[
https://issues.apache.org/jira/browse/CASSANDRA-4449?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sylvain Lebresne updated CASSANDRA-4449:
----------------------------------------
Attachment: 4449-v2.txt
bq. Why not use CLHM and get actual LRU behavior instead of strict
remove-oldest behavior?
Not only is there no good reason for doing so (I dumbly copy/pasted the
existing code) but it's actually wrong since the prepared map is now accessed
by multiple threads. Changed in v2 to CLHM.
bq. What does continuing to use clientState for Thrift clients buy us? If we
give unique IDs (and I agree w/ the change to md5) I don't see how our internal
implementation details matter.
That was mostly laziness on my part to change the minimum relating to thrift.
But I can agree that I doubt anyone rely on the fact that statement are not
global, so changed in v2 to make it always global for CQL3.
> Make prepared statement global rather than connection based
> -----------------------------------------------------------
>
> Key: CASSANDRA-4449
> URL: https://issues.apache.org/jira/browse/CASSANDRA-4449
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Sylvain Lebresne
> Assignee: Sylvain Lebresne
> Labels: binary_protocol
> Fix For: 1.2.0 beta 2
>
> Attachments: 4449.txt, 4449-v2.txt
>
>
> Currently, prepared statements are connection based. A client can only use a
> prepared statement on the connection it prepared it on, and if you prepare
> the same prepared statement on multiple connections, we'll keep multiple
> times the same prepared statement. This is potentially inefficient but can
> also be fairly painful for client libraries with pool of connections (a.k.a
> all reasonable client library ever) as this means you need to make sure you
> prepare statement on every connection of the pool, including the connection
> that don't exist yet but might be created later.
> This ticket suggests making prepared statement global (at least for CQL3),
> i.e. move them out of ClientState. This will likely reduce the number of
> stored statement on a given node quite a bit, since it's very likely that all
> clients to a given node will prepare the same statements (and potentially on
> all of their connection with the node). And given that prepared statement
> identifiers are the hashCode() of the string, this should be fairly trivial.
> I will note that while I think using a hash of the string as identifier is a
> very good idea, I don't know if the default java hashCode() is good enough.
> If that's a concern, maybe we should use a safer (bug longer) hash like md5
> or sha1. But we'd better do that now.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
