Ah ok. No that was not the case.
The client which did the long running scan didn't wait for the slowest node.
Only other clients that asked the slow node directly were affected.
Sorry about the confusion.
On 04.12.10 05:44, Jonathan Ellis wrote:
That makes sense, but this shouldn't make requests last for the
timeout duration -- at quorum, it should be responding to the client
as soon as it gets that second-fastest reply. If I'm understanding
right that this was making the response to the client block until the
overwhelmed node timed out, that's a bug. What version of Cassandra
is this?
On Fri, Dec 3, 2010 at 7:27 PM, Daniel Doubleday
<[email protected]> wrote:
Yes.
I thought that would make sense, no? I guessed that the quorum read forces
the slowest of the 3 nodes to keep the pace of the faster ones. But it cant.
No matter how small the performance diff is. So it will just fill up.
Also when saying 'practically dead' and 'never recovers' I meant for the
time I kept the reads up. As soon as I stopped the scan it recovered. It
just was not able to recover during the load because for that it would have
to become faster that the other nodes and with full queues that just
wouldn't happen.
By changing the node for every read I would hit the slower node every couple
of reads. This forced the client to wait for the slower node.
I guess to change that behavior you would need to use something like dynamic
snitch and ask only as many peer nodes as necessary to satisfy quorum and
only ask other nodes when reads fail. But that would probably increase
latency and cause whatever other problems. Since you probably don't want to
run the cluster at a load at which the weakest node of a replication group
can't keep up I don't think this is an issue at all.
Just wanted to prevent others shooting their own foot as I did.
On 03.12.10 23:36, Jonathan Ellis wrote:
Am I understanding correctly that you had all connections going to one
cassandra node, which caused one of the *other* nodes to die, and
spreading the connections around the cluster fixed it?
On Fri, Dec 3, 2010 at 4:00 AM, Daniel Doubleday
<[email protected]> wrote:
Hi all
I have found an anti pattern the other day which I wanted to share,
although its pretty special case.
Special case because our production cluster is somewhat strange: 3
servers, rf = 3. We do consistent reads/writes with quorum.
I did a long running read series (loads of reads as fast as I can) with
one connection. Since all queries could be handled by that node the overall
latency is determined by its own and the fastest second node (cause the
quorum is satisfied with 2 reads). What will happen than is that after a
couple of minutes one of the other two nodes will go in 100% io wait and
will drop most of its read messages. Leaving it practically dead while the
other 2 nodes keep responding at an average of ~10ms. The node that died was
only a little slower ~13ms average but it will inevitably queue up messages.
Average response time increases to timeout (10 secs) flat. It never
recovers.
It happened all the time. And it wasn't the same node that would die.
The solution was that I return the connection to the pool and get a new
one for every read to balance the load on the client side.
Obviously this will not happen in a cluster where the percentage of all
rows on one node is enough. But the same thing will probably happen if you
scan by continuos tokens (meaning that you will read from the same node a
long time).
Cheers,
Daniel Doubleday
smeet.com, Berlin