> I can’t imagine how we would have solved this problem without being able to sample Node.js stacks and visualize them with flame graphs.
This has me scratching my head. The diagrams are pretty, maybe, but I can't read the process calls from them (the words are truncated because the graphs are too narrow). And I can't see, visually, which calls are repeated. They're stacked, not grouped, and the color palette is quite narrow (color brewer might help here?).
At least, I _can_ imagine how you could characterize this problem without novel eye-candy. Use histograms. Count repeated calls to each method and sort descending. Sampling is only necessary if you've got -- really, truly, got -- big data (which Netflix probably does), but I don't think the author means 'sample' in a statistical sense. It sounds more like 'instrumentation', decorating the function calls to produce additional debugging information. Either way, once you have that, there are various common ways to isolate performance bottlenecks. Few of which probably require visual graphs.
There's also various lesser inefficiencies in the flame graphs: is it useful (non-obvious) that every call is a child of `node`, `node::Start`, `uv_run`, etc.? Vertical real-estate might be put to better use with a log-scale? Etcetera, etc.
> The diagrams are pretty, maybe, but I can't read the process calls from them (the words are truncated because the graphs are too narrow).
Flame Graphs provide SVGs by default. You should be able to zoom if your broser supports it. The current version also supports "zooming" in to any frame in the stack, resetting that frame as the base of the display. Also WRT the base frames of 'node' et al its because Flame Graphs are a general use tool for stack visualization, it might be 'main' for a c program or the scheduler looking at a system.
> They're stacked, not grouped, and the color palette is quite narrow (color brewer might help here?).
Colors by default have no meaning and the palette is configurable. The current lib can also assign colors by instruction count/ipc and width by call count, if you have access to that.
> Sampling is only necessary if you've got -- really, truly, got -- big data (which Netflix probably does), but I don't think the author means 'sample' in a statistical sense.
It is sampling. Flame graphs re typically used with something like perf/dtrace/oprofile which dumps stacks at a couple hundred to thousand hertz. Actual call tracing is (typically) not feasible for large/prod stacks.
You're looking at a screenshot. The actual diagram isn't static. Hover over it and it will expand each box with all the info you need to see what was run, what called it, etc.
This has me scratching my head. The diagrams are pretty, maybe, but I can't read the process calls from them (the words are truncated because the graphs are too narrow). And I can't see, visually, which calls are repeated. They're stacked, not grouped, and the color palette is quite narrow (color brewer might help here?).
At least, I _can_ imagine how you could characterize this problem without novel eye-candy. Use histograms. Count repeated calls to each method and sort descending. Sampling is only necessary if you've got -- really, truly, got -- big data (which Netflix probably does), but I don't think the author means 'sample' in a statistical sense. It sounds more like 'instrumentation', decorating the function calls to produce additional debugging information. Either way, once you have that, there are various common ways to isolate performance bottlenecks. Few of which probably require visual graphs.
There's also various lesser inefficiencies in the flame graphs: is it useful (non-obvious) that every call is a child of `node`, `node::Start`, `uv_run`, etc.? Vertical real-estate might be put to better use with a log-scale? Etcetera, etc.