Benedict created CASSANDRA-6812:
-----------------------------------
Summary: Iterative Memtable->SSTable Replacement
Key: CASSANDRA-6812
URL: https://issues.apache.org/jira/browse/CASSANDRA-6812
Project: Cassandra
Issue Type: Improvement
Components: Core
Reporter: Benedict
Fix For: 3.0
In an ideal world we wouldn't flush any memtable until we were almost
completely out of room. The problem with this approach (and in fact whenever we
currently *do* run out of room) is that flushing an entire memtable is a slow
process, and so write latencies spike dramatically during this interval.
The solution to this is, in principle, quite straight forward: As we write
chunks of the new sstable and its index, open them up immediately for reading,
and free the memory associated with the portion of the file that has been
written so that it can be reused immediately for writing. This way whilst
latency will increase for the duration of the flush, the max latency
experienced during this time should be no greater than the time taken to flush
a few chunks, which should still be on the order of milliseconds, not seconds.
This depends on CASSANDRA-6689 and CASSANDRA-6694, so that we can reclaim
arbitrary portions of the allocated memory prior to a complete flush.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
