dh

A few adjustable distractions

What happens if you…

Small controls, visible consequences.
Browser simulations are self-contained; live cards use bounded demo endpoints.

01

Room for one more?

Raise CPU demand and watch the replica target change, then see which pods fit.

Kubernetes HPA / SVG simulation

Illustrative preview
Explore Room for one more?

Turn up CPU demand. HPA changes the target, then the scheduler tries to place each pod.

HPA sets the replica target. Placement uses each pod's CPU request, so a pod can stay pending when there is no room.

One CPU snapshot from two ready replicas. Each node has 2,000m available CPU. Replica bounds 1–12; 10% HPA tolerance. Illustrative least-used placement, not the full Kubernetes scheduler. No memory, affinity, readiness delays or HPA stabilisation. Nodes are added manually here.

02

Who’s allowed to do what?

Choose a scenario, set the call budget and decide whether a simulated write is allowed.

Agents + MCP / scripted SVG simulation

Illustrative preview
Explore Who’s allowed to do what?

Follow the scripted agent as it checks context, calls tools and stops at a timeout, missing context or denied write.

The script proposes a call. The host checks context, budget and approval before a simulated write.
Follow the trace
  1. Receive the task and check the available context.
  2. Propose a read-only tool call within the host's budget.
  3. Review the result. A proposed write needs host approval.
  4. Report the findings, or stop when context, permission or budget is missing.

A scripted agent loop, not a live model or MCP server. Approval and budgets belong to this host policy, not automatic guarantees provided by MCP. Failed attempts count towards the budget.

03

A little interference.

Compare a clean wave with the same wave after interference is added.

Canvas 2D / synthetic interference

Illustrative preview
Explore A little interference.

The upper trace stays clean. Change the lower trace's noise, frequency and amplitude.

Synthetic interference, not signal recovery or incident detection.

04

The average looks fine.

Keep most requests steady, then stretch the slowest ten.

SVG / synthetic latency

Illustrative preview
Explore The average looks fine.

Move the tail slider and watch the median stay put while the slow end grows.

The median stays steady; the slowest requests move the upper percentiles.

Sorted samples, fixed 0–2,000 ms scale. The readout gives the final distribution; the animation draws it in. No live traffic.

05

Requests in flight.

Change how many requests run together and see which finish before the deadline.

SVG / concurrency simulation

Illustrative preview
Explore Requests in flight.

Run the lookup with a different number in flight. Queued time counts against the deadline.

The readout will show how many of 12 requests meet the deadline.

Uniform duration, no retries or upstream contention. The readout gives the result at the deadline. Outlines are planned work; hatching marks work cut short.

06

A small connected world.

Rotate the graph or select a node to see its direct connections.

Canvas 2D / imaginary 3D topology

Illustrative preview
Explore A small connected world.

Drag to rotate. Select a node to highlight its neighbours and dim the rest.

All connections visible.
See the connections
  • Browser: API, Cache
  • API: Browser, Queue, Store
  • Queue: API, Worker
  • Worker: Queue, Store
  • Store: API, Worker
  • Cache: Browser

Further afield

The Internet has a pulse.

Pick a country and scrub through its observed week.

Radar / Workers KV

Cloudflare Radar

Observed data loads when the interactive view is ready.

Internet atlas01 / 08
Country markers, not network routesMap: Natural Earth (opens in a new tab)
Observed traffic, normalised to 100GB

United Kingdom

Not loaded/ 100No observed reading

No observations loaded yet.

Observed weekHourly · UTCNot loaded

Relative HTTP request volume across this country's observed week.

Reported disruptions

No observations loaded yet.

Relative traffic, not uptime or absolute country volumes. No reported events does not mean no outages.

How to read itRead the shape, not the size. Each country's peak is 100; a dip alone does not establish an outage.

Data: Cloudflare Radar (opens in a new tab) · CC BY-NC 4.0 (opens in a new tab) · Custom visualisation; traffic normalised by country.

A few live requests

Outside the browser.

These cards make bounded requests to Cloudflare, then show exactly what comes back.

07

You are here.

Your request, at the edge.

Workers / request.cf metadata

Illustration; open to inspect the metadata returned for this request
Explore You are here.
Country
Not loaded
Serving edge
Not loaded
Protocol
Not loaded
TLS
Not loaded
Connection round trip
Not loaded
Network
Not loaded

Open to inspect this request.

Country is approximate. Your IP address isn’t returned to this page.

08

Still answering?

A day of scheduled checks, including gaps.

Cron Triggers + D1 / asset checks

Open to load observed samples.
Scheduled checks; open to inspect recorded response headers
Explore Still answering?

Gaps are unmeasured intervals; crosses are failed checks.

Open to load the measurement window.

One check every five minutes through this environment’s asset binding. Response headers only, excluding public routing, Access and browser rendering. Not uptime monitoring or an SLO.

09

Bit by bit.

Start a response and watch six chunks arrive.

Workers / Web Streams API

Illustration; run it to receive six real chunks
Explore Bit by bit.

No chunks received.

No request until you press Start.

A Cloudflare Worker sends authored text in six small frames. Each marker lights up as a frame arrives. Buffering can group frames together; this is not a network-speed test. No AI call or stored data.

10

On the same page.

Connect two windows and send a numbered pulse between them.

Durable Objects / WebSockets

Shared sequence; connect to send a real pulse
Explore On the same page.
Shared sequenceNot connected

Not connected.

A small shared demo room, not chat. No names or messages. Short sessions, bounded pulses, and connections close when this view is hidden. Idle connections use hibernation.

11

A second opinion.

Choose a fictional incident and get a hypothesis, two checks and the unknowns.

Workers AI + Turnstile / synthetic incidents

Scenario facts

Synthetic API: median latency 80 ms, p99 1,800 ms. Error rate steady at 0.2%. Database connection wait increased after a deployment. No trace samples have been checked yet.

What could explain the slow requests, and what should be checked next?

No model request until you press Run triage.

Reviewed interpretation

Database wait is a lead, not a confirmed cause. Check traces and pool measurements. A change after deployment does not prove causation.

This is the authored baseline for comparison. The model sees only the selected facts and question.

Llama 3.3 70B on Cloudflare Workers AI. Fictional incidents, real inference. No tools or access to live systems. Turnstile verifies requests; D1 enforces a daily request limit.

No real systems
were harmed.

The little simulations run in your browser. The Internet atlas looks further afield, using public observations from Cloudflare Radar. The edge experiments use their own bounded services, separate from the browser-local simulations.

Explore the source ↗ (opens in a new tab)