>> Often failures in cl-test-grid are "just" the result of using too >> little memory, > > not little memory, but SBCL's conservative (i.e. imprecise) garbage > collection - it doesn't always recognize that an object is a garbage. > > I saw cases where CCL compiles a project under a 100MB heap, > while SBCL reproducibly runs out of 1GB heap. > Maybe it's an artefact of SBCL using too much memory *while compiling* and would go away if you used e.g. POIU to compile inside forks.
>> or batching system loads, > > There is no batch loading, each test is run in a fresh CL process. > Interesting. The memory load might still be sensitive to compilation order, if you reuse fasls shared with previous builds versus rebuild new ones. >> or some other reason, and >> have to be retried. > > I would put it differently: some failures may be ignored because > on the previous version they were successful by a chance, > and could faile the same way too. > OK. PS: The newer test results look good. I'd like to see the results with the syntax-control branch, though. There also are a few minor bug fixes in the pipe that mean we'll have to re-test before release, anyway. —♯ƒ • François-René ÐVB Rideau •Reflection&Cybernethics• http://fare.tunes.org Foolishness is rarely a matter of lack of intelligence or even lack of information. — John McCarthy