[vitpro2213]

If you discard the old array (and allocate a bigger one before), you must also copy all enqueued material.

Also this one enqueue will be mega expensive - a clear "fat tail" in the latency histogram.

In MultiArrayQueue you keep all already enqueued material in-place, "just" allocate the bigger array, register the new diversion, enqueue the one new element - and done.

Thanks

[Szpadel]

why not free previous smaller chunk after reader finished reading from it?

for me it would be better to allocate new buffer but allow reading from old one until it contains data and after that dealocate it and keep only new one in use

[zamalek]

You might run into ABA, though that isn't an issue for managed languages.

[jauntywundrkind]

Yeah, leaving small old buffers behind seems like a major no-no to me. It could be useful if you think you'll shrink back down, but it feels like cache locality suffering and iteration/tracking penalties strongly incentive getting rid of the old buffer asap.

One other thing I want to shout out, I saw what I thought was a really neat multicast ring buffer the other day where the author has an atomic for each element, rather than the typical reader/writer atomics. The promise was having much less contention on any given atomic, in most cases. https://github.com/rezabrizi/SPMC-Queue https://news.ycombinator.com/item?id=40410172

[vitpro2213]

Removing anything in non-blocking structures is problematic, see e.g. the referenced lecture of Professor Scott.

You never know how many concurrent threads still "are" on the place you wish to remove.

You would have to deal with stuff like hazard pointers, limbo lists and the like.

Better to keep the small arrays there.

[Szpadel]

Isn't this only issue when you allow referencing data in queue?

If queue only allows to copy out data you can increase reader pointer after data were copied to different buffer, therefore nothing can be at the place we are removing

[vitpro2213]

With ConcurrentMultiArrayQueue, there can be N threads INSIDE of the program code of the Queue, running or preempted (for a not predictable time) and you cannot control it.

[jauntywundrkind]

I'm sorry but that doesn't seem at all like any kind of fundamental constraint. At most basic, simply not have your readers advance until after they're done reading? This seems trivial as fuck. I've seen a lot of protestations that have seemed like, ok, maybe perhaps perhaps perhaps I'm missing some factor? But no, it really seems like the problem of understanding when data is done being read isn't anywhere near as hard as the rebuttals here make it seem, and it seems like everything works much better when we can accept this constraint.

Perhaps we want to have any of the given multicast readers able to read more than one element at a time, and that does complicate things somewhat. But hardly impossible to handle.

Again: deeply disagreeing with the premise here that this can't be done. And it isn't even really a significant penalty, if your consumers do have to be async consumers that need to hold open their reading for a while. Unclear what the protests are.