Earlier quoted context omitted.
> we've still avoided adding another potentially crappy dependency to our project. No, you've just added the crappy dependency == your off the cuff implementation (which is even more work since you have to handle issues and unit testing for it too). > What the hell sort of library would you crack out to do a breadth first search traversal? In Python I would use networkx probably, unless there was a good reason I coul…
In the case where you'd require a bfs to traverse a graph, you're already doing something that is algorithmic, and will otherwise be prohibitively costly if explored in a dfs manner. The most complicated part of this algorithm is highly unlikely to be the bfs itself, with it's highly likely 10 lines of code in total. In this case, someone who didn't have the time to visualise exactly how the bfs is operating shouldn'…
Yes, this kind of hubris is often the root of the problem when so-called 'simple' algorithms get homebrewed, then later on unexpected corner cases pop up, testing isn't adequate, you didn't implement it in an extensible way, your implementation is unacceptably inefficient, etc. etc.
> On top of this, who writes this code and doesn't test that it does what you expected it to do on sample data. If you bfs implementation was bust, then you'd miss nodes or revisit them, which becomes immediately apparent from testing the algorithm you're writing.
Yes, but the point is that testing isn't free. If you commit to writing your own implementation of something to interface with whatever larger business / domain-specific project created the need for it in the first place, then you have to dedicate not just the time to write it, but also the time to test it, and to test how it integrates, and to maintain the API for it and maintain its integration (possibly backward compatibility, etc. etc.)
If you adopt a standard to use a certain library, then up to the degree to which you trust the library, those problems are mostly offloaded to others. For things like classical algorithms, the libraries are extremely trustworthy, so many of the pitfalls of adding a dependency don't apply.
> In this case, someone who didn't have the time to visualise exactly how the bfs is operating shouldn't be writing the algorithm.
It's not about being too pressed for time. Even if you had all day, it's a poor use of time to create more future work for yourself by yoking yourself to all the extra responsibilities that come along with rolling your own. Yoking yourself to those responsibilities should require an exception reason for doing so.
> Saying writing it yourself must be an exceptional circumstance sounds about as crazy to me as saying no-one should ever use for loops, they expose you to making mistakes with an index you may forget is zero-based, please use anonymous iterators instead.
This is an extremely fallacious comparison. Basic control structure are built in features of the language. You don't have to test or maintain the runtime execution of a for loop, that's the job of language designers. Making mistakes in the usage of something (like an off-by-one index) is completely non sequitur to this entire discussion. I don't see how you would think that type of bug is related to anything I'm saying.
In general, you also assume a much higher fidelity of testing than what happens in the real world. In 99% of software jobs, you'll be extremely lucky if someone even documents their homemade BFS algorithm, let alone writing even the most superficial of tests. Expecting them to give you an adequate bank of automated tests is like believing in the Easter Bunny.
> Try writing bugs into this, it's actually difficult to screw up.
Ugh.