On 22/07/2011 18:08, Yang wrote:
btw, this "issue" of not knowing whether a write is persisted or not
when client reports error, is not limited to counters, for regular
columns, it's the same: if client reports write failure, the value may
well be replicated to all replicas later. this is even the same with
all other systems: Zookeeper, Paxos, ultimately due to the FLP
theoretical result of "no guarantee of consensus in async systems"
yes, but with regular columns, retry is OK, while counter is not.
---------- Forwarded message ----------
From: Sylvain Lebresne<sylv...@datastax.com>
Date: Fri, Jul 22, 2011 at 8:03 AM
Subject: Re: Counter consistency - are counters idempotent?
To: user@cassandra.apache.org
On Fri, Jul 22, 2011 at 4:52 PM, Kenny Yu<kenny...@knewton.com> wrote:
As of Cassandra 0.8.1, are counter increments and decrements idempotent? If,
for example, a client sends an increment request and the increment occurs,
but the network subsequently fails and reports a failure to the client, will
Cassandra retry the increment (thus leading to an overcount and inconsistent
data)?
I have done some reading and I am getting conflicting sources about counter
consistency. In this source
(http://cassandra-user-incubator-apache-org.3065146.n2.nabble.com/clarification-of-the-consistency-guarantees-of-Counters-td6421010.html),
it states that counters now have the same consistency as regular
columns--does this imply that the above example will not lead to an
overcount?
That email thread was arguably a bit imprecise with its use of the
word 'consistency'
but what it was talking about is really consistency level. That is, counter
supports all the usual consistency levels (ONE, QUORUM, ALL, LOCAL_QUORUM,
EACH_QUORUM) excepted ANY.
Counter are still not idempotent. And just a small precision, if you
get a TimeoutException,
Cassandra never retry the increment on it's own (your sentence
suggests it does),
but you won't know in that case if the increment was persisted or not,
and thus you
won't know if you should retry or not. And yes, this is still a
limitation of counters.
If counters are not idempotent, are there examples of effective uses of
counters that will prevent inconsistent counts?
Thank you for your help.
--
Donal Zang
Computing Center, IHEP
19B YuquanLu, Shijingshan District,Beijing, 100049
zan...@ihep.ac.cn
86 010 8823 6018
