cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Hiller, Dean" <>
Subject Re: Prepared Statement - cache duration (CQL3 - Cassandra 1.2.4)
Date Tue, 23 Apr 2013 17:35:45 GMT

From: Sylvain Lebresne <<>>
Reply-To: "<>" <<>>
Date: Tuesday, April 23, 2013 11:31 AM
To: "<>" <<>>
Subject: Re: Prepared Statement - cache duration (CQL3 - Cassandra 1.2.4)

On Tue, Apr 23, 2013 at 6:02 PM, Hiller, Dean <<>>
Out of curiosity, why did cassandra choose to re-invent the wheel instead of using something
like google protobuf which spans multiple languages?
I see it as a step better than thrift since it is really only defining message format and
has all sorts of goodies with it.  I think you only need to frame it and that may exist already
as well actually but I can't remember.

The serialization/deserialization involved in the binary protocol is not a big deal tbh, so
I guess we chose to avoid the dependency. I personally don't think using protobufs would have
simplified things much in practice, I don't think there is that much wheel reinventing, and
so I'm reasonably happy to have something tailored to your needs. I'll admit there is some
subjectivity in that opinion however, your mileage may vary.

Lastly, does the java-driver have an asynch nature to it at all?

The java driver is completely asynchronous, from the protocol to it's implementation, so yes.

 It would be nice to be able to call driver.put(myData, myCallbackSuccessHandler);

In case you look at the driver API, its execute method returns a future, that happens to extends
guava's ListenableFuture, and so you can add a callback/listener through that.


From: Sylvain Lebresne <<><<>>>
Reply-To: "<><<>>"
Date: Tuesday, April 23, 2013 9:55 AM
To: "<><<>>"
Subject: Re: Prepared Statement - cache duration (CQL3 - Cassandra 1.2.4)

When we speak of "binary protocol", we talk about the protocol introduced in Cassandra 1.2
that is an alternative to thrift for CQL3. It's a custom, binary, protocol, that has not link
to thrift whatsoever.

That protocol is defined by the document here:;a=blob_plain;f=doc/native_protocol_v1.spec;hb=HEAD

Of course, this is just a protocol, and unless you have the time and willingness to write
a proper library using that protocol, you should just use an existing driver implementing
it. If you are using Java (some of your example above seems to be in Java), then you could
for instance pick If you're not using java, then
well, since said protocol is fairly recent, there isn't an existing driver for every languages,
but a bunch of drivers are in the work.

That being said, I'm not saying you *should* use a driver that uses the binary protocol, just
that at least for exceptions handling, said binary protocol has a slightly cleaner handling
of them than what's available through thrift. I'll not that even if you do want to use thrift,
it's usually advised to use a high level client rather than raw thrift. Unless you have no
choice or like suffering that is.


On Tue, Apr 23, 2013 at 5:38 PM, Stuart Broad <<><<>>>
Hi Edward,

My understanding was that thrift supports a number of protocols (binary being one of them).
 I don't understand what switching to "binary protocol" but not using thrift means.  Can you
point me to any code examples?



On Tue, Apr 23, 2013 at 4:21 PM, Edward Capriolo <<><<>>>
Having to catch the exception and parse it is a bit ugly, however this is close to what someone
might do with an SQLException to determine if the error was transient etc.  If there is an
error code it is possible that it could be added as an optional property of the InvalidRequestException
in future versions.

Switching to the "binany protocol" is not a method in thrift, it means your not using thrift
at all.

On Tue, Apr 23, 2013 at 11:13 AM, Stuart Broad <<><<>>>
Hi Edward,

Thanks for your reply - I was already using the prepare/execute cql methods that you suggested.
 My problem is that these methods 'mask' the PreparedQueryNotFoundException as an InvalidRequestException.
 At present I catch the InvalidRequestException (when cassandra has been re-started) and check
the message text to figure out if I need to rebuild the prepared queries (rather than building
each time I call).

Sylvain had suggested that I use the binary protocol as the exceptions are more explicit so
I am trying to determine how this can be done (I don't see any obvious methods other than
the cql ones for calling prepared statements).



On Tue, Apr 23, 2013 at 4:05 PM, Edward Capriolo <<><<>>>
Thrift has a prepare_cql call which returns an ID. Then it has an exececute_cql call which
takes the id and a map or variable bindings.

On Tue, Apr 23, 2013 at 10:29 AM, Stuart Broad <<><<>>>
Hi all,

I just realised that the binary protocol is the low-level thrift api that I was originally
using (Cassandra.Client>> get / insert ...).  How can a prepared statement be called
through the thrift api (i.e. not the cql methods)?



On Tue, Apr 23, 2013 at 11:48 AM, Stuart Broad <<><<>>>
Hi Sylvain,

Thanks for your response.  I am handling the 'PreparedQueryNotFoundException' more for the
case of a cassandra re-start (rather then expecting to build 100000 statements).

I am not familiar with the binary protocol - which class/methods should I look at?



On Tue, Apr 23, 2013 at 11:29 AM, Sylvain Lebresne <<><<>>>
In thrift, a lot of exceptions (like PreparedQueryNotFoundException) are simply returned as
InvalidRequestException. The reason for that was a mix of not wanting to change the thrift
API too much and because we didn't knew how to return a lot of different exception with thrift
without making it horrible to work with. So you'll probably have to parse strings here indeed.

This will be cleaner/less fragile if you use the binary protocol as exceptions are more fined
grained there.

Though taking a step back (and without saying that you shouldn't handle the case where a query
is not prepared on the node you contact), if you're really considering preparing more than
100000 statements, I'd suggest that it might be worth benchmarking whether using prepared
statements in your case is really going to be worth the trouble. Just saying.


On Tue, Apr 23, 2013 at 12:14 PM, Stuart Broad <<><<>>>
Hi Sorin,

The PreparedQueryNotFoundException is not thrown from Cassandra.Client>>execute_prepared_cql3_query
method.  I created some prepared statements and then re-started cassandra and received the
following exception:

InvalidRequestException(why: Prepared query with ID 1124421588 not found (either the query
was not prepared on this host (maybe the host has been restarted?) or you have prepared more
than 100000 queries and queries 1124421588 has been evicted from the internal cache))

The best I have been able to come up with is the following:

            try {
                client.execute_prepared_cql3_query(psId, bindValues, ..);
            } catch (InvalidRequestException invEx) {
                String why = invEx.getWhy();
                if(why.startsWith("Prepared query with ID")) {
                    client.execute_prepared_cql3_query(psId, bindValues, ..);
                } else {
                    throw invEx;

Obviously this is pretty fragile and would break if the cassandra message was changed...but
it least it works for now!



On Sun, Apr 21, 2013 at 11:51 AM, Sorin Manolache <<><<>>>
On 2013-04-19 13:57, Stuart Broad wrote:

I am using Cassandra.Client
prepare_cql3_query/execute_prepared_cql3_query to create and run some
prepared statements.  It is working well but I am unclear as to how long
the server side 'caches' the prepared statements.  Should a prepared
statement be prepared for every new Cassandra.Client?  Based on my
limited testing it seems like I can create some prepared statements in
one Cassandra.Client and use in another but I am not sure how
reliable/lasting this is i.e.  If I called the prepared statement again
the next day would it still exist?  What about if cassandra was re-started?


I am creating prepared statements for batch updates of pre-defined
lengths (e.g. 10000, 1000, 500, 250, 50, 10, 1) and wanted to know if
these could just be set up once.  We felt that using the prepared
statements was easier than escaping values within a CQL statement and
probably more performant.

Thanks in advance for your help.

I've looked in Cassandra's code (v1.2.3). The cache of prepared statements has a size of 100,000.
So if you prepare more than 100 thousand statements, the least recently used ones will vanish.
You'll get the exception PreparedQueryNotFoundException, code 0x2500.


View raw message