[
https://issues.apache.org/jira/browse/CASSANDRA-1379?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Javier Canillas updated CASSANDRA-1379:
---------------------------------------
Attachment: CASSANDRA-1379.patch
This issue has been around my head for some time and since we are using
Cassandra on live environment (not exactly 0.7 version) I have taken the time
to make a patch for you to check. Here you will find some roadmap of my change:
Since Table.getRow was checking if the asked Key was in the RowCache before
going directly into the Memtables and SSTables, I have made another method
called getCachedRow that only checks for the key on the RowCache. If its not
there, then a null object (actually a Row.Cf = null) is returned. If you see
the code, you will notice that I didn't remove the Cache check on the
Table.getRow since I prefer a double check on cache, than a second miss if, for
some reason, someone put that row on cache after first check.
Getting that in mind, I have added another public method to ReadCommand and
implemented on SliceByNamesReadCommand and SliceFromReadCommand. This new
method is actually only calling the Table.getCachedRow instead of Table.getRow.
Doing so, now I went straight to see ReadVerbHandler, and I cloned it for a new
ReadCacheVerbHandler (then I turned ReadVerbHandler into an abstract class and
created a ReadPhyisicalVerbHandler to avoid duplicate code). This new class
(ReadCacheVerbHandler) will call the ReadCommand.getCachedRow, if this method
returns a Row with cf equals to null, then the command must be forwarded to
another Stage in order to fulfill the requirements by making a physical read.
This new stage was named PHYSICAL_READ and a new verb (PHYSICAL_READ) was
created. The new stage was actually cloned from READ Stage (please check if
this stage needs further tunning).
In order to be able to manipulate the original message to change the Verb,
ReadCommand.makeReadMessage was changed and extended to support overrides on
messageId and Verb itself.
So, after getting an empty result on a cache search, the read command is
forwarded into the PHYSICAL_READ stage (freeing a READ thread), and it is the
last one that actually respond the row; otherwise, the cached row is returned.
Sorry for the large response and my bad english, I tried to explain the main
idea, but you probably will understand all this mechanism by inspecting the
code.
Long life to Cassandra!
> Uncached row reads may block cached reads
> -----------------------------------------
>
> Key: CASSANDRA-1379
> URL: https://issues.apache.org/jira/browse/CASSANDRA-1379
> Project: Cassandra
> Issue Type: New Feature
> Components: Core
> Reporter: David King
> Priority: Minor
> Fix For: 0.7.1
>
> Attachments: CASSANDRA-1379.patch
>
>
> The cap on the number of concurrent reads appears to cap the *total* number
> of concurrent reads instead of just capping the reads that are bound for
> disk. That is, given N concurrent readers if all of them are busy waiting on
> disk, even reads that can be served from the row cache will block waiting for
> them.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
