Open sstrickl opened 5 years ago
Tried reproducing locally at the front and end of the affected range of commits, didn't trigger at any point over 100 runs of each. However, I expect my dev machine has more memory than the builder bots. Given that, it's likely something that was exposed due to memory pressure. If lazy constant evaluation shifted the memory use during reload so that memory was less constrained at that point, that could explain why it stopped happening after that commit.
We're seeing flaky crashes on
vm/dart/trigger_gc_in_native_test
on one of the reload/rollback bots when the test is run with both the--no_concurrent_mark
and--use_compactor
flags. It started showing up in the ignored flaky test failure logs starting at commit da8cb470cc9 and showed up there consistently until commit b32d196f. I'm not sure whether the latter commit fixed the issue that was causing the crash, though, or if it's a still-existing issue that's just masked for some reason now.Failures from example logs:
There are also crashes on Android, see #37185. While that issue doesn't have any comments explaining why it was closed, looking at the logs I expect it's an OOM killer and unrelated.