Arc Forumnew | comments | leaders | submitlogin
2 points by akkartik 5524 days ago | link | parent

Very cool.

The nilsym is an interesting idea. Why do you prefer it to the nilcache approach in memo?

I liked seeing the invalidate/reset functions; one problem with the current defmemo and defcache is that a cache can grow unboundedly as stale values accumulate. In your case you could actually set cach.key to nilsym to avoid that. I've been thinking about more global policies than deleting individual keys.

Idea #1: Caches have a max size and when it is exceeded you delete the least recently used (LRU) element. Was this what you meant by the FIFO mechanism above? If you have a systems background as I do, this is the first approach you think of. (http://en.wikipedia.org/wiki/Cache_algorithms)

Idea #2: A function to clear the cache. I like this approach because it's simpler than watching what was accessed when, and it fits naturally to the way we do things.

  (defcache parse-fragment A fragment
    ..)

  (defcache check-fragment A fragment
    ..)

  (whilet (document-fragments doc)
    (clearcache A)
    ... ;; lots of stuff happens here every iteration, relying on fragment operations being cached.
I've added a notion of multiple caches; defcache specifies a cache, and clearcache simply clears everything in a cache regardless of which function it's for. Helps group related functions together.


1 point by palsecam 5524 days ago | link

1. nilsym over nilcache: because I prefer to have only one table to store all the cached values. And I suppose it makes the code shorter. And maybe more efficient. And it's a new approach, and I like new things.

2. In your case you could actually set cach.key to nilsym to avoid that. I suppose you mean "set cach.key to (the real) nil (to actually delete the entry)"

3. I've been thinking about more global policies than deleting individual keys. Me too. As I said, this is just a draft. For instance, '(fstat)-cache-reaper is a common pattern, it needs to be automated, you should not write it manually.

4. delete the least recently used (LRU) element. Was this what you meant by the FIFO mechanism above? Yes.

5. A function to clear the cache. Yes, good idea. See above, as I said, ...-cache-reaper is a common pattern.

-----

2 points by akkartik 5524 days ago | link

Ah, I totally missed your reaper thread when I wrote http://arclanguage.org/item?id=10699

I suppose you mean "set cach.key to (the real) nil"

Yep :}

-----