Chaos · Injected failure
An intentional failure tests how the architecture responds.
Starts at a healthy baseline — enable "Traffic Recovery" in the workspace when you want to run that optional phase.
What will happen
Follow a modeled AWS database failover. After the five-step creation warm-up, the traffic ramp rises from 3,000 to 8,000 RPS and the scheduled severe-overload phase begins at post-warmup step 15. At the 8,000 RPS plateau, writer CPU is in the low-80% range and modeled P95 is roughly 1.8× baseline, not a guaranteed doubling. Database CPU above 80% is warning; above 95% is critical, and connection pressure can independently degrade health. A healthy, explicitly related standby is promoted after a modeled one-step transition and takes the database workload. A manually injected primary crash uses the same promotion window; provisioned primary and standby resources remain billed during the outage, and the standby reflects replication work until promotion. The optional Traffic Recovery ramp is inactive by default; enable it at creation to compare recovery. This is a simulation, not a live AWS test or guarantee of provider failover timing.
Modeled demand
Traffic Ramp — watch RDS Primary approach its modeled limit · ramp · steps 0–25 · 3,000→8,000 RPS
Optional phases
Traffic Recovery · ramp · steps 60–80 · 8,000→2,500 RPS · enable in the workspace
CWM injects severe database overload at RDS Primary · Start step 15 · Duration 35 steps · End step 50.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
Starts at a healthy baseline — enable "Traffic Recovery — see GKE scale-in" in the workspace when you want to run that optional phase.
What will happen
Explore a containerized application on Google Kubernetes Engine (GKE) Autopilot with Memorystore for Redis and Cloud Pub/Sub. The model focuses on how the configured Kubernetes capacity responds to a 2,500-to-15,000 RPS ramp.
Modeled demand
Traffic Ramp — watch GKE autoscale · ramp · steps 0–30 · 2,500→15,000 RPS
Optional phases
Traffic Recovery — see GKE scale-in · ramp · steps 60–90 · 15,000→2,500 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
Starts at a healthy baseline — enable "Traffic Recovery — see AKS scale-in" in the workspace when you want to run that optional phase.
What will happen
See an Azure application handle rising demand through Kubernetes, a cache, and a queue. The simulation shows how the configured capacity and autoscaling respond; it does not contact Azure or predict production performance.
Modeled demand
Traffic Ramp — watch AKS autoscale · ramp · steps 0–30 · 2,500→15,000 RPS
Optional phases
Traffic Recovery — see AKS scale-in · ramp · steps 60–90 · 15,000→2,500 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
Global LB (anycast) · Cloud SQL auto-failover
Starts at a healthy baseline — enable "East Scale-Out — simulate full central drain to us-east1" in the workspace when you want to run that optional phase.
What will happen
Examine a bounded GCP routing-layer failure: Cloud DNS and the Global Load Balancer are modeled as unavailable while the east path remains a possible destination. The 2019 incident supplies provenance; this is not a historical timing replay.
Modeled demand
Mid-failover baseline — Global LB draining central, loading east · Hold · From step 0 · 500 RPS
Optional phases
East Scale-Out — simulate full central drain to us-east1 · ramp · steps 60–90 · 500→3,000 RPS · enable in the workspace
CWM injects severe availability-zone outage at zone usc1-zone-a · Start step 0 · Duration 9999 steps · End step 9999.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
Choose whether to include Traffic Recovery before opening the workspace. The 5-second autoscaling cooldown lets the full comparison fit in a demo run; traffic patterns lock after the first Run.
What will happen
See how a load-balanced web tier responds as demand rises, and compare scale-out with the optional later scale-in phase. This is an illustrative capacity walkthrough, not a forecast of a particular AWS deployment.
Modeled demand
Healthy baseline — hold at 800 RPS · Hold · steps 0–5 · 800 RPS
Traffic Ramp — watch autoscaling fire · ramp · steps 5–20 · 800→4,000 RPS
Optional phases
Traffic Recovery — see scale-in · ramp · steps 30–50 · 4,000→800 RPS · choose before starting the run
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Predictive · Capacity simulation
Modeled load and capacity constraints reveal bottlenecks or failure.
Starts at a healthy baseline — enable "Message Burst — overwhelm subscribers" in the workspace when you want to run that optional phase.
What will happen
Model a Google Cloud Pub/Sub subscriber fleet under rising event demand. The focus is backlog, subscriber capacity, and delivery pressure—not a live Pub/Sub outage or a guaranteed acknowledgement deadline.
Modeled demand
Steady Publisher Load · Hold · From step 0 · 1,200 RPS
Optional phases
Message Burst — overwhelm subscribers · ramp · steps 0–20 · 1,200→5,000 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
Starts at a healthy baseline — enable "Traffic Recovery — enable after step 60" in the workspace when you want to run that optional phase.
What will happen
Compare a modeled ECS Fargate service starting and replacing tasks with the same workload. This shows how ready capacity affects the response; it is not a test of Lambda Managed Instances, RDS Proxy behavior, or universal ECS limits.
Modeled demand
Idle — establish the warm task baseline · Hold · steps 0–3 · 0 RPS
Cold burst — watch tasks and application readiness + Warm hold — compare ready capacity and utilization · Hold · steps 3–28 · 2,400 RPS
Quiesce — invoke rolling deploy here if desired · Hold · steps 28–31 · 0 RPS
Re-burst — observe replacement and readiness + Post-deployment observation — hold the 2,400-RPS workload · Hold · steps 31–60 · 2,400 RPS
Optional phases
Traffic Recovery — enable after step 60 · ramp · steps 60–90 · 2,400→0 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
What will happen
Explore a two-server web application behind an Oracle Cloud Infrastructure (OCI) Load Balancer, with an Autonomous Database and object storage. The scenario starts with a steady modeled 1,400 RPS baseline; it does not contact OCI or predict production performance.
Modeled demand
Steady Baseline · Hold · From step 0 · 1,400 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
What will happen
Explore OCI’s smaller but non-zero modeled idle footprint. The ~$90/month amount and the 10 Mbps load-balancer minimum are scenario assumptions; the comparison is not a live OCI quote.
Modeled demand
Zero traffic — the bill keeps running anyway · Hold · From step 0 · 0 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
Starts at a healthy baseline — enable "Traffic Recovery — see DOKS scale-in" in the workspace when you want to run that optional phase.
What will happen
See how a DigitalOcean Kubernetes (DOKS) stack with a Load Balancer, Managed Redis, and Managed Kafka responds to rising demand. The scenario is an illustrative capacity walkthrough, not a claim about every DOKS cluster.
Modeled demand
Traffic Ramp — watch DOKS autoscale · ramp · steps 0–30 · 2,000→15,000 RPS
Optional phases
Traffic Recovery — see DOKS scale-in · ramp · steps 60–90 · 15,000→2,000 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
Starts at a healthy baseline — enable "Traffic Recovery" in the workspace when you want to run that optional phase.
What will happen
Follow modeled requests from CloudFront through ALB and EC2 to RDS. Assumed mix: 80% cacheable traffic with a 93.75% hit rate, giving 75% overall edge absorption. The origin receives 5% static misses plus 20% dynamic requests; only the dynamic requests query RDS. Watch the live request-flow breakdown. These are scenario assumptions, not measured CloudFront behavior or guarantees.
Modeled demand
Traffic Spike — scenario assumes 75% edge absorption · ramp · steps 0–25 · 2,000→8,000 RPS
Optional phases
Traffic Recovery · ramp · steps 60–80 · 8,000→2,000 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
Traffic Mgmt (DNS) · Data Guard manual SWITCHOVER
Starts at a healthy baseline — enable "Scale test — confirm Phoenix absorbs full production load" in the workspace when you want to run that optional phase.
What will happen
Explore how traffic steering and a standby database can help an OCI application during a regional network failure. Promotion and connection changes are modeled choices for the user, not an automatic switchover or a guarantee of zero data loss.
Modeled demand
us-ashburn-1 degraded — Traffic Management routing to us-phoenix-1 · Hold · From step 0 · 800 RPS
Optional phases
Scale test — confirm Phoenix absorbs full production load · ramp · steps 60–90 · 800→4,000 RPS · enable in the workspace
CWM injects severe availability-zone outage at zone iad-ad-1 · Start step 0 · Duration 9999 steps · End step 9999.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
What will happen
Explore an unaffiliated, illustrative database-primary failover affecting modeled Actions and Pages paths. It is inspired by a status update, not an official GitHub product, exact forensic reconstruction, or set of published impact measurements.
Modeled demand
Cutover Retry Storm — failover already underway · Hold · steps 0–8 · 620 RPS
Backlog Drains — replica finishes taking over · ramp · steps 8–20 · 620→120 RPS
CWM injects severe database overload at Primary Metadata Database · Start step 0 · Duration 30 steps · End step 30.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
Starts at a healthy baseline — enable "Traffic Recovery — see the stack settle back down" in the workspace when you want to run that optional phase.
What will happen
Compare the modeled cost and capacity response of a Google Cloud SaaS stack with Cloud CDN, App Engine, Cloud SQL, and Cloud Operations Suite. The 80% cache assumption is specific to this scenario, not a promise about real traffic.
Modeled demand
Business-hours ramp — watch App Engine climb into the warning zone · ramp · steps 0–60 · 90→680 RPS
Optional phases
Traffic Recovery — see the stack settle back down · ramp · steps 60–90 · 680→90 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
What will happen
Inspect DigitalOcean’s modeled residual charges after traffic reaches zero. The ~$90/month and $10/month load-balancer values are illustrative scenario inputs; they are not a live invoice or universal DigitalOcean pricing claim.
Modeled demand
Zero traffic — the bill keeps running anyway · Hold · From step 0 · 0 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Predictive · Capacity simulation
Modeled load and capacity constraints reveal bottlenecks or failure.
Starts at a healthy baseline — enable "Order Burst — overwhelm consumers" in the workspace when you want to run that optional phase.
What will happen
Model an Azure Service Bus order queue as demand rises beyond the configured consumer capacity. This is a queueing and throughput exercise; it does not call Azure or imply that Premium upgrade, auto-forwarding, or consumer scaling happens automatically.
Modeled demand
Steady Order Volume · Hold · From step 0 · 1,200 RPS
Optional phases
Order Burst — overwhelm consumers · ramp · steps 0–20 · 1,200→5,000 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
Starts at a healthy baseline — enable "Ramp to Autoscale — observe recovery behavior after fault injection" in the workspace when you want to run that optional phase.
What will happen
Start healthy with one two-pool AKS cluster behind Application Gateway and Redis in East US Zone 1, with SQL in Zone 2. Nothing self-triggers. Inject a Zone 1 outage to observe loss of the only ingress/front-door and compute path: SQL survives but receives no application reads. In a separate run, fail Redis alone while ingress and compute remain live to observe full read demand reaching SQL and rejection above its modeled 300-connection limit. This is an illustrative simulation, not measured Azure recovery timing.
Modeled demand
Wave Baseline — correlated faults cascade at steady load · wave · From step 0 · configured traffic
Optional phases
Ramp to Autoscale — observe recovery behavior after fault injection · ramp · steps 60–100 · 1,500→8,000 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Predictive · Capacity simulation
Modeled load and capacity constraints reveal bottlenecks or failure.
What will happen
Investigate a GitHub-inspired, unaffiliated retry cascade with fixed external traffic and modeled internal retry amplification. Compare the configured unprotected and protected dependency policies without treating the result as an official GitHub reconstruction.
Modeled demand
Healthy 8K RPS baseline and recovery hold · Hold · steps 0–60 · 8,000 RPS
External traffic stays at 8,000 RPS; when modeled failures trigger the configured dependency retries, effective internal request volume can be higher. Retry attempts are internal, not additional external traffic.
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
What will happen
Explore a small DigitalOcean web stack: a Load Balancer, one Droplet, and Managed PostgreSQL. It starts at a modeled 400 RPS baseline to show the topology and capacity trade-off; the result is not a promise of a specific monthly bill or production throughput.
Modeled demand
Steady Baseline · Hold · From step 0 · 400 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
What will happen
Configuration walkthrough: use independent hostnames for SPA → CloudFront → S3 and API → ALB → EC2 → RDS. A SPA 4xx → index.html rewrite cannot change errors on a genuinely separate API hostname. The diagram illustrates that desired split, but this engine runs one 150 RPS stream with 80% CDN hits: 120 RPS finish at the edge and 30 reach the application tier (15/server, about 6.2% modeled CPU). The ALB is explicitly marked as this modeled stream's origin. It does not model separately set API RPS, S3 request demand, DNS hostnames, or HTTP rewrites. A hard CDN outage blocks this single stream; it does not create an API fallback. S3 cost is an estimate with missing usage inputs, not a verified bill.
Modeled demand
Single-stream walkthrough — 150 RPS, 80% CDN hits · Hold · From step 0 · 150 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
What will happen
Learn how Redis and a standard SQS queue change modeled load as an AWS microservices stack scales from 1,000 to 5,000 RPS and recovers. The API fleet scales out at 80% CPU and in below 55% CPU, with a one-step cooldown and at most one server removed per eligible step. New API nodes are ready immediately. Once database connection demand exceeds its modeled limit, the estimated error fraction follows 1 − limit / demand (50% at twice the limit). Near the generic 75% modeled fleet-capacity envelope, capacity-overload metrics can show a small estimated share before API CPU reaches its 80% scale-out threshold; this is a model estimate, not measured loss or a queue drop. Values are illustrative, not production measurements.
Modeled demand
Traffic Spike — watch cache absorb the burst · ramp · steps 0–20 · 1,000→5,000 RPS
Traffic Recovery — drain the SQS backlog · ramp · steps 30–50 · 5,000→1,000 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Predictive · Capacity simulation
Modeled load and capacity constraints reveal bottlenecks or failure.
Starts at a healthy baseline — enable "Event Burst — overwhelm workers" in the workspace when you want to run that optional phase.
What will happen
Start healthy at 1,200 RPS with four fixed SQS consumers draining up to 1,500 messages/second. The optional 5,000 RPS burst starts at step 6 by default; simulation.create can set scenarioOverrides.sqsQueueBacklogSaturation.eventBurstStartStep to a later absolute step after the five-step creation warm-up. At burst load, the 4,000-message queue grows by about 3,500 messages/second and throttles sends when full. Producer CPU is capacity-based (about 23% at 5,000 RPS on the declared m6i.4xlarge); the producer scales out at the same 80% CPU threshold as the alert and scales in below 30% after its five-second cooldown. Producer scale-outs do not add consumers. During an RDS outage, failed work remains unacknowledged, becomes visible after the visibility timeout, and moves to the DLQ after the receive limit; modeled provisioned cost does not fall just because requests fail. The model reports queue depth, in-flight messages, oldest-message age, redeliveries, retention-expiry deletions, DLQ transfers, send rejections, and acknowledged RPS routed to RDS. Change load with Workspace traffic controls, `PATCH /api/simulations/{simulationId}` using `traffic`, or `simulation.inject_traffic`; for a finalized SQS run, these traffic edits preserve the replay snapshot and apply at runtime. This is an illustrative simulation, not a live SQS measurement.
Modeled demand
Steady Producer Load · Hold · From step 0 · 1,200 RPS
Optional phases
Event Burst — overwhelm workers · burst · steps 1–20 · peak 5,000 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
Starts at a healthy baseline — enable "Traffic Recovery — enable after step 60" in the workspace when you want to run that optional phase.
What will happen
Compare a Lambda application that must start new execution capacity with one that begins with capacity already available. The scenario shows how startup delay affects a fixed workload. It is a simplified CWM comparison, not a reproduction of AWS Lambda Managed Instances or published benchmark results.
Modeled demand
Idle — establish the cold baseline · Hold · steps 0–3 · 0 RPS
Cold burst — watch application readiness reject traffic + Warm hold — compare ready capacity and utilization · Hold · steps 3–28 · 2,400 RPS
Quiesce — clear traffic before the re-burst · Hold · steps 28–31 · 0 RPS
Re-burst — observe readiness on a second cold start + Post-deployment observation — hold the 2,400-RPS workload · Hold · steps 31–60 · 2,400 RPS
Optional phases
Traffic Recovery — enable after step 60 · ramp · steps 60–90 · 2,400→0 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
What will happen
See how provider limits, retries, and queueing affect a batch of AWS EC2 changes at the selected concurrency. This is a catalog-backed provider-limit simulation separate from the workspace resource state.
CWM models 240 AWS EC2 mutating requests in us-east-1 at a concurrency of 8. Some requests may be slowed, queued, or retried. The results show how provider limits, retries, and queueing affect the batch.
No AWS APIs are called and no cloud resources are created.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
What will happen
Compare a modeled hybrid path that connects AWS and OCI compute, databases, load balancers, and object storage. The goal is to inspect configuration trade-offs at a steady 1,200 RPS, not to measure a real cross-cloud network.
Modeled demand
Steady Baseline · Hold · From step 0 · 1,200 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
Route 53 (DNS) · RDS Multi-AZ auto-failover
Starts at a healthy baseline — enable "Traffic Recovery — watch us-east-1 scale in as us-west-2 recovers" in the workspace when you want to run that optional phase.
What will happen
Start healthy at 120 RPS split 50/50 between us-west-2 (Oregon) and us-east-1 (N. Virginia), with four m5.xlarge web servers at about 6% modeled CPU. Each m5.xlarge has a scenario-assumed 500 RPS serving ceiling; typical-app CPU is estimated per server. The overall metrics.cpuUsage is a system-wide aggregate; individual resource CPU and scaling-history group means are separate values. Cost breakdown rows are per provisioned resource, and failed resources can remain billed. Latency adds the modeled queueing term 10 ms × CPU/(1−CPU), with CPU as a fraction capped at 0.99 for finite output; this is not a provider benchmark. West fails at evaluation step 5: the first user step after the five-step creation warm-up. Route 53 uses 10-second probes, three consecutive failures or recoveries, and a 60-second resolver-cache TTL; detection and cache drain are separate clocks. Each step models 10 seconds. Offered traffic rises from 120 to 2,600 RPS during evaluation steps 10–40 (user steps 6–36), reaching its 2,600 RPS peak on response step 41 (user step 36), then holds. During failback, DNS share is held to routable regional capacity until the returning fleet can serve it. If a database outage overlaps stale DNS, requests rejected at stale endpoints fail before reaching the database and are counted only as stale-DNS errors; database errors apply only to requests that reach the database. metrics.route53Failover.staleDnsTrafficShare separately reports the share still pointed at unavailable endpoints. The CPU alert fires above 80%; AWS scale-out starts at a 70% mean CPU threshold for the per-server group, with scale-in below 30% group CPU and 25% throughput utilization (25% of declared group capacity). The 300-second scale cooldown is 30 simulation steps at this 10-second clock. Scaling-history rows show the response step when the decision takes effect and the earlier evaluation step when the policy ran. West remains down through evaluation step 59; regional recovery starts at evaluation step 60 (user step 56), and the optional recovery ramp starts then but is inactive by default. The modeled, not measured, replication lag estimate accumulates elapsed step time in proportion to offered traffic (full rate at 2,600 RPS) while the west writer is unavailable, then freezes at promotion. DNS failover does not automatically promote the RDS replica; inject a west writer failure to exercise explicit cross-region promotion. This is not a historical timing replay.
Modeled demand
Pre-outage baseline — Route 53 splits traffic between west and east · Hold · steps 0–9 · 120 RPS
East Scale-Out — DNS TTLs expire, us-east-1 absorbs us-west-2 drain · ramp · steps 10–40 · 120→2,600 RPS
Optional phases
Traffic Recovery — watch us-east-1 scale in as us-west-2 recovers · ramp · steps 60–90 · 2,600→120 RPS · enable in the workspace
CWM injects severe availability-zone outage at zone usw2-az1 · Start step 5 · Duration 55 steps · End step 60.
CWM injects severe availability-zone outage at zone usw2-az2 · Start step 5 · Duration 55 steps · End step 60.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
What will happen
Explore how an application responds when a Bahrain region and part of a UAE deployment become unavailable while a remote backup remains available. CWM shows the difference between keeping services running and recovering lost data; it is an illustrative exercise, not a replay of a real AWS recovery.
Modeled demand
Healthy baseline — illustrative demand · Hold · From step 0 · 100 RPS
CWM injects severe region outage at aws region me-south-1 (Bahrain) · Start step 10.
CWM injects moderate network latency at UAE AZ 1 — Illustrative Surviving App · Start step 10.
CWM injects severe availability-zone outage at zone mec1-az2 · Start step 10.
CWM injects moderate network latency at UAE AZ 3 — Illustrative Surviving Data · Start step 10.
CWM injects severe permanent data loss at aws region me-south-1 (Bahrain) · Start step 20 · Terminal data loss; infrastructure recovery cannot restore original data.
CWM injects severe permanent data loss at zone mec1-az2 · Start step 20 · Terminal data loss; infrastructure recovery cannot restore original data.
Initial incident snapshot: app and primary SQL warning states describe incident impact, not local backend failure or measured CPU pressure. In the loaded workspace, local health and routed availability are separate: healthy apps can be unreachable behind failed Front Door, with zero routed request traffic and idle CPU. Recovery updates runtime state. Reset clears runtime history and deactivates failures while preserving workspace configuration; reload the scenario to restore the starting incident.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
Front Door (anycast) · Active Geo-Replication auto-failover
Starts at a healthy baseline — enable "Traffic Manager failover — direct bypass traffic floods West US origin" in the workspace when you want to run that optional phase.
What will happen
Learn why a healthy multi-region backend does not protect against failure of the global proxy in front of it. The 2025 Front Door incident is context; the Traffic Manager path is a modeled, optional bypass exercise. Initial incident snapshot: app and primary SQL warning states describe incident impact, not local backend failure or measured CPU pressure. In the loaded workspace, local health and routed availability are separate: healthy apps can be unreachable behind failed Front Door, with zero routed request traffic and idle CPU. Recovery updates runtime state. Reset clears runtime history and deactivates failures while preserving workspace configuration; reload the scenario to restore the starting incident.
Modeled demand
Stale-DNS clients, unreachable behind failed Front Door · Hold · From step 0 · 50 RPS
Optional phases
Traffic Manager failover — direct bypass traffic floods West US origin · ramp · steps 60–90 · 50→3,000 RPS · enable in the workspace
CWM injects severe global ingress outage at Azure Front Door — Global Anycast (global ingress) · Start step 0 · Duration 9999 steps · End step 9999.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
Floating IP (DNS) · pg_logical manual promotion
Starts at a healthy baseline — enable "Scale test — confirm SFO3 Droplet handles full NYC3 production load" in the workspace when you want to run that optional phase.
What will happen
See how an application can keep serving when its New York site fails and traffic moves to San Francisco. The database still needs manual cross-region recovery, so this exercise shows the difference between routing traffic and restoring data; it does not promise automatic failover or a provider SLA.
Modeled demand
NYC3 degraded — DNS update in progress, traffic routing to SFO3 · Hold · From step 0 · 600 RPS
Optional phases
Scale test — confirm SFO3 Droplet handles full NYC3 production load · ramp · steps 60–90 · 600→2,500 RPS · enable in the workspace
CWM injects severe availability-zone outage at zone nyc3-az1 · Start step 0 · Duration 9999 steps · End step 9999.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
What will happen
Study a bounded, incident-inspired AWS regional routing failure that is structural rather than demand-driven. Creation loads the crisis at S5, after the five initialization ticks: three web servers consume assumed local retry CPU (88%, 85%, 82%) but serve no external requests, the primary starts at 62% CPU, and the cache is critical. The related RDS standby promotes on S7; it cannot bypass the isolated ALB and compute fleet. Authored failure windows 0–40 (RDS) and 0–45 (zones) are shifted once by initialization to evaluation windows 5–45 and 5–50. They remain effective through responses S45/S50; expiry is first evaluated on S46/S51, followed by recovery warm-up. The default recovery traffic ramps automatically from 300 RPS at S60 to 2,000 at S90 and then holds. The historical date and incident basis provide context; the card’s steps, local retry CPU assumptions and recovery behavior are an illustrative CWM model, not a historical replay.
Modeled demand
Steady outage load — structural crisis at low baseline · Hold · steps 0–59 · 300 RPS
Traffic Recovery — watch services stabilize · ramp · steps 59–89 · 300→2,000 RPS
CWM injects severe availability-zone outage at zone us-west-2-regional · Start step 0 · Duration 45 steps · End step 45.
CWM injects severe availability-zone outage at zone us-west-2-compute · Start step 0 · Duration 45 steps · End step 45.
CWM injects severe database overload at RDS MySQL Primary · Start step 0 · Duration 40 steps · End step 40.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
What will happen
Find the residual AWS charges that remain in this configured zero-traffic topology. The ~$310/month figure is the scenario’s illustrative starting estimate, not a bill or price forecast.
Modeled demand
Zero traffic — the bill keeps running anyway · Hold · From step 0 · 0 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
What will happen
Inspect modeled GCP residual charges in a decommissioned, zero-traffic project. The ~$120/month amount is an illustrative configuration estimate; the exercise is not a live GCP invoice.
Modeled demand
Zero traffic — the bill keeps running anyway · Hold · From step 0 · 0 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
What will happen
See which Azure resources can continue contributing to a modeled residual bill after application traffic reaches zero. The ~$175/month figure is an illustrative starting estimate, not an Azure invoice.
Modeled demand
Zero traffic — the bill keeps running anyway · Hold · From step 0 · 0 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Predictive · Capacity simulation
Modeled load and capacity constraints reveal bottlenecks or failure.
What will happen
Model an API with an Application Load Balancer, Lambda, and DynamoDB. Traffic holds at 100 RPS, ramps to 2,000 RPS from steps 20–35, then stays at the peak. Lambda scales toward a modeled 70% utilization target with a 5,000 RPS fleet ceiling. Scale-out exposes initializing environments and about 300 ms of modeled latency on affected requests; DynamoDB request-based modeled cost rises with traffic. This is a simulation, not a live AWS measurement.
Modeled demand
Quiet period — Lambda idles · Hold · steps 0–19 · 100 RPS
Traffic burst — cold starts fire · ramp · steps 19–34 · 100→2,000 RPS
Peak hold — observe steady capacity · Hold · From step 34 · 2,000 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
Starts at a healthy baseline — enable "Traffic Drain — reduce load during incident response" in the workspace when you want to run that optional phase.
What will happen
Explore modeled cache loss with a routable CloudFront proxy and two fixed web servers. At 2,000 total RPS, origin load rises from 500 to 2,000 RPS total — 250 to 1,000 per server. CPU is modeled, not measured incident telemetry. A hard edge outage instead blocks forwarding; this exercise is not an incident replay.
Modeled demand
Steady client load — cache loss forwards 2,000 origin RPS · Hold · From step 0 · 2,000 RPS
Optional phases
Traffic Drain — reduce load during incident response · ramp · steps 60–80 · 2,000→500 RPS · enable in the workspace
CWM injects severe cdn cache loss at CloudFront CDN · Start step 0 · Duration 9999 steps · End step 9999.
Chaos · Injected failure
An intentional failure tests how the architecture responds.
Global LB (anycast) · Paxos consensus (zero RPO)
What will happen
See how an application responds when all configured Cloud Spanner replicas lose their serving path. CWM models the resulting service interruption and recovery choices; it does not reproduce the real incident’s cause, duration, or customer impact.
Modeled demand
Pre-outage baseline — 500 RPS hitting Spanner hard · Hold · From step 0 · 500 RPS
CWM injects severe availability-zone outage at zone usc1-zone-b · Start step 0 · Duration 9999 steps · End step 9999.
CWM injects severe availability-zone outage at zone use1-zone-c · Start step 0 · Duration 9999 steps · End step 9999.
CWM injects severe availability-zone outage at zone use4-zone-a · Start step 0 · Duration 9999 steps · End step 9999.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
Starts at a healthy baseline — enable "Traffic Recovery — see the stack settle back down" in the workspace when you want to run that optional phase.
What will happen
Compare the modeled cost and capacity response of an Azure SaaS stack with Front Door, App Service, PostgreSQL Flexible Server, Monitor, and Defender. The 75% cache assumption and warning-band outcome belong to this configuration, not all Azure SaaS workloads.
Modeled demand
Business-hours ramp — watch App Service climb into the warning zone · ramp · steps 0–60 · 200→1,550 RPS
Optional phases
Traffic Recovery — see the stack settle back down · ramp · steps 60–90 · 1,550→200 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Chaos · Injected interruption
A scheduled AWS Spot interruption tests whether affected workloads become ready before the termination deadline.
What will happen
Practice a modeled Amazon EKS Spot interruption with demand autoscaling disabled. The ramp overloads the worker CPU capacity before the notice. Four workloads reschedule in two 18-second waves; compare workload readiness at the two-minute deadline (simulation second 150) with ready-node capacity through actual reclaim, replacement startup, and recovery. Service is bounded by both ready nodes and the workload-ready share of configured-fleet capacity. Initial CPU figures are modeled 150 RPS estimates: mean serving-worker CPU for EKS and database CPU. Spot worker pricing uses a 0.3 modeled estimate, not a live quote; the control-plane fee is undiscounted.
Modeled demand
Steady application traffic · Hold · From step 0 · 150 RPS
Pre-interruption demand ramp — CPU overload · ramp · steps 6–10 · 150→1,100 RPS
Overload hold — compare serving capacity through notice · Hold · steps 10–52 · 1,100 RPS
Post-migration recovery — return to healthy baseline · ramp · steps 52–55 · 1,100→150 RPS
Interruption: EKS Spot Worker Fleet · AWS Spot interruption · Start step 10 · 120-second deadline at 3 simulated seconds/step; 4 workloads must be rescheduled.
Predictive · Capacity simulation
Modeled load and capacity constraints reveal bottlenecks or failure.
What will happen
See Fargate scale from one task to three, wait for startup, and reach its ceiling. An assumed application FIFO holds 300 requests before shedding overflow. Recovery drains the queue and a modeled application idle controller scales to zero—not a native CPU-only AWS alarm. Running tasks bill vCPU/memory; calibrated latency percentiles are unavailable.
Modeled demand
Traffic Ramp — watch Fargate tasks start · ramp · steps 0–30 · 30→500 RPS
Traffic Recovery — drain the queue and scale to zero · ramp · steps 38–53 · 500→0 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Predictive · Capacity simulation
Modeled load and capacity constraints reveal bottlenecks or failure.
What will happen
See how a 25→250 RPS launch ramp crosses the scenario-configured 180 RPS upload-server and 40 RPS transcoding-worker ceilings, triggering modeled capacity-overload signals and load shedding. The queue does not back up in this run. These are bounded planning assumptions, not OCI/provider limits or real-world upload thresholds; this is not a production forecast.
Modeled demand
Launch Day Surge (10×) · ramp · steps 0–30 · 25→250 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
Starts at a healthy baseline — enable "Traffic Recovery — see the stack settle back down" in the workspace when you want to run that optional phase.
What will happen
Compare the modeled cost and capacity response of an AWS SaaS stack with CloudFront, App Runner, Aurora PostgreSQL, and CloudWatch. The 80% cache assumption and warning-band outcome are scenario configuration, not a universal AWS result.
Modeled demand
Business-hours ramp — watch App Runner climb into the warning zone · ramp · steps 0–60 · 120→850 RPS
Optional phases
Traffic Recovery — see the stack settle back down · ramp · steps 60–90 · 850→120 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
What will happen
Understand how a global single-page application (SPA) error rewrite can change API status codes when the SPA and API share one CloudFront distribution. This is a configuration walkthrough, not a load failure.
Modeled demand
Steady API load — routing bug is live, not a load problem · Hold · From step 0 · 500 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Educational · Guided scenario
A predefined, illustrative walkthrough guides you through modeled behavior; it is not a production forecast.
What will happen
Start with healthy Redis and PostgreSQL at 5,000 RPS by default; a create-time traffic value sets the actual starting load. The model assumes 80% of requests are cacheable and 80% of those hit, so about 1,800 RPS reach PostgreSQL (roughly 360 of its 500 modeled connections). Fail Redis manually to send all offered traffic to PostgreSQL. Above the 500-connection serving limit, errors represent the excess fraction of demand; at 2× demand, PostgreSQL still serves half the requests. Provisioned costs do not fall merely because errors reduce successful throughput. After recovery, the first request-flow step remains at 0% hits, then hit rate warms to 40% and 80% over the next two steps. Change load at any point with Inject Traffic (`simulation.inject_traffic` or `POST /api/simulations/{simulationId}/inject-traffic`); this scenario has no active traffic pattern, so a controlled target remains usable after replay identity is finalized. No failure or recovery is scheduled by default.
No traffic pattern is scheduled.
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Predictive · Capacity simulation
Modeled load and capacity constraints reveal bottlenecks or failure.
Starts at a healthy baseline — enable "Traffic Recovery — see per-token cost climb as GPUs idle" in the workspace when you want to run that optional phase.
What will happen
Study how a modeled GKE GPU inference pool spreads fixed node cost across tokens as utilization changes. The T4, utilization, autoscaling, and cost-per-million-token relationships are this configured model, not a forecast for every large-language-model workload.
Modeled demand
Inference ramp — watch cost/M tokens fall as GPU utilization rises · ramp · steps 0–40 · 200→2,000 RPS
Optional phases
Traffic Recovery — see per-token cost climb as GPUs idle · ramp · steps 60–90 · 2,000→200 RPS · enable in the workspace
No failure is scheduled; any degradation should emerge from modeled load and capacity.
Optimization · Configuration comparison
The engine compares configuration trade-offs and surfaces the better option.
What will happen
Show the modeled cost of an EKS GPU pool with no inference traffic. With zero requests there are no generated tokens, so cost per million tokens is undefined/infinite while configured GPU capacity continues to carry cost; the result is a waste-detection exercise, not a provider bill.
Modeled demand
Zero traffic — the GPUs bill anyway · ramp · steps 0–60 · 0 RPS
No failure is scheduled; any degradation should emerge from modeled load and capacity.