Explain tells you what you drew: it reads your Grid and describes it back in words. Simulate tells you what it would do: it pushes traffic through the graph and reports where it saturates. A clear description can belong to a design that falls over, so run both.
Explain: a description for one reader
Open Explain from the top bar to describe the whole Grid, or select nodes and use Explain on the selection bar to describe only those. Then pick who the text is for.
- Engineer: components and protocols.
- Manager: outcomes and capabilities, minimal jargon.
- Onboarding: little assumed, domain terms defined, why each part exists.
- Interview review: what the diagram communicates, for someone assessing a candidate, with no judgement.
The result has five sections: overview, request flow, data flow, responsibilities and external dependencies.
A run needs no plan and costs 8 credits, which the welcome grant covers (a very large Grid costs more); the Pro plan and above add a monthly allowance, and AI credits: what they buy and how they are spent has the rest. You need to be signed in on a saved Grid.
Simulate: what it needs from you
Traffic Simulation is free on every plan: no AI, just arithmetic over the graph. It is not available on Pattern Grids. Every catalog node carries a capacity profile by type, so there is nothing to configure per diagram. It needs catalog components, a client or frontend node for the traffic to enter through, and a traffic number, typed or set with the slider. A rectangle labelled "database" has no capacity: a Grid of nothing but shapes does not offer Simulate, and on a mixed Grid a shape is left out of the numbers, so an imported or hand-drawn Grid needs real components.
Remove the Client and the panel reports no entry point: no traffic enters, however complete the rest looks.
Type 800 requests a second into this Grid and Feed Service sits at 80%, because the catalog gives an application service 1,000 a second per replica and a new node starts with one. Type 8,000 and it is at 800%: the model serves those 1,000 a second and counts the rest as unanswered, so about one request in eight gets an answer. A cache serves the share of reads its hit rate sets, a queue defers, and replicas divide the load. These figures model relationships between components; they are not benchmarks of your system.
Two dials change the result:
- Read/write mix. The total stays the same, and caches absorb only the reads. With ten Feed Service replicas at 8,000 requests, Posts DB (10,000 a second in the catalog) goes from 29% at 80% reads to 48% at 50%: every write reaches it, and the cache stops 80% of reads.
- Replica count. Hover a node for its − N + stepper, or use Replicas in the Node Inspector. It moves a bottleneck without redrawing anything: ten replicas bring Feed Service back to 80% at 8,000. On a single-primary database, replicas serve reads only, so writes still stop at one primary's capacity.
Predict first, on News Feed
On the News Feed Challenge, open Simulate before you submit. It starts at the load the Challenge states, 8,000 requests a second with 90% reads. Before you read anything, write down which node you expect to saturate first, then check the Bottlenecks section. Type 80,000 and predict again. The panel marks that run as a what-if, since the grade runs at the stated load; it is still a fair test of whether you can read your own design.