cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jonathan Ellis <jbel...@gmail.com>
Subject Re: repeated timeouts on quorum reads
Date Tue, 20 Oct 2009 02:23:14 GMT
On Mon, Oct 19, 2009 at 9:08 PM, Edmond Lau <edmond@ooyala.com> wrote:
> Why do we need 2 nodes to compute hashes in addition to another node
> for retrieving the data?  Then the system is reading all three copies
> instead of just a quorum of 2.  Why not just one node for a hash and
> one node for hash + data?

That way you can take whichever hash finishes first.

> Here are my debug logs from the node with 3 concurrent reads.  5
> seconds after the get_slice, the rpc times out.  3 seconds after the
> timeout, all the callbacks return.  I know that the entire supercolumn
> (~15MB) has to be read into memory, but 8 seconds for 3 reads seems
> extremely slow.

I agree.

A couple more questions:

How many data files are there across all the directories (on a single
node) for the CF in question?

Does it still take ~8s if you direct a CL.ONE query at one of the
nodes you know has the data (i.e., a local read)?

What does JConsole show about jvm memory usage during the query?  Is
it right up against the heap size limit?

Mime
View raw message