tis 2008-04-08 klockan 23:50 +1200 skrev Amos Jeffries: > If its not easy and its not simple it means more bugs. I don't think 'c' > is a good choice here without a full ground-up recode. Nay on that from > me on general principles.
True. I view 'c' as more of a last resort if the other turn out to be unmaintainable. Initially I didn't even consider it an option.. > 'a' gains on grounds of easy, but can it be done and gain performance? I am not so worried about performance. Yes, the naive implementation using cbdata just as it is today will probably not be very good (but not bad), but that's possible to address. Maybe by addressing cbdata, maybe by replacing it with a similar mechanism but more optimized for shortlived interest registrations and cancellations of the same. The underlying problem is rapid frequent construction/destruction of objects. > 'b' sounds like the assert() jungle squid currently has Yes. 'b' is nearly impossible to audit for correctness. So is 'c'. But 'c' reduces the risk of this kind of race windows happening to nearly 0. Regards Henrik