Middleware Redux
In the Guide, we introduced middleware as a stack of functions. While it is not wrong that you can use middleware in this linear fashion (also in grammY), calling it just a stack is a simplification.
Middleware in grammY
Commonly, you see the following pattern.
const bot = new Bot("<token>");
bot.use(/* ... */);
bot.use(/* ... */);
bot.on(/* ... */);
bot.on(/* ... */);
bot.on(/* ... */);
bot.start();
Looks pretty much like a stack, except, behind the scenes, it really is a tree. The heart of this functionality is the Composer class (reference) that builds up this tree.
First of all, every instance of Bot is an instance of Composer. It’s just a subclass, so class Bot extends Composer.
Also, you should know that every single method of Composer internally calls use. For example, filter just calls use with some branching middleware, while on just calls filter again with some predicate function that matches updates against the given filter query. We can therefore limit ourselves to looking at use for now, and the rest follows.
We now have to dive a bit into the details of what Composer does with your use calls, and how it differs from other middleware systems out there. The difference may seem subtle, but wait until the next subsection to find out why it has remarkable consequences.
Augmenting Composer
You can install more middleware on an instance of Composer even after installing the Composer itself somewhere.
const bot = new Bot("<token>"); // subclass of `Composer`
const composer = new Composer();
bot.use(composer);
// These will be run:
composer.use(/* A */);
composer.use(/* B */);
composer.use(/* C */);
A, B, and C will be run. All this says is that once you have installed an instance of Composer, you can still call use on it and this middleware will still be run. (This is nothing spectacular, but already a main difference to popular competing frameworks that simply ignore subsequent operations.)
You may be wondering where the tree structure is in there. Let’s have a look at this snippet:
const composer = new Composer();
composer.use(/* A */);
composer.use(/* B */).use(/* C */);
composer.use(/* D */).use(/* E */).use(/* F */).use(/* G */);
composer.use(/* H */).use(/* I */);
composer.use(/* J */).use(/* K */).use(/* L */);
Can you see it?
As you can guess, all middleware will be run in order from A to L.
Other libraries would internally flatten this code to be equivalent to composer and so on. On the contrary, grammY preserves the tree you specified: one root node (composer) has five children (A, B, D, H, J), while the child B has one other child, C, etc. This tree will then be traversed by every update in depth-first order, hence effectively passing through A to L in linear order, much like what you know from other systems.
This is made possible by creating a new instance of Composer every time you call use that will in turn be extended (as explained above).
Concatenating use Calls
If we only used use, this would not be too useful (pun intended). It gets more interesting as soon as e.g. filter comes into play.
Check this out:
const composer = new Composer();
composer.filter(/* 1 */, /* A */).use(/* B */)
composer.filter(/* 2 */).use(/* C */, /* D */)
On line 3, we register A behind a predicate function 1. A will only be evaluated for updates which pass the condition 1. However, filter returns a Composer instance that we augment with the use call on line 3, so B is still guarded by 1, even though it is installed in a completely different use call.
Line 5 is equivalent to line 3 in the respect that both C and D will only be run if 2 holds.
Remember how bot calls could be chained in order to concatenate filter queries with AND? Imagine this:
const composer = new Composer();
composer.filter(/* 1 */).filter(/* 2 */).use(/* A */);
2 will only be checked if 1 holds, and A will only be run if 2 (and thus 1) holds.
Revisit the section about combining filter queries with your new knowledge and feel your new power.
A special case here is fork, as it starts two computations that are concurrent, i.e. interleaved on the event loop. Instead of returning the Composer instance created by the underlying use call, it returns a Composer that reflects the forked computation. This allows for concise patterns like bot. A will now be executed on the parallel computation branch.