I started writing this as an answer to someone on some discord, but it would not fit the channel topic, but I'd still love to see people's views on this.

So I'll quote the comment but just as a primer:

The safest pattern to use is to not use any pattern at all and write the most straight forward code. Apply patterns only when the simplest code is actually causing real problems.

First and foremost: Many paths to hell are paved with design patterns applied willy-nilly. (A funny aside: OO community seems to be more active and organized in describing them (and often not warning strongly enough about dangers of inheritance, the true lord of the pattern rings), which leads to the lower-level, simpler patterns being underrepresented.)

But, the other extreme is not without issues, by far.

I've seen too many FastAPI endpoints talking to db like there's no tomorrow. That is definitely "straight forward" approach but the first problem is already there: it's pretty much untestable, and soon enough everyone is coupling to random DB columns (and making random assumptions about their content, usually based on "well let's see who writes what there" analysis) which makes it hard to change without playing a whack-a-bug.

And what? Our initial DB design was not future proof? Tough luck changing it now. So new endpoints will actually be trying to make up for the obsolete schema, using pandas everywhere to do what SQL or some storage layer (perhaps with some unit-of-work pattern) should be doing -- and further cementing in the obsolete design. Eventually it's close to impossible to know who writes/expects what, so now everyone better be defensive, adding even more cruft (and space for bugs).

My point is, I guess, that by the time when there are identifiable "real problems" to be solved by pattern, it's far too late.

Look, in general, postponing a decision to have more information can be a great strategy. But that depends on the quality of information you get by postponing. If that extra information is going to be just new features you added in the meantime, that is going to be heavily biased by the amount of defensive / making-up-for-bad-db junk that you forced yourself to keep adding. It's not necessarily going to be easier to see the right pattern.

So the tricky part is, which patterns are actually strong enough yet not necessarily obtrusive, so that you can start applying them early on? That's a million dollar question.

I don't think "straight forward" gets you towards answering that question. (Well, to be fair, I'm sure people have made $1M with "straight forward code", so that's that, but is that a good bet?)

(By the way, real world actually has a nice pattern specifically for getting out of that hole, and it's called "your competitor moving faster & being cheaper than you" so in a healthy market the problem should solve itself eventually...)

So what are your ideas? Do you have design patterns / disciplines that you tend to apply generally, with new projects?

I'm not looking for actual patterns (although it's fine to suggest your favorites, or link to resources), I'm mainly interested in what do people think about patterns in general, and how to apply them during the lifetime of the project.