> Shenandoah implicitly relies on collecting faster than application allocates
That's what the documentation tells you that you need in order to deploy Shenandoah. If you can't meet that then you don't get the claimed performance characteristics.
If you can meet that requirement then GC pause times are dependent on root set sizes, not heap sizes.
> Shenandoah implicitly relies on collecting faster than application allocates
What precisely does this even mean? How is the programmer or user supposed to provision for this? By blocking on every allocation until enough time has passed since the previous one... which merely shifts the blame for the pause from the GC onto the application? And so if there are A allocations/second and D deallocations/second and we have A < D then suddenly there's a guarantee of constant-time pauses? I don't buy that either. I'm pretty confident I could construct a counterexample if I spend some time to create a nontrivial object graph. Do you believe it?
There is no shift here, this has always been how things work. GCs are not a magic wand that allow you to do anything with reckless disregard towards efficiency with zero drawbacks (neither do malloc/free). Improved GC algorithms strive to cover more and more real-world workloads. Sometimes they fall short. Sometimes the developer did inefficient things that could benefit from optimizations to decrease allocation rates. A classic example is the O(n²) string appending (+=) behavior for which one should use StringBuilders instead. It just puts unnecessary stress on the GC.
G1GC had issues with LRU-like in-memory cache workloads. Some of those were improved with ongoing G1 development, but ultimately the new concurrent collectors (ZGC, shenandoah) solve those problems better since their pause times do not explode when many inter-region pointers need to be updated.
Those performance goals are only valid when given enough spare resources (heap > live set size, cpu cores). If there are insufficient resources then the VM may eventually enter those "failure modes" to keep the application running, albeit at reduced performance, instead of crashing. I think there are flags to make it just exit if that's preferable over latency spikes.
Ultimately you cannot ask it to do the impossible, this much should be obvious.
How do you know if allocation is slower than collection? The logs tell you. This is the same as any performance requirement - you need to test it empirically.
If you think there's an issue with the claims of Shenandoah then you should publish your work rather than just abuse.
> Pause times with Shenandoah are independent of heap size, meaning you will have the same consistent pause times whether your heap is 200 MB or 200 GB.
You can't just write an unconditional "summary" like that with a straight face. It's practically dishonest to present it as if the giant asterisk isn't there. And this has nothing to do with the quality of the GC... I'm sure they've done great work, and if they're proud of it they probably very well should be, but that doesn't give us (or them) an excuse to advertise it as doing something it admits it doesn't do.
> Usual latency induced: >100 ms, but can be more, especially on a very occupied heap
[1] https://wiki.openjdk.java.net/display/shenandoah/Main#Main-F...