Performance Lab

A model of one request-serving system: client → app → cache → database → external API. Move traffic, cores, hit rate, pool size, timeouts and retries, and watch which number moves — and which one refuses to.

SIMULATEDEverything below is produced by a queueing model inside Engineer Atlas.

The absolute numbers describe no real system. What is faithful is the behaviour: latency bending upward near saturation, the tail moving before the median, the bottleneck relocating after every fix, and retries feeding back into arrival rate. Real traffic is burstier than this model assumes, so real systems reach the knee earlier than the sliders suggest.

Controls

Traffic1,000 req/sOffered load, before any retries. Watch what happens as this crosses the knee rather than at any single value.App CPU cores16 coresTotal cores across the app tier. This is the capacity that CPU demand queues for.CPU per request8 msHow expensive your handler is. Raising this is what an inefficient algorithm looks like from the outside.Cache hit rate90 %Every point you lose here multiplies database traffic. Drop it from 96 to 60 and watch the pool, not the cache.Cache latency2 msService time of a hit. Cheap by design — if this is not cheap, the cache is not helping.DB query latency25 msHow long a query holds a connection. This is the service time the pool queues for, so it sets pool capacity.DB connection pool20 connsServers in the pool queue. Too small and requests wait while the database sits idle.Requests calling external API20 %A dependency you do not own and cannot scale. It adds latency without adding local utilisation.External API latency120 msTheir mean. The model gives it a fat tail, because third-party p99 usually runs several times its mean.Client timeout2,000 msRequests slower than this count as errors. Lower it and errors rise; raise it and slow requests hold capacity longer.Retry rate on failure0 %The feedback loop. Harmless while the system is healthy; past the knee it turns degradation into collapse.

SIMULATEDModel output

p50

37 ms

p95

61 ms

p99

85 ms

throughput

1,000/s

errors

0.0%

in flight

Resource utilisation

App CPU0.50

DB connection pool0.12

External APIbottleneckno local queue — it consumes latency, not capacity

The external API dominates: 24 ms of the 37 ms mean, with no local queue to show for it. Utilisation dashboards will look healthy while users wait.

Change one variable

Each experiment snapshots the current settings as a baseline, then applies exactly one idea. Predict the result before you click.