[
https://issues.apache.org/jira/browse/CASSANDRA-4539?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13447834#comment-13447834
]
Sylvain Lebresne commented on CASSANDRA-4539:
---------------------------------------------
bq. I read that section as saying that <value> was a [value]
Hum, I see how one could be confused :). Though the idea is really that the two
are not related at all. I've renamed [value] to [bytes] in commit 06269b9 to
hopefully make if more clear that those are not related.
Now the main _raison d'être_ of [option] is to have a compact way to describe
the type in the query response metadata. For that, and given that it's
something sent very frequently, I really think we should keep it the way it is.
I'm all for making client library implementers job easy as much as possible but
adding even one byte per type is just not worth it in that case. And I don't
think it's much of a problem in the case we add new data type as long as we fix
the option ids for a given protocol version. And if we add a new type, the
server can very well use the 'custom' option id until we decide to bump the
protocol version to add a specific option id for that new type.
Now for STARTUP, I'll admit that reusing [option] might have been a bit
overkill. It would probably be simpler/more flexible to change STARTUP to take
some map<string, string> (and to have OPTIONS messages to return a map<string,
[string list]>. I'll make that change.
> potential backwards incompatibility in native protocol
> ------------------------------------------------------
>
> Key: CASSANDRA-4539
> URL: https://issues.apache.org/jira/browse/CASSANDRA-4539
> Project: Cassandra
> Issue Type: Improvement
> Components: API
> Affects Versions: 1.2.0 beta 1
> Reporter: paul cannon
> Assignee: Sylvain Lebresne
> Priority: Minor
> Labels: cql, native_protocol
> Fix For: 1.2.0 beta 1
>
>
> In the text of the native_protocol.spec document, it explains the format for
> a notation called {{[option]}}, which should represent "{{a pair of
> <id><value>}}".
> In doing a first-draft implementation of the protocol for the python driver,
> though, I found that I had a misunderstanding. I read that section as saying
> that {{<value>}} was a {{[value]}}, and that it could have a length of 0
> (i.e., the {{[int]}} on the front of the {{[value]}} could be 0). However, it
> turns out that {{<value>}} might not be there at all, or might be *two*
> {{[value]}}'s, depending on the option id and message context.
> I'm not a fan of this, since
> * A protocol parsing library can't simply implement a single function to
> read in {{[option]}}'s, since the length of the value part is dependent on
> the message context
> * If we add a new native data type (a new option id which could be used
> inside a {{<col_spec_i>}} in a RESULT message), then older clients will not
> know how to read past the value part. Of course they won't know how to
> interpret the data or deserialize later rows of that unknown type - that's
> not the problem - the problem is that the older client should be able just to
> mark that column as unparseable and still handle the rest of the columns.
> Can we make {{<value>}} be a {{[value]}}, the contents of which can be
> re-interpreted as a {{[string]}}, an {{[option]}}, two {{[option]}}'s, or
> whatever?
--
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
