compaction: Fail split of new sstable if manager is disabled

If manager has been disabled due to out of space prevention, it's
important to throw an exception rather than silently not
splitting the new sstable.

Not splitting a sstable when needed can cause correctness issue
when finalizing split later.
It's better to fail the writer (e.g. repair one) which will be
retried than making caller think everything succeeded.

The new replica::table::add_new_sstable_and_update_cache() will
now unlink the new sstable on failure, so the table dir will
not be left with sstables not loaded.

Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
This commit is contained in:
Raphael S. Carvalho
2025-12-08 16:27:28 -03:00
parent ee3a743dc4
commit 992bfb9f63
2 changed files with 7 additions and 2 deletions

View File

@@ -2290,8 +2290,12 @@ compaction_manager::maybe_split_new_sstable(sstables::shared_sstable sst, compac
if (!split_compaction_task_executor::sstable_needs_split(sst, opt)) {
co_return std::vector<sstables::shared_sstable>{sst};
}
if (!can_proceed(&t)) {
co_return std::vector<sstables::shared_sstable>{sst};
// Throw an error if split cannot be performed due to e.g. out of space prevention.
// We don't want to prevent split because compaction is temporarily disabled on a view only for synchronization,
// which is uneeded against new sstables that aren't part of any set yet, so never use can_proceed(&t) here.
if (is_disabled()) {
co_return coroutine::exception(std::make_exception_ptr(std::runtime_error(format("Cannot split {} because manager has compaction disabled, " \
"reason might be out of space prevention", sst->get_filename()))));
}
std::vector<sstables::shared_sstable> ret;

View File

@@ -376,6 +376,7 @@ public:
// Splits a single SSTable by segregating all its data according to the classifier.
// If SSTable doesn't need split, the same input SSTable is returned as output.
// If SSTable needs split, then output SSTables are returned and the input SSTable is deleted.
// Exception is thrown if the input sstable cannot be split due to e.g. out of space prevention.
future<std::vector<sstables::shared_sstable>> maybe_split_new_sstable(sstables::shared_sstable sst, compaction_group_view& t, compaction_type_options::split opt);
// Run a custom job for a given table, defined by a function