You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Child compaction is another killer feature of go-leveldb project.
Although the concurrent compaction is already enabled by default
which speeds up the compaction efficiency a lot, but the write stall
is still not fixed.
What's write stall
Write stall refers to the phenomenon that the system write is disabled
temporarily because of the SLOW COMPACTION. The root issue is
the level0 compaction is not fast enough.
But currently the concurrent level0 compaction is not feasible because
the files in level0 are overlapped with each other. In this case we need
some additional mechanism to improve the efficiency of level0 compaction.
I would say the efficiency of level0 compaction is the KEY of the leveldb
write performance.
Pebble has the idea called SubLevel0Compaction. The idea is quite
straightforward: create indexes for level0 files whenver the level0 content
is changed, try to schedule concurrent level0 compactions for these
"non-overlapped" files. However in practice, it's not so easy to find such
files.
So we had this idea called ChildCompaction. Let's write it down for further
discussion.
The text was updated successfully, but these errors were encountered:
Child compaction is another killer feature of go-leveldb project.
Although the concurrent compaction is already enabled by default
which speeds up the compaction efficiency a lot, but the write stall
is still not fixed.
What's write stall
Write stall refers to the phenomenon that the system write is disabled
temporarily because of the SLOW COMPACTION. The root issue is
the level0 compaction is not fast enough.
But currently the concurrent level0 compaction is not feasible because
the files in level0 are overlapped with each other. In this case we need
some additional mechanism to improve the efficiency of level0 compaction.
I would say the efficiency of level0 compaction is the KEY of the leveldb
write performance.
Pebble has the idea called
SubLevel0Compaction
. The idea is quitestraightforward: create indexes for level0 files whenver the level0 content
is changed, try to schedule concurrent level0 compactions for these
"non-overlapped" files. However in practice, it's not so easy to find such
files.
So we had this idea called
ChildCompaction
. Let's write it down for furtherdiscussion.
The text was updated successfully, but these errors were encountered: