While DAG (Direct Acyclic Graphs) and their variants are good to start with - in practice you may also want to ensure the 'directed' representation implementation does not become bottleneck later (if you go with implementing it as a graph, as-is).

For example, you may want to answer questions like "what does this component depends on" as well as "what all components depends on this" (the opposite direction). So, it is better to ensure that your actual implementation of the DAG allows such reverse traverals too right from the ground up.

One right way in such case would be to go with 'boolean matrix' representation. Once there, the techniques of Sparse and triangulation help optimize the memory / access further.

Also, the boolean vector / matrix representations are easily scalable across multiple machines (thus leading to ease of big-data readiness, just in case).

http://gk.palem.in/

I prefer a direct acyclic graph (it might not exist, it will depend on the dependences). But I don't understand what you mean with 'service dependency graph'

Direct graph is the basic representation. But in recent researches, more people start to use the multi-edge direct graph, in which two or more edges between two components can represent more complex relations. Is ‘Service Dependency Graph’ means the graph with edges representing resource dependency between services?

While DAG (Direct Acyclic Graphs) and their variants are good to start with - in practice you may also want to ensure the 'directed' representation implementation does not become bottleneck later (if you go with implementing it as a graph, as-is).

For example, you may want to answer questions like "what does this component depends on" as well as "what all components depends on this" (the opposite direction). So, it is better to ensure that your actual implementation of the DAG allows such reverse traverals too right from the ground up.

One right way in such case would be to go with 'boolean matrix' representation. Once there, the techniques of Sparse and triangulation help optimize the memory / access further.

Also, the boolean vector / matrix representations are easily scalable across multiple machines (thus leading to ease of big-data readiness, just in case).

http://gk.palem.in/

@Zhiling Luo:  Yes it is.

@Gopalakrishna Palem: I should start from the question "what does this component depends on"  or what does this service depends  on. A componente would have one or more services. But I should start  modeling dependencies between components.