So my questions are:
- Why isn't diagram generation automated as part of the build process (UML or otherwise)?
- Why aren't code visualization tools more popular? The options out there seem outdated
- Would you want to use these tools? What would be your ideal tool?
Edit: looks like this is a duplicate question https://news.ycombinator.com/item?id=31569646
I can't delete it so feel free to discuss more
In practice it's the same problem as "noodle graph" visual programming. It works well in some niches (e.g. creating shaders in graphics programming, or sometimes describing AI tasks in game programming), but it completely breaks down outside those niches.
One type of diagram I have found to be truly useful, though, is the sequence diagram. I needed to integrate someone else's library into my application, and having this was a huge help.
If anyone has a pointer to a good sequence-diagram generator (that runs on Mac, preferably), I'd be happy to hear about it!
That is nontrivial, and it is very hard to do well. But it is also the essential job necessary when designing software and then communicating the essence of that design.
My favourite tool btw is plantuml. It lets you describe diagrams (class, sequence, deployment) with text/algebra. Plantuml works well up to a point where the diagrams becomes to complex for the layout algorithm to do well.
I used to think of this as an annoyance, but now I think of it as a feature: It is a way for the universe to tell me that the model is becoming too complex. The layout algorithm serves as a proxy for everyone else that should parse the diagram, and if I can make the diagram better by simplifying, so be it.
Now, a human can do diagram layout better than plantuml, so a human can easily concoct diagrams that are both more complex and better looking than plantuml, but it is my firm belief that this usually not a good thing: It more often than not means that the message is lost in the complexity of the diagram.
Keep it simple!
At the moment, codeatlas is just the static gallery, but we're only a few weekends away from releasing a Github action that deploys this diagram on github pages for your own repos - if you're interested, feel free to watch this repo: https://github.com/codeatlasHQ/codebase-visualizer-action
OP, how close is this to what you had in mind in your question?
EDIT: fixed broken links :o
I feel that people have just embraced Agile blindly and simply forgot about basic modelling.
Ultimately, you just create and endless amount of complex work that keeps developers continuously busy. On the bright side there's a never ending amount of tedious work wrestling systems back into some manageable form, on the downside that work is miserable, in my opinion because much of it can be removed whth proper planning. At some point, expectations eventually meet reality no matter how many developers management burns through, at some point it's clearly not an issue with technology, it's an issue with approach and project management. By that time the organization has had enough turnover in those above and below those pushing agile that those issues too can be hand waived away and the cycle repeats.
Very Good Point!
If you only look at things piecemeal and never holistically, the need for modeling and corresponding tools decreases.
For instance, I frequently build small paper diagrams of different code paths through a component and nearly always find leaky abstractions, mixed layers of abstractions, weird cyclical dependencies, etc. etc. and there really is no clear way to diagram this. Instead, you sort of have to make judgements and assumptions to make the diagram concise and understandable; the sort of decisions that machines just aren't that good at.
On the other side, when code is simple and easy to follow then the pay off of building a diagram just isn't there.
Exactly. I've found that there is no perfect tool and I'm closing in on 40 years of writing software. I use a blend of UML, IE (James Martin's Information Engineering), etc. Anything that makes it "easy-to-grok". Draw.io generally does what I want it to do; there is no feedback loop back to the codebase. It's really just to help onboard new helping hands.
The closest I found that solves this problem is https://c4model.com/ but you still need the code to turn your code into these markups. Can this be well inferred from code alone without framework specific interpreters? I doubt it.
And then you still need a frontend to zoom and navigate the ridiculous amount of hierarchy found within any modern software architecture, e.g microservices.
It also doesn't help microservices patterns also prescribe that you don't share repositories or code. So now you also need to pattern match untyped references across these codebases.
This is a lot of convention and tooling that I'm not sure exists.
Edit: and this is before even getting into version control and reconciling the target->as-is iterative loop.
People do want it (contrary to the common HN refrain of “well, I don’t want it so clearly nobody wants it”). We’ve had customers where I work specifically ask for these kinds of tools. They’re just harder than they seem to write, not only for the parsing reason I mention above. For many codebases you see a giant ball of spaghetti if you look at the full graph, or the layout algorithm gives you something gigantic and hard to browse. That’s a deficiency in graph visualization tools: again, a hard problem with little good tooling out there.
I’d love to see more work in this area since there do exist people who see value in it, contrary to the skeptics.
What you need is the ability to filter the graph. Narwhal and the nx mono-repo toolset has a pretty cool dependency graph feature built in. Here's a video of how they use it:
But, of course, it turns out, someone still needs to understand and be able to debug all the nuances that makes complex logic systems complex, especially when they're cobbled together from many underlying systems.
The real goal should be to take good programmers and magnify what they can do. But since the industry bought so hard into the naive vision, the industry is behind where it should be on a smarter vision.
The industry couldn't get the simple one of those (diagram -> code) to work well enough, how would they do the much more complex reverse?
Architectural and even library nuances can't be easily quantified to a particular UML symbol unless you go through the effort of classifying it for every library/design pattern, and keeping that up to date as well.
Diagrams are sometimes unnecessary overhead early in a project. Sometimes I’ve used them and seen other people use them for initial design planning, especially if management needs to be involved or approve the plans & schedule. But by a year later, the design has grown and changed, and everyone on board is so familiar with the code, but also so pressed for time and feature delivery, that making diagrams doesn’t make sense: nobody involved at this point needs them. Two years later, when the code is getting complicated and slowing down, and you’re onboarding some new people, that’s when it might help to sketch the flow of code.
FWIW, sometimes a good profiling tool will show you and let you explore call stacks, call graphs, execution charts, etc. I often reach for a profiler when I’m new to a codebase. Flame charts are a fave of mine. You can find flame charts in Chrome’s debug tools, or in compiled language profilers like vtune or valgrind. Here’s a decent article on how to use them https://www.brendangregg.com/FlameGraphs/cpuflamegraphs.html
Another issue is that well designed code bases diagram themselves by their module structure, while diagrams for poorly designed code bases may not help understand them at all. When code has too many side effects, or things are poorly or misleadingly named, when class boundaries aren’t well defined or the code has a lot of spaghetti, diagrams might not really help.
IMO two things worth doing are: get a mentor in any new codebase any time you can, and 2) start building your own arsenal of code diagramming tools, rather than wondering why or waiting for others to do it. Demonstrate the value of diagramming code to people around you and see if you can get it to catch on.
IMO there’s no real substitute for getting the design explained by the people who designed it. (Of course this isn’t always possible, but when it is, take advantage.) Automated tools can never prioritize the explanation nor summarize what parts are critical or tricky vs what parts are incidental or trivial, and they can’t tell you which parts should be redesigned because they were slapdash vs which parts look ugly to a newbie but have a long list of hard to see requirements, and touching things should be done with extreme care. Good diagrams are very helpful, but best used as a supplement to in-person stories, in my opinion.
To answer your question, people do use various tools to extract Class Hierarchies, Call Graphs, Cross-Reference listing etc. The other HN thread that you have linked to contains some details. Lots of people do use them. You can easily add Doxygen/CFlow etc. to your make files to generate the diagrams during every build. The key thing for usage is that do not try to comprehend the entire system as a whole (all but impossible for large systems) but localize your study to a module at a time. Once you have the different pieces mapped out, you can combine them by hand.
PS: Came across the book Software Visualization - Visualizing the Structure, Behaviour and Evolution of Software by Stephan Diehl which seems to provide a good survey of the field.
The main problem with low-level code visualization was that it did not add much to the well-formatted code representation in most cases. As for the high-level architecture extraction tool, which is more close to the question in the article, many links on the diagram do not just involve header inclusion, module import, method calls etc that are relatively easy to extract (not without its own challenges with virtual and indirect calls though). Users wanted to see Inter process communications (socket, queues, pipes, http connections) and extracting those is an uphill battle though we introduced some of it (lots of custom, platform specific code). Between this and knowing which connections are important and which are less so, automatically extracted diagrams were of limited value.
Let me know if that works in your scenario :)
After trying various diagramming tools and dragging around boxes and lines, I settled on PlantUML which makes diagrams much much easier to create and modify. It cuts out a lot of the pain of diagramming with the mouse which means there is less resistance to creating diagrams and I do it more.
To your question, "Why isn't diagram generation automated as part of the build process" - one thing I've found that would be difficult to solve is the level of detail you need in the diagram. For instance, in a very complex system with many decision branches, a diagram with every branch would not be helpful. There are cases where I want a high-level component overview but don't want to clutter up the diagram with lots of details. And yet there may b cases where I do want some more detail but may be only in a certain section of the code. I think this judgement of detail tradeoffs is what would be the hardest problem to solve for diagram generation tools. You want enough detail to be useful but not too much to be overkill.
Generation of legible diagrams could be accomplished on a domain or framework basis where code is subject to local patterns and can be structured "for" generation. We see this with things like OpenAPI schema generation.
Ultimately I think diagramming isn't prioritized because diagrams themselves aren't that valuable. They're just a medium for the actually valuable thing: high level representations.
I want this diagram from my code because it’s a simple way to understand the system, ok then you better refactor your code to match that simple abstraction…
I vaguely recall Visual Studio has this option where you can generate some sort of class diagram. It looked like shit the last time I used it (~2019) especially as your classes get more and more functions built into it. I also can't imagine how shitty it looks for codebases that have a significant coupling problem.
Furthermore, creating a UML diagram is a documentation process rather than something that should be automatically built in. I put it on the same level as writing a document in a word doc or something that's done as the project gets closer to being finished. Some places can live with it, a lot of places (actual software companies) probably do not as they move unreasonably fast (Agile) which does not even allow time for documentation or they just purposely neglect documentation.
> Why aren't code visualization tools more popular? The options out there seem outdated
Because they look like shit. I tried mermaid with markdown, I was not happy with the results, I tried plantUML back in 2019, I hated how it ended up looking, I hated how I have to install java for it, and I gave up on it pretty quickly.
The only code visualization tool I ever use is either draw.io or MS Visio. At lease there's a plugin for that for VS Code.
> Would you want to use these tools? What would be your ideal tool?
Markdown with vim option. It also must have an option to force a top-down flow approach and not freaking forcing it to be a left-right layout
That's the point, right? Visually representing the complexity of the system. I've used IntelliJ to do this before to show why modifying certain behavior was so slow and error-prone. In that case there were 3-4 classes with heavily overlapping functionality because, surprise, in the past there were multiple teams contributing to the same codebase that all did their own thing.
With a manually constructed diagram, I have leverage to handwave irrelevant details away.
Perhaps to compare with documentation: it's easy to automatically describe things like types, and maybe callgraphs, but there's value in having prose which explains details about the interface which the program's type doesn't reveal. - With diagrams to visualise a system, the significance (or incidental nature) of the relationships may be hard to pick automatically.
Most of a system is either uninteresting or trivial. You need someone to tell you where the interesting part is.
This is very hard. And since it's hard, it's not automated. And since it's not automated, it goes out of date very quickly. I think that's the fundamental issue: keeping around evergreen documentation is a lot of overhead. There is no connection between the code and the diagrams, so it's too easy to change the code and not realize that the diagram needs to be updated too.
Another thing is that it's really the most useful for new members. If you've been working on the infra for a while, you already know the structure and you don't need the diagram. So teams tend to just avoid the diagrams altogether.
https://docs.microsoft.com/en-us/visualstudio/modeling/map-d...
In many cases it might really be faster and easier to just diagram things out with a pad of paper and a pencil compared to setting up a tool like this and getting all the parts working correctly without any bugs.
That said, a virtual reality 3D tool for visualizing code base dependencies, internal structure, what parts call what other parts, internal exception handling etc. would be pretty cool. Maybe it's an area where AI machine learning could do something.
This may help with adoption.
But more seriously, it depends on how complex the system is and how it’s modelled. The case I was working with then transferred excellently to such diagrams (shallow and deep inheritance, and other forms of composition and linkage, with every box a link) and key-value property sheets about the types and the likes, but I don’t think I’ve encountered another system where anything even vaguely like that would work particularly well.
I keep bookmarking threads like this one, but I haven't found anything useful for me. The closest one was SourceTrail, which is unfortunately not developed anymore.
With so much hiring and onboarding going on, I am surprised there isn't a market (or an offering) for such tools.
NDepend Dependency Graph https://www.youtube.com/watch?v=23fBxM2v22k
But not everyone thinks about designs as 2-D pictures. I'd argue there's selection pressure that favors programmers being good spellers and symbol-manipulators. Some people find it intuitive to think of concepts as boxes and the relationships between those concepts as lines. Or maybe to think of a 2-D grid of actions/dependencies with deliverables on the vertical axis and work-steps on the horizontal.
But not everyone likes these kinds of visualizations. Some prefer a text-based description. And some designs don't fit any obvious pictorial representation.
I've had a related thought/desire percolating... roughly: I wonder what interesting levers we could build if it was normalized (for both toolchains and projects) to create and publish the plaintext relationship graphs in a common easily-reused format.
I'll self-reply to elaborate a bit.
There's a devilish sub-problem, which is that any given executable can potentially exec arbitrary arguments. For now I handle this with a very crude automated binary/executable analysis that needs to be augmented by human source analysis. Deep multi-language source analysis wouldn't be very scalable, but I suspect fairly-standardized structural annotations could improve the results in a scalable way.
I have to imagine there are other applications of the same information.
I wouldn't use these tools anyway, to be honest. They have some limited utility when constrained to small components/parts of an application (e.g., self-contained libraries), but for understanding systems as a whole there is too much to have effective reverse-engineering into a visualization (in my opinion).
In most complex systems the part where the magic happens is likely impossible for a tool to identify so would get lost in the noise of the cruft around it -- even for monoliths using frameworks and especially for anything distributed across microservices, and it's usually that aspect that is of most interest
It's another thing that can break, another element that needs to be maintained. In my experience there are very few pieces of code that will be able to run indefinitely without ever being updated, fixed, or re-examined at some point. The cost of adding more processes is not one-time, and it can be difficult to figure out what the time bounds are.
- Why aren't code visualization tools more popular? The options out there seem outdated
People who are interested in the structure of code are typically engineers, capable of writing and reading the codebase of interest. A UML diagram may be a way to understand an element of the system, but things such as in-line comments in the codebase itself are often more instructive on structure and function.
- Would you want to use these tools? What would be your ideal tool?
When I was in high school, if I didn't want to read, say Crime & Punishment, I could buy the Cliff's Notes version, and get a chapter-by-chapter summary of major characters, events, and literary techniques. In many ways, it contained all of the information of the book without the substance.
But importantly, it took significantly less time to read and fully process than the book, while being written in the same language. In code, it is already extremely easy to look thru header files, or collapse every function in your IDE to get a high-level overview of what data and methods exist. You can then dive in immediately to anything you would like to understand better ("what does 'UpdateSignificanceValue' really mean") and there's no mental overhead in translating from an encoded diagram into whatever your mental model is. This is why I do not personally see value in code visualization -- outside of notes I take that are relevant to any specific problem I am working on.
As someone who is not a very visual person at all, I found it really nice to use to make my design docs more comprehensible to visual learners. I've gotten good feedback about designs every time I've used the tool.
Fundamentally I think that the useful kind of software or system diagrams are always abstractions of the actual code. Figuring out the correct abstraction for the intended purpose requires either experience or a lot of trial and error. It may be possible for very specific applications, but I kind of doubt there is an algorithm to generate the content for a useful system diagram from the raw code.
Then there is the problem of rendering and layout out the diagram automatically. We have Graphviz and Mermaid, and probably others I haven't heard of, and while these do an okay job, I've never found their layout algorithms to be particularly great.
Overall, I don't think anything is going to be as useful as a manually drawn diagram, made with a specific intent in mind.
In my experience these tools generally exist to facilitate bikeshedding. The academic nature of UML makes it pretty useless in the real-world.
Something that could be useful is having a tool that uses knowledge about code to help you build a mindmap (but does not just puke the whole thing out). Huge bonus points for allowing the user to create late-bound relations and conceptual boundaries. Finally, one of these tools should be able to compare its output with the source, and indicate what has changed (through deletion/addition, or via VCS diff).
A visualization can be helpful as an artifact for non-technical colleagues, but I always end up hand-rolling those diagrams to highlight a specific aspect of the system and hide irrelevant features.
It has been used for database schemas, game story creation, cutting up sprites among other things...
Lately I made my own node database so I don't need this tool any longer, but I'm sure it will prove useful eventually again!
Maybe you'd need a three dimensional model (really it's likely n-dimensional/hyperdimensional), 2D might not be enough.
Programming models get so convoluted with regards to state and interactions, both in-memory/in-process state and the stored state in databases/files.
Jurassic Park's 3D filesystem was a pie in the sky idea, what, 30 years ago? Holy crap it was 29 years ago or so. We've had REVOLUTIONS in 3D processing and games, and never even stratched the surface of basic 3D visualizations of code or data or filesystems or machine networks or the like.
And then even if you represent a diagram, it's useless without time visualization/traces, as kind of referred to by the RR debugger post. So for active code, you'd need simulation or actual run data to show what it does visually to be effective.
Really what's being dealt with here is probably related to theory of computation, and various results like the undecidability of the halting problem. The halting problem shows that even for very basic languages that are minimally Turing complete, the complexity shoots VERY QUICKLY to massive degrees of infinity/uncomputability.
So some catch-all visualizer for even general classes of Turing complete languages is probably impossible.
Maybe something like "this is a java spring app with well regimented separationg of data/domain classes and service classes"...
Even then once you get to database persistence ... wow.
And the amount of data you'd need to store for test runs.
Spring + TDD enforces a certain simplicity to a codebase, so perhaps you could make effective classes of visualization and tracing/replay visualization for that.
But it is telling these tools don't really exist, and attempts like UML were largely abandoned.
Compare that to something like a call graph, or a module dependency diagram. The last will be more complete, but will convey *much less* information than the later.
This varies with technology, some will be more friendly than others to this kind of tool, I think that the more dynamic, the worse, but even in very static and consistent language, I would not bet on any tool being better than the brain's parser for a long time.
I was so happy not being the one doing this useless task..
I find creation of a sequence diagram with class instances as columns and method names as arrows can help visualise things.
In the future mainstream languages will probably have annotations to describe "role" of various things, specifically to enable diagram generators.
Also I have rarely seen diagrams generated from code, the main exception being database ERDs ("reverse engineering.") Usually, those diagrams are also a mess.
Also, I almost forgot to mention: with "Agile", there usually is no design process. We'll just "fix it in the next sprint."
https://www.google.com/search?q=Prograph+CPX&rlz=1C5CHFA_enG...
You'd have high level classes, and low-level nitty gritty. You could edit your code as it was running (and then continue).
People prefer text (weirdly).
That said, diagrams can either be rare, focused, and useful; or common, unfocused, and distracting. Automated processes tend to generate the latter.
Edit: another way of thinking about time is mutability so perhaps functional languages are more amenable to graphing.
http://logan.tw/posts/2015/03/10/trace-source-code-with-vim-...
Honestly, freehand diagrams are best. You’ll exercise your own understanding of the code base as you draw.
One reason it is slower is that it is is difficult to create a map like diagram where you zoom in to get greater detail.
dot graphs are popular with many tools but often barely or not interactive at all.
I want to see the big picture, what they can generate are direct translations of the code down to the line level.
Sticking to the spirit of the question, I will categorize the tools mentioned in the comments. They solve very different problems, so the "ideal tool" absolutely depends on the task at hand. I think this rough categorization can help to select the right tool for the job.
The first class of tools are drawing tools. These tools aim to aid design and communication. plantUML http://www.plantuml.com/ is best suited to manual use. I use plantUML when sketching out designs. It's great for thinking about and communicating state machines, system architecture, and server-client or multiprocess interactions. The problems solved by plantUML can also be solved by general GUI diagramming software like Visio or LibreOffice Draw. I'll not dawdle on the pros and cons.
A second class of tools are static analysis tools. These are more of a microscope than they are a sketch pad. They operate on existing code and the user input comes as filters or options to get the desired level of detail. The examples here all produce graphviz .dot files:
bazel query for dependency graphs - https://bazel.build/docs/query-how-to
Doxygen C call graphs and struct inheritance graphs - https://www.doxygen.nl/manual/diagrams.html
radare2 generates call graphs from an .elf file - https://reverseengineering.stackexchange.com/a/9120
A DBC visualizer (CAN bus protocol specification file) - https://github.com/driftregion/dbcview (mine)
These tools are great not only for probing the depths of an unfamiliar codebase but also because they give insights to people already familiar with that codebase. For example: statically generated callgraphs showed that some debugging hooks had been left in. They've also shown duplicate code paths.
A third class of tools mentioned in the comments are dynamic analysis tools: profilers, tracers, debuggers. These are the oscilloscopes and signal analyzers of software engineering. These seem to be outside OP's query.
> Why isn't diagram generation automated as part of the build process (UML or otherwise)?
The output of class 1 and 2 tools is used in builds. Doxygen supports plantUML (class 1) and callgraph generation (class 2). Class 1 is obviously manual, and class 2 is also manual for reasons that others have mentioned here: namely adjusting to get the right level of detail. The text output of class 3 tools is often used in CI as a pass-fail indicator.
> Why aren't code visualization tools more popular? The options out there seem outdated.
My anecdotal experience is that these tools are specialized and therefore have a small audience. They solve problems related to design and analysis of software which is a small part of "real world" (sorry) software development.
edit:formatting
