2020-06-22

Eval Spotted in the Wild

(#lisptips on the dynamic nature of CLOS magnified by eval)

Since starting programming in Lisp, I always had an impression that using eval is a taboo. Or rather, a cul-de-sac that you never want to touch. When I was only learning Lisp, I had a couple of unsuccessful and rather stupid attempts of utilizing it to bend the language to my view of how it should function — only to learn how it is really intended to function. After that, it occupied its rightful place on my mind's shelf of "low-level constructs only needed to implement the language".

Yet, recently, I saw a legitimate use case for it and even wrote a piece of production code containing eval! That was such a revelation that I wanted to share it in this short post.

So, here is the case I needed to solve: I was developing a parser for a new data format that had to fit into an existing set of parsers. The parsers not only decode the data but also store it in the datastore using the CLOS machinery for datastore access. I.e. there's a generic function to store an individual piece of data that is specialized for different connection/datastore types. Now, my parser had to prepare the individual pieces and, eventually, they would be fed to this function. But that may happen independently of the parser operation: when the data store commit is performed.

Yet, there was another issue at play: the data format allows the individual items to be interdependent, i.e. reference one another via an implicit reference. And when the data is persisted, due to the properties of the data store, these references should be changed to the internal ids of the referenced items. And those are not known before the commit happens. I.e. I was in the following situation:

  • my parser produces an array of items that are to be persisted to the dataset at some later time
  • the order of their addition matters as the dependent items should be added after the ones they reference
  • and as the referenced item is added its id should be saved
  • and assigned to a field of the dependent item before that item is also added

This program logic is quite normal, apart from the fact that my parser doesn't have control over the whole workflow. Actually, the data persistence stage operates in the inversion of control paradigm, i.e. I can only override (rather, augment) the part of the program that is responsible for processing an individual item. Needless to say that I had no desire or intention to reimplement all the different (I believe, 7) ways of interaction with the datastore that had their own methods plus a number of before/after/around-methods.

In fact, CLOS is very flexible and provides a way, using an object of my own mixin-class to hold the state and around-method specialized on it, to achieve my goal of fitting into this whole machinery without distracting it or having to reimplement anything. If not for one issue: limited facilities for dynamic creation of classes.

So, here's what I wanted to do and to avoid:

  1. I wanted to define a mixin-class and an around-method for it to augment the data storing procedure with saving of the ids of specified items and assigning them to fields in other items before persisting them. Here's the sketch of the relevant code:
    
    (defclass my-data-store-mixin ()
      ((linked-items-table :reader my-mixin-table
                           :initform (make-hash-table))))
    
    (defmethod add-item :around ((db my-data-store-mixin) item)
      (let ((linked-items-table (my-mixin-table db))
            (item-id (call-next-method)))
        (dolist (it (gethash item linked-items-table))
          (remhash it linked-items-table)
          (setf (reference it) item-id))
        (remhash item linked-items-table)
        item-id)) 
    
  2. Yet, I didn't want this code to run when other data formats are imported, hence my mixin should have been "activated" if and only if my specific format is parsed.
  3. In other words, I needed a way to dynamically add this mixin to an existing connection object, in the context of the parser call, and then restore the connection to its previous state. In general, CLOS also provides such a facility with its change-class operator. I would say, this would have been a manifestation of a dynamic object system in all its glory if not for one deficiency.
  4. You can't just dynamically define a temporary class: the one that will inherit from the class of the current connection and my mixin. defclass is a macro that's expected to deal with names known ahead-of-time and coded as part of its call: it doesn't evaluate variables. Usually, all such APIs in Lisp have a make-function counterpart. I.e. what I needed was:
    
    (let ((temp-class (gensym))
          (current-db-class (class-of *db*)))
      (make-class temp-class (list (class-name current-db-class) my-data-store-mixin) nil)
      (unwind-protect (progn (change-class *db* temp-class)
                             ;; execute my code
                      )
        (change-class *db* current-db-class)))
    
    But CLOS just doesn't have an API for that. (Which might be perfectly reasonable — and I don't want to delve into the discussion of those reasons in this post). Actually, there's MOP for that. But I'd prefer not to take the MOP route here for another set of reasons I want to skip discussing now :) Suffice to say that it is complicated and, from my experience with the MOP, I developed a stance that it's another area intended for language implementation usage — not for user-level code.
  5. And here's where eval comes to the rescue. In place of the nonexisting make-class I could just put this piece:
    
    (let ((class (intern (format nil "my-mixed-~a" (class-name current-db-class)))))
      (when (not (find-class class nil))
        (eval `(defclass ,class (,(class-of *db*) my-data-store-mixin) ()))))
    

So, eval is an escape hatch into the world of ultimate dynamism. This operator can add it anywhere: whether an appropriate API was left out due to lack of foresight or even when it was not intended to exist... :)