Open DerGuteMoritz opened 2 years ago
Just found https://github.com/clj-commons/byte-streams/issues/10 which also touches on the issue of the hidden initialization cost which resulted in the introduction of a precache-conversions
API. This didn't solve the early capturing issue but at least is some prior art in the spirit of my suggestion. However, it later got removed again without further explanation. Hm!
The main issue here could also be solved by invalidating the memo whenver new conversions are declared. However, this wouldn't also address the performance gotcha, so I decided to break it out into its own issue.
It's important to articulate the problem; is this a real concern for anyone?
The only non-toy/demo/example I found of using def-conversion
anywhere on Github was for clj-fdb, and that was for a new conversion, and in the tests.
This issue is currently theoretical afaict, and I don't think anyone should spend time on it. 😄
Declaring new conversions via
def-conversion
which would make conversion between two types less costly are ineffective if the conversion in question has occurred at least once before. This is caused by the global converter memoization which captures the state of the conversion graph at the point in time of the very first invocation for a given pair of source and dest types.Reproducer:
Output:
In contrast, moving both
run
invocations after the lastbs/def-conversion
outputs:This is a bit of a gotcha which might at least be worth documenting. Alternatively,
def-conversion
could reset the memo which should solve this issue.