cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Björn Hegerfors (JIRA) <>
Subject [jira] [Created] (CASSANDRA-8361) Make DTCS split SSTables to perfectly fit time windows
Date Fri, 21 Nov 2014 17:42:34 GMT
Björn Hegerfors created CASSANDRA-8361:

             Summary: Make DTCS split SSTables to perfectly fit time windows
                 Key: CASSANDRA-8361
             Project: Cassandra
          Issue Type: Improvement
            Reporter: Björn Hegerfors
            Priority: Minor

The time windows that DTCS uses are what the strategy tries to align SSTables to, in order
to get the right structure, for best performance. I added the ticket CASSANDRA-8360, taking
SSTables one step closer to aligning with these windows in a 1:1 manner.

The idea in this ticket is to perfectly align SSTables with the DTCS time windows, by splitting
SSTables that cross window borders. This can lead to certain benefits, perhaps mostly in consistency
and predictability terms, where it will be very well defined where every value is stored that
is old enough to have stabilized.

Read queries can be aligned with windows in order to guarantee a single disk seek (although
then the client needs to know the right window placements). Basically, SSTables can be made
to align perfectly on day borders, for example. Right now, there would be an SSTable that
almost represents a day, but not perfectly. So some data is still in another SSTable. 

It could also be a useful property for tombstone expiration and repairs.

Practically all splits would happen only in the latest time windows with the newest and smallest
SSTables. After those are split, DTCS would never compact SSTables across window borders.
I have a hard time seeing when this could cause an expensive operation except for when switching
from another compaction strategy (or even from current DTCS), and after a major compaction.
In fact major compaction for DTCS should put data perfectly in windows rather than everything
in one SSTable.

This message was sent by Atlassian JIRA

View raw message