Pinging @elastic/es-distributed (Team:Distributed)

Unfortunately this is somewhat performance sensitive due to the brute-force nature of the linearizability checker. The tests risk timing out with this change. I wonder if there is an alternative map implementation available (i.e., that we already use perhaps) that avoids wrapping longs? Or else we will need to build or use one here.

Would it be a viable option to leave the dependency on hppc as a test only dependency?

The tests risk timing out with this change.

While I understand the concern, I question whether this change is that impactful. Java has become quite good at inlining boxed objects. Also, HashMap's table is a singular array, while hppc's map is two arrays (keys and values), so getting at a value still requires a second dereference once the index is retrieved.

I wonder if there is an alternative map implementation available (i.e., that we already use perhaps) that avoids wrapping longs? Or else we will need to build or use one here

Could we try out this change, and if it results in timeouts, then we can investigate how to fix?

Would it be a viable option to leave the dependency on hppc as a test only dependency?

This would not be ideal. It would still leave open for hppc use in many places, especially since this checker is in the test framework.

@elasticmachine run elasticsearch-ci/part-2

@henningandersen I ran 20 iterations of ConcurrentSeqNoVersioningIT.testSeqNoCASLinearizability (with a fixed seed, though I know that doesn't matter much give the test implementation). Baseline on master took 5m17s. This PR took 4m43s. I don't claim this PR is somehow better, but it is certainly not leading to OOMs or slower checks.

Thanks Ryan. Perhaps also run it 20 times with no fixed seed just to check that it does not OOM with a random seed. The seed does determine the disruption type, size of cluster, number of keys, number of threads and more so might reveal a different result.

Sorry, I was not clear. The fixed seed was the root seed. Each individual test iteration had a unique seed, based on the root seed. I did the fixed seed so that each iteration used the same seed, but they were all still unique.

@henningandersen I ran 100 more iterations and did not see any timeouts or OOMs with this PR.