The Start, Part 1

I was going to start this article differently. But today, as I’m writing these lines, a new ticket about Proxy has been opened on Rakudo issue tracker. And once again I was amazed how coincidental some coincidences are! There was also some good about the ticket: it gave me the subject I was missing for a few days to make this article accomplished.

The Start, Part 2

So far the mainline of this series was focused on the metamodel and MOP. Things are likely to remain this way until the end of the project, but I can’t avoid interesting topics from other language areas. This time the plan is to elaborate on Proxy container, mentioned earlier in Containers And Symbols.

The Purpose

The official documentation for the class is rather succinct but you should read it first unless have done so already. My own definition of Proxy purpose winds down to the following: it provides means of controlling and manipulating objects fetched from or stored into a container. To start with, I would borrow an example from the documentation:

sub double() is rw {
    my $storage = 0;
    Proxy.new(
        FETCH => method ()     { $storage * 2    },
        STORE => method ($new) { $storage = $new },
    )
}
my $doubled := double();
$doubled = 4;
say $doubled;           # 8
say $doubled.VAR.^name; # Proxy

The last line is added to the example to make it clear what is different comparing to the following code:

my $doubled = 4;
say $doubled.VAR.^name; # Scalar

The Use

You likely have noticed already the use of anonymous methods as values of Proxy constructor profile keys in the above example. The reason for this approach is Proxy instance passing itself as the first argument to the FETCH and STORE code objects.

Wether you knew it already or not, but a method is an instance of Method class which inherits directly from Routine. And a sub is an instance of Sub class which also inherits from Routine. Other than been of different classes they share almost every aspect of their implementation. What actually tell one from another is that method’s first argument is implicit and it’s the object the method is called upon. Apparently, the argument is what self lexical is then bound to. Thus, method and sub are more about the way of using a code object than about what the object is. A good demo for this would be this code:

sub foo($self) {
    say $self.WHICH;
}
"42".&foo

Here foo is invoked as a method of a string object "42". The .& invocation syntax is almost equivalent to the foo("42") form where "42" is implicitly passed on to foo. The snippet can be rewritten to use a method declaration:

my &foo = method { say self.WHICH };
"42".&foo;

The result is the same except for one less parameter in &foo signature.

Getting back to Proxy, use of anonymous methods makes FETCH and STORE code objects behave as if they’re methods of a Proxy instance. Does it provide us with any advantages? Not really. Partly, because Proxy is a very simplistic class with no really useful public methods but new. Partly due to specifics of Proxy behavior I’ll talk about later in this text.

It can make a little difference in a case of subclassing:

class FooProxy is Proxy {
    has $!storage;
    method foo { say "foo" }
}

But the thing to be always remembered is that we’re dealing with a container here. Containers are highly transparent entities. Thus, even though self points back to a Proxy instance, any attempt to invoke a method on it would result in reading from the container and, consequently, in calling FETCH and eventually invoking the method on its return value:

my $p := FooProxy.new(
        FETCH => method () { 42 },
        STORE => method ($v) { say self.foo; }
    );
$p = 12; # No such method 'foo' for invocant of type 'Int'

For this reason I would discourage anyone from using self within FETCH or otherwise the routine is very likely to end up with an infinite recursion.

There is a legit question though: how does one access the method foo of FooProxy from above? Apparently, via self.VAR.foo as we normally do with containerized stuff. Also worth noting that within methods of FooProxy it is totally safe to reference attributes with $!storage syntax. Yet, any direct use of self brings us back into the pitfall of using the container itself; even if a private method is invoked with self!private() despite of the invocation syntax akin to referencing a private attribute.

And to get over with this subject, for those of you who doesn’t like wordy code, FETCH and STORE are not obliged to be method declarations. After all, Raku is about TIMTOWTDI. The uniformity of internal conventions let us have things done differently from syntax point of view while preserving the semantics:

Proxy.new(
    FETCH =>            { $storage * 2 },
    STORE => -> $, $new { $storage = $new },
)

or, if we’d need to to use the FooProxy class:

FooProxy.new(
    FETCH => { $storage * 2 },
    STORE => { $^a.VAR.foo($^b) }
    # STORE => -> \proxy, $new { proxy.VAR.foo($new) }
)

Though I personally not a big fan of using self-declared formal positional parameters this way. Pointy blocks are my favorites most of the time, as shown in the comment.

The Pitfalls

Proxy is expensive. Or, rather: it is expensive. At the very beginning of this article I have mentioned a ticked. It is about FETCH been called multiple times under certain circumstances. Unfortunately, I expect same kind of tickets to be opened again and again in the future. Sometimes Raku’s real complexity is so well hidden behind the curtain that it causes a confusion in inexperienced developers. Let’s see how it happens with an example:

my $p := Proxy.new(
        FETCH => { say "FETCH"; 42 },
        STORE => -> $, $v { }
    );
say $p;

Do you find anything criminal in this code? Likely not. Until it’s been ran:

FETCH
FETCH
FETCH
FETCH
FETCH
FETCH
FETCH
42

“Wow, wow, slow down!” – isn’t it the first impression one gets when see this for the first time? I was there, felt the same! Back then my thinking was that one day I’ll grow up, become a big and strong Raku developer and get it fixed! Not that I’m so much grown up by now, but at least I know that the problem is not about too many calls to FETCH been made but in our assumptions.

At the first glance say $p looks like a perfectly atomic operation. But, in fact, it is not. Far from it! The argument it takes is actually gets passed through a number of stages. To mention just a few:

  1. say is a multi sub. In order to determine what candidate to choose Raku needs to know the types of its arguments. Therefore it reads from $p to find a candidate with matching signature.
  2. to stringify $p say invokes method gist on it.
  3. gist itself is a multi, so we get the dispatcher involved again.
  4. etc.

Depending on a chosen candidate, an argument could also become the subject for testing by additional parameter constraints like definedness or alike. Apparently, this would result in extra calls of FETCH. And, in fact, there is no way we could optimize away all those reads. Once we consider the situation carefully it becomes evident that it is the dynamic nature of Proxy which doesn’t let us to skip a read or two; or even eliminate almost all of them altogether reducing the total count to one. Because sometimes the side effects of FETCH or STORE are what really matters. The most simple case I can come up with is when we need Proxy exactly for the reason of counting the number of references to a variable in our code.

Contrary to the static Scalar which would always return same value until is assigned with a new one, Proxy is unpredictable about it’s value too. Therefore the compiler can’t make any assumptions about it and ought to generate a call to FETCH for any reference to the container. Maybe at some point a way to hint the compiler about possible changes in the return value would be invented and implemented and this would make additional optimizations possible. Though such hinting would only make sense under the obligation of producing no side-effects by the FETCH/STORE code. Until then we should just keep in mind this peculiarity of proxies.

Does it make proxies harmful? Yes, sometimes. And yet, sometimes they’re indispensable and even irreplaceable. It means just one thing: to use or not to use a proxy should be carefully considered.

The above mentioned ticket refers to a situation where a Proxy masks a network operation behind it. While being rather critical performance degradation cause if used in a straightforward way, it could be very handy if the object it gets from the network is cached. In this case it would even pay back by speeding up operations while keeping the code tidy by hiding something like $server.fetch('key_foo') behind a simple $key_foo variable.

Another approach would be implementation of lazy operations where a symbol could be bound to a Proxy and, when and if eventually referenced, get its value and then re-bound to either the value itself or to a Scalar containing it. In this case we only pay the performance price once in the life time of the symbol. And if obtaining the initial value is an expensive operation on its own then use of a proxy would pay back again by postponing it to the moment when it is really necessary possibly sparing on code initial setup time; or even eliminating the initialization of the symbol altogether if it remains unused.

The End

It is really amazing how much does it take to describe such a simple class as Proxy! Did you know that its implementation is just about 33 lines long including a comment? But to comprehend the full power of the concept it takes us to the neighboring areas of the Raku language. This pattern I often find myself following to: no matter what subject is being studied, a couple of adjustent subjects would get involved. Consider it another remarkable feature of Raku: it would hardly let one get bored easily.


I would be very thankful for any report about errors found on this page!

Updated:

Comments