[
https://issues.apache.org/jira/browse/CASSANDRA-6906?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13958114#comment-13958114
]
Tyler Hobbs commented on CASSANDRA-6906:
----------------------------------------
bq. milliseconds? Something is wrong if so.
Yes, 0.1 seconds. Here's an example trace of {{select * from system.local}} on
a 15 node cluster with vnodes enabled:
{noformat}
activity
| timestamp | source | source_elapsed
-------------------------------------------------------------------------------------------------+--------------+-----------+----------------
execute_cql3_query | 15:05:55,295 | 127.0.0.1 | 0
Parsing select * from
system.local LIMIT 10000; | 15:05:55,297 | 127.0.0.1 | 1884
Preparing statement | 15:05:55,299 | 127.0.0.1 | 3772
Determining
replicas to query | 15:05:55,300 | 127.0.0.1 | 4435
Executing seq scan across 3 sstables for [min(-9223372036854775808),
min(-9223372036854775808)] | 15:05:55,437 | 127.0.0.1 | 141652
Seeking to partition
beginning in data file | 15:05:55,440 | 127.0.0.1 | 145212
Seeking to partition
beginning in data file | 15:05:55,440 | 127.0.0.1 | 145274
Seeking to partition
beginning in data file | 15:05:55,440 | 127.0.0.1 | 145309
Read 1 live and 0
tombstoned cells | 15:05:55,451 | 127.0.0.1 | 156072
Scanned 1
rows and matched 1 | 15:05:55,451 | 127.0.0.1 | 156269
Request complete | 15:05:55,455 | 127.0.0.1 | 160916
{noformat}
I agree that this could probably be greatly optimized in other ways as well.
> Skip Replica Calculation for Range Slice on LocalStrategy Keyspace
> ------------------------------------------------------------------
>
> Key: CASSANDRA-6906
> URL: https://issues.apache.org/jira/browse/CASSANDRA-6906
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: Tyler Hobbs
> Assignee: Tyler Hobbs
> Priority: Minor
> Fix For: 2.0.7, 2.1 beta2
>
> Attachments: 6906.txt
>
>
> For vnode-enabled clusters, the "Determining replicas to query" portion of
> range slice commands can be expensive. When querying LocalStrategy
> keyspaces, we can safely skip this step. On a 15 node cluster with vnodes,
> skipping this saves about 100ms. This makes a big difference for the
> drivers, which frequently execute queries like "select * from system.peers"
> and "select * from system.local".
--
This message was sent by Atlassian JIRA
(v6.2#6252)
