feat: CallGraph API #3178
feat: CallGraph API #3178
Conversation
| message CallGraphNode { | ||
| // Values associated to this node. Self is exclusive to this node, | ||
| // total includes all descendents. | ||
| uint64 total = 1; | ||
| uint64 self = 2; | ||
| string name = 3; | ||
| } |
There was a problem hiding this comment.
Note that we use strings without indirection as in flame graph or pprof formats. The reason is that call graph nodes never duplicate; therefore, the names are meant to be unique.
In contrast to the call graph implementation in the pprof tool, we only keep the display name of the function. Function details such as file name, line number, versions, etc. should be retrieved via separate API (shared with flame graphs).
|
/cc @petethepig @aocenas |
|
Could we use this to replace what was done in #2808? |
Yeah, ideally we should be using the new API there as well |
|
For my 2 cents here, this looks good. Structurally the same shape as what we send to NodeGraph already so would not need too much transformation. We can also either "hide" it behind the flamegraph query so the user won't know there are 2 API calls being made. |
Resolves #3219
The PR introduces a new endpoint for querying profiling data in the call graph format. The format is greatly inspired by the pprof's implementation: https://github.com/grafana/pyroscope/blob/main/pkg/frontend/dot/graph/graph.go#L50
Despite the fact that a call graph can be built from a flame graph or a profile in pprof format, the new use-case specific API will significantly help us integrate call graphs into the Pyroscope UI.
The issue lies in the fact that truncating flame graphs/pprof profiles is not suitable for call graphs: the heuristics we use to trim insignificant nodes make the call graph incorrect. A workaround could be keeping all the nodes (disabling the truncation); however, this will not work well with large profiles. Thus, it's preferable to build call graphs in the backend, trimming nodes in a way optimized for call graphs; otherwise, we will be juggling between incorrect call graph structures/numbers and very poor performance.
TODO: