗. Experiments are independently implemented. Cloudflare Radar retains its own attribution and licence.