design of gluehow and why |
|
A small, highly useful bag of tricks.
This API must be compact, not complete. This is the one API that should fit comfortably in someone's head. Adding rarely used or too specialized functions is thus harmful.
Every function should come with compelling real-world examples in its defense. In fact, the library should only be permitted to grow from real-world examples which should be used to document and justify the inclusions.
The idea is to find the most popular, familiar and short names for each and every function (no underscores and no capitals). Python gets this right, so does UNIX. A function with an unheard of name or alien semantics will be avoided. People rather recall known names/semantics rather than learn unfamiliar new names/semantics.
For non-obvious O(n) operations, use verbs to emphasize the process rather than the result, eg. scan(t,e) not indexof(e,t), count(t) not size(t).
They follow the general api-design rules.
Don't provide data structures like list and set in a glue library, or a way to do OOP. Instead just provide the mechanisms as functions working on bare tables. Don't do both either: if your list type gets widely adopted, your programs will now be a mixture of bare tables (this is inevitable) and lists so now you have to decide which of your lists has a sort() method and which need to be wrapped first. A global glue.sort(t) spares you the trouble.
String lambdas, callable strings, list comprehensions are all fun, but even the simplest DSLs have a learning curve, complex semantics and performance corner cases, and become a barrier to reading other people's code. I remember loathing having to use Lua's unfamiliar regex language with non-linear performance to do all of my string chores.
Adding DSLs make reading code frustrating. Why not give them an identity of their own and ship them as individual libraries and see if they can claim the required learning time from programmers by themselves.
Don't add shortcut functions except when calling the shortcut function makes the code more readable than the equivalent Lua code.
If something is an [tricks idiom], don't add a function for it, use it directly. Chances are its syntax will be more popular than its name. Eg. it's harder to recall and trust semantic equivalence of isnan(x) to the odd looking but mnemonic idiom x ~= x. That doesn't mean a < b and a or b is an idiom for min though, min itself is the idiom as we know it from math (wish there were a sign() function too).
Functional programming sugars like compose and bind makes code harder to read because brains are slow to switch between abstraction levels unless it's a self-contained DSL with radically different syntax and semantics than the surrounding code. Eg. it's relatively easy to read a Lua string pattern or an embedded SQL string more than it is reading expressions involving bind and compose which force you to simulate the equivalent Lua syntax in your head.
Sugars like "message %s" % arg are the good ones: % is odd enough to put after a string constant that it has an idiomatic quality, and its semantics is self-evident by reading the format string literal, even for someone who never heard of python's % operator. Also, a prefix notation is generally more readable than a function call.
Keep the code readable (didactically when possible) and compact. Code changes that compromise these qualities for optimization should come with a benchmark to justify them.
Document the limits of the algorithms involved with respect to input, like when does it have non-linear performance and how it is stack bound. Performance characteristics are not an implementation detail.