[
https://issues.apache.org/jira/browse/CASSANDRA-1608?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Benjamin Coverston updated CASSANDRA-1608:
------------------------------------------
Attachment: 1608-v8.txt
First the good:
1. Modified the code s.t. tombstone purge during minor compactions use the
interval tree to prune the list of SSTables speeding up compactions by at least
an order of magnitude where the number of SSTables in a column family exceeds
~500.
2. Tested reads and writes. Write speeds (unsurprisingly) are not affected by
this compaction strategy. Reads seem to keep up as well. The interval tree does
a good job here making sure that bloom filters are only queried only for those
SSTables that fall into the queried range.
3. Three successive runs of stress inserting 10M keys resulted in ~3GB of data
stored in leveldb. By comparison, the same run using the tiered (default)
strategy resulted in ~8GB of data.
The Meh:
Compactions do back up when setting the flush size to 64MB and the leveled
SSTable size to anywhere between 5-10MB. On the upside, if your load has peaks
and quieter times this compaction strategy will trigger a periodic check to
"catch up" if all event-scheduled compactions complete.
Interestingly this extra IO has an upside. For datasets that frequently
overwrite old data that has already been flushed to disk there is the potential
for substantial de-duplication of data. Further, during reads the number of
rows that would need to be merged for a single row is bound by the number of
levels + the number of un-leveled sstables.
> Redesigned Compaction
> ---------------------
>
> Key: CASSANDRA-1608
> URL: https://issues.apache.org/jira/browse/CASSANDRA-1608
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: Chris Goffinet
> Assignee: Benjamin Coverston
> Attachments: 0001-leveldb-style-compaction.patch, 1608-v2.txt,
> 1608-v3.txt, 1608-v4.txt, 1608-v5.txt, 1608-v7.txt, 1608-v8.txt
>
>
> After seeing the I/O issues in CASSANDRA-1470, I've been doing some more
> thinking on this subject that I wanted to lay out.
> I propose we redo the concept of how compaction works in Cassandra. At the
> moment, compaction is kicked off based on a write access pattern, not read
> access pattern. In most cases, you want the opposite. You want to be able to
> track how well each SSTable is performing in the system. If we were to keep
> statistics in-memory of each SSTable, prioritize them based on most accessed,
> and bloom filter hit/miss ratios, we could intelligently group sstables that
> are being read most often and schedule them for compaction. We could also
> schedule lower priority maintenance on SSTable's not often accessed.
> I also propose we limit the size of each SSTable to a fix sized, that gives
> us the ability to better utilize our bloom filters in a predictable manner.
> At the moment after a certain size, the bloom filters become less reliable.
> This would also allow us to group data most accessed. Currently the size of
> an SSTable can grow to a point where large portions of the data might not
> actually be accessed as often.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
