Commit 8c4b5e4283 introduced an optimization which only
calculates max purgeable timestamp when a tombstone satisfy the
grace period.
Commit 'repair: Get rid of the gc_grace_seconds' inverted the order,
probably under the assumption that getting grace period can be
more expensive than calculating max purgeable, as repair-mode GC
will look up into history data in order to calculate gc_before.
This caused a significant regression on tombstone heavy compactions,
where most of tombstones are still newer than grace period.
A compaction which used to take 5s, now takes 35s. 7x slower.
The reason is simple, now calculation of max purgeable happens
for every single tombstone (once for each key), even the ones that
cannot be GC'ed yet. And each calculation has to iterate through
(i.e. check the bloom filter of) every single sstable that doesn't
participate in compaction.
Flame graph makes it very clear that bloom filter is a heavy path
without the optimization:
45.64% 45.64% sstable_compact sstable_compaction_test_g
[.] utils::filter::bloom_filter::is_present
With its resurrection, the problem is gone.
This scenario can easily happen, e.g. after a deletion burst, and
tombstones becoming only GC'able after they reach upper tiers in
the LSM tree.
Before this patch, a compaction can be estimated to have this # of
filter checks:
(# of keys containing *any* tombstone) * (# of uncompacting sstable
runs[1])
[1] It's # of *runs*, as each key tend to overlap with only one
fragment of each run.
After this patch, the estimation becomes:
(# of keys containing a GC'able tombstone) * (# of uncompacting
runs).
With repair mode for tombstone GC, the assumption, that retrieval
of gc_before is more expensive than calculating max purgeable,
is kept. We can revisit it later. But the default mode, which
is the "timeout" (i.e. gc_grace_seconds) one, we still benefit
from the optimization of deferring the calculation until
needed.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Closes#13908
38b226f997