[Object spilling] Fix race condition that deletes files at the wrong timing.

[rkooo567]

Why are these changes needed?

TODO: cpp unit test

The issue was that we add ref count “after” spilled url is updated to the owner (after RPC). So this could’ve happened.

file A: object A, B, C
object A, B, C are concurrently waiting for url to be updated to the owner
object A url is updated
object A updates ref count
object A is evicted
we try to delete the file. Check ref count to see if it is 0
ref count == 1 -> 0, so we delete the file A  (which shouldn’t happen because object B and C are not evicted yet)

I think we could easily fix the issue by updating ref counting before updating urls

Related issue number

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[stephanie-wang]

Nice catch! Yes, a regression test would be good.