alt Hacker NewsThe "Early Return" sections describe it more, I don't think it's as bad as it sounds in that first bullet. They're returning after 15 out of 20 complete,and it sounds like even if only those 15 end up succeeding it'll still generally be fine. (Exactly how fine / is that violating some common expectations and will cause problems: I dunno. Not familiar enough with IPFS's internals)
That said:
>In practice, at least one of the 20 follow-up requests fails in the vast majority of operations, and a single unresponsive peer can stall the entire phase waiting for a timeout.
It continually surprises me how often systems lack a Fast Fallback-like strategy¹. Or at least sound like it. Just an absolute flood of apps and websites and systems that try to do something once and then never tried an alternate route until that finishes, something like a minute or two later... for a process that usually takes less than a second. It's maddening. By the time you're considering one to be "stalled" and delaying everything unnecessarily, you probably should've already started trying two or three alternate routes!
> (Exactly how fine / is that violating some common expectations and will cause problems: I dunno. Not familiar enough with IPFS's internals)
I felt the article addressed that a bit further down. 20 copies is a somewhat arbitrary knob in the Kademlia DHT design IPFS is based on and this lab's research suggested that 15 was probably closer to good enough for GET requests to succeed at about the same time cost. Rather than dropping the knob for the entire DHT, because redundancy is always useful in the long run they went with the Early Return and a secondary process called the Reprovide Sweep that still tries to push the network towards the 20 live copies minimum it desires.
I'm assuming the Reprovide Sweep was work previously done/documented because it seems like something that might have been more interesting to discuss at longer length in relevant parts of the article.