rocksdb enter fatal error model after recovery from no space error

[lee-qiu]

See #11643.

I found that when the disk free size is large enough, the db calls the db_->ResumeImpl() function, but if the memory buffer is empty, the SwitchMemtable() function is not triggered.

So rocksdb will not create a new wal file. When the next write request comes, the write to wal file will get a IOStatus::IOError("Writer has previous error.") error, which makes the engine get a fatal error and you must reopen the database to recover.

I wanna know can we trigger SwitchMemtable() function when the flush reason is "kErrorRecovery" and memory buffer is empty?

I found that can help me to escape from the above fatal error. I'd like to make a push request to resolve this issue If you guys agree it.

Waiting for your reply.