Summary: Currently, when rocksdb tries to run manual compaction to refit data into a level, there's a ReFitLevel() process that requires no bg work is currently running. When RocksDB plans to ReFitLevel(), it will do the following: 1. pause scheduling new bg work. 2. wait until all bg work finished 3. do the ReFitLevel() 4. unpause scheduling new bg work. However, as it pause scheduling new bg work at step one and waiting for all bg work finished in step 2, RocksDB will stop flushing until all bg work is done (which could take a long time.) This patch fix this issue by changing the way ReFitLevel() pause the background work: 1. pause scheduling compaction. 2. wait until all bg work finished. 3. pause scheduling flush 4. do ReFitLevel() 5. unpause both flush and compaction. The major difference is that. We only pause scheduling compaction in step 1 and wait for all bg work finished in step 2. This prevent flush being blocked for a long time. Although there's a very rare case that ReFitLevel() might be in starvation in step 2, but it's less likely the case as flush typically finish very fast. Test Plan: existing test. Reviewers: anthony, IslamAbdelRahman, kradhakrishnan, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D55029main
parent
68189f7e1b
commit
a7d4eb2f34
Loading…
Reference in new issue