Botond Dénes

> I would check if there are evictable objects in the cache_tracker's LRU cache tracker's LRU has 13593 elements

Seen again in https://jenkins.scylladb.com/job/scylla-master/job/scylla-ci/16632/testReport/junit/cluster/test_tablets/Tests___Unit_Tests___test_two_tablets_concurrent_repair_and_migration_repair_writer_level_dev_3_2/.

@Deexie please look into fixing this test.

@avikivity what is the status of this PR?

Based on feedback from the plan review, we won't implement any of the more sophisticated algorithms suggested here, instead we will simply reject everything that generates writes when disk usage...

Yes, all solutions proposed here assume tablets and additional disk added by scale-out. All proposed solutions revolve around speeding up tablet migration (out of the node) and managing remaining disk...

> > Follow-up idea: now that we don't need to track highest generation anymore, can we make generation generator completely stateless and remove the generator from make sstable factory method...

> so if RBNO is used for removenode, we shouldn't do a full cluster repair ? If you use RBNO, you don't need extra repair, repair happens during the node...

Seen again in https://jenkins.scylladb.com/job/scylla-master/job/next/8753/testReport/junit/(root)/non-boost%20tests/Build___aarch64___Unit_Tests_aarch64___cqlpy_cassandra_tests_validation_entities_frozen_collections_test_debug_1/

Seen again in https://jenkins.scylladb.com//job/scylla-master/job/scylla-ci/14971/testReport/junit/%28root%29/non-boost%20tests/Tests___Unit_Tests___cqlpy_cassandra_tests_validation_entities_frozen_collections_test_debug_2.