sstable::create_links does not clean up after itself on error

[bhalevy]

This issue is split off of #5714, focusing on create_links error handling.

As mentioned in a4d8e99 FIXME comment if link_file fails on any of the components (e.g. when running out of space) the function returns the error but it does not clean up after itself, leaving a mess behind that requires human intervention, since retrying the link operation will fail.

[bhalevy]

Patchset [scylladb-dev] [PATCH v1 00/10] sstables: make move_to_new_dir idempotent was sent to the ML, fixing this issue.

[avikivity]

Not a regression, rare error, not fatal, so not backporting.

[bhalevy]

@avikivity the issue prevents nodes that experienced a crash from starting when move from staging hits the remainders of partially moved sstables. This is seen in the field by a user that runs 4.2.1.

[slivne]

@bhalevy was this backported to 2021.1 ?

[bhalevy]

@bhalevy was this backported to 2021.1 ?

No, 5b312a1 was merged in 4.4 and it was decided not to backport it (see #7429 (comment))