As an update to this thread, I modified the monitors exactly as Nate had suggested, and there have been no alerts since then (except in a case where we insanely loaded our cluster in test trying to determine our capacity)

 

The “All time blocked” numbers are still pretty high though, eg. on one of the nodes, tpstats reports:

Pool Name                    Active   Pending      Completed   Blocked  All time blocked

FlushWriter                       0         0          30044         0              5413

 

I monitored iostat during one of our regular test stress tests, and the "await" times reported by iostat seem pretty high, with averages over a minute being in the 20-30ms range, and going as high as 100ms. The "svctm" or service time is mostly around 2-3 ms. So the average time an IO request is waiting around compared to the time it takes to be serviced is very high.

The disk setup is three 1TB disks with software RAID 0, to appear as a single disk of 3TB.

Does this confirm that I have a slow disk problem? If so, any other ways to help other than moving to SSDs and adding more Cassandra nodes.

 

Thanks,

Arindam

 

From: Arindam Barua [mailto:abarua@247-inc.com]
Sent: Monday, July 01, 2013 2:48 PM
To: user@cassandra.apache.org
Subject: RE: about FlushWriter "All time blocked"

 

 

Thanks guys, these sound like good suggestions, will try those out.

 

Aaron, we have around 80 CFs.

 

From: aaron morton [mailto:aaron@thelastpickle.com]
Sent: Friday, June 28, 2013 10:05 PM
To: user@cassandra.apache.org
Subject: Re: about FlushWriter "All time blocked"

 

We do not use secondary indexes or snapshots

Out of interest how many CF's do you have ?

 

Cheers

 

-----------------

Aaron Morton

Freelance Cassandra Consultant

New Zealand

 

@aaronmorton

 

On 28/06/2013, at 7:52 AM, Nate McCall <zznate.m@gmail.com> wrote:

 

Non-zero for pending tasks is too transient. Try monitoring tpstats
with a (much) higher frequency and look for sustained threshold over a
duration.

Then, using a percentage of the configuration values for the max - 75%
of memtable_flush_queue_size in this case - alert when it has been
higher than '3' for more than N time. (Start with N=60 seconds and go
from there).

Also, that is a very high 'all time blocked' to 'completed' ratio for
FlushWriter. If iostat is happy, i'd do as Aaron suggested above and
turn up the memtable_flush_queue_size and play around with turning up
memtable_flush_writers (incrementally and separately for both of
course so you can see the effect).

On Thu, Jun 27, 2013 at 2:27 AM, Arindam Barua <abarua@247-inc.com> wrote:

In our performance tests, we are seeing similar FlushWriter, MutationStage, MemtablePostFlusher pending tasks become non-zero. We collect snapshots every 5 minutes, and they seem to clear after ~10-15 minutes though. (The flush writer has an 'All time blocked' count of 540 in the below example).

We do not use secondary indexes or snapshots. We do not use SSDs. We have a 4-node cluster with around 30-40 GB data on each node. Each node has 3 1-TB disks with a RAID 0 setup.

Currently we monitor the tpstats every 5 minutes, and alert if FlushWriter or MutationStage has a non-zero Pending count. Any suggestions if this is a cause of concern already, or, should we alert only if that count becomes greater than a bigger number, say 10, or if the count remains non-zero greater than a specified time.

Pool Name                    Active   Pending      Completed   Blocked  All time blocked
ReadStage                         0         0       15685133         0                 0
RequestResponseStage              0         0       29880863         0                 0
MutationStage                     0         0       40457340         0                 0
ReadRepairStage                   0         0         704322         0                 0
ReplicateOnWriteStage             0         0              0         0                 0
GossipStage                       0         0        2283062         0                 0
AntiEntropyStage                  0         0              0         0                 0
MigrationStage                    0         0             70         0                 0
MemtablePostFlusher               1         1           1837         0                 0
StreamStage                       0         0              0         0                 0
FlushWriter                       1         1           1446         0               540
MiscStage                         0         0              0         0                 0
commitlog_archiver                0         0              0         0                 0
InternalResponseStage             0         0             43         0                 0
HintedHandoff                     0         0              3         0                 0

Thanks,
Arindam

-----Original Message-----
From: aaron morton [mailto:aaron@thelastpickle.com]
Sent: Tuesday, June 25, 2013 10:29 PM
To: user@cassandra.apache.org
Subject: Re: about FlushWriter "All time blocked"

FlushWriter                       0         0            191         0                12


This means there were 12 times the code wanted to put an memtable in the queue to be flushed to disk but the queue was full.

The length of this queue is controlled by the memtable_flush_queue_size https://github.com/apache/cassandra/blob/cassandra-1.2/conf/cassandra.yaml#L299 and memtable_flush_writers .

When this happens an internal lock around the commit log is held which prevents writes from being processed.

In general it means the IO system cannot keep up. It can sometimes happen when snapshot is used as all the CF's are flushed to disk at once. I also suspect it happens sometimes when a commit log segment is flushed and their are a lot of dirty CF's. But i've never proved it.

Increase memtable_flush_queue_size following the help in the yaml file. If you do not use secondary indexes are you using snapshot?

Hope that helps.
A
-----------------
Aaron Morton
Freelance Cassandra Consultant
New Zealand

@aaronmorton
http://www.thelastpickle.com

On 24/06/2013, at 3:41 PM, yue.zhang <yue.zhang@chinacache.com> wrote:

3 node
cent os
CPU 8core memory 32GB
cassandra 1.2.5
my scenario: many counter incr, every node has one client program, performance is 400 wps /every clicent (it’s so slowly)

my question
Ø  nodetool tpstats
---------------------------------
Pool Name                    Active   Pending      Completed   Blocked  All time blocked
ReadStage                         0         0           8453         0                 0
RequestResponseStage              0         0      138303982         0                 0
MutationStage                     0         0      172002988         0                 0
ReadRepairStage                   0         0              0         0                 0
ReplicateOnWriteStage             0         0       82246354         0                 0
GossipStage                       0         0        1052389         0                 0
AntiEntropyStage                  0         0              0         0                 0
MigrationStage                    0         0              0         0                 0
MemtablePostFlusher               0         0            670         0                 0
FlushWriter                       0         0            191         0                12
MiscStage                         0         0              0         0                 0
commitlog_archiver                0         0              0         0                 0
InternalResponseStage             0         0              0         0                 0
HintedHandoff                     0         0             56         0                 0
-----------------------------------
FlushWriter “All time blocked”=12I restart the nodebut no useit’s normally ?

thx

-heipark