I've emailed you a raw log file of an instance of this happening.
I've been monitoring more closely the timing of events in tpstats and the
logs and I believe this is what is happening:
- For some reason, C* decides to provoke a flush storm (I say some reason,
I'm sure there is one but I have had difficulty determining the behaviour
changes between 1.* and more recent releases).
- So we see ~ 3000 flush being enqueued.
- This happens so suddenly that even boosting the number of flush writers
to 20 does not suffice. I don't even see "all time blocked" numbers for it
before C* stops responding. I suspect this is due to the sudden OOM and GC
occurring.
- The last tpstat that comes back before the node goes down indicates 20
active and 3000 pending and the rest 0. It's by far the anomalous activity.
Is there a way to throttle down this generation of Flush? C* complains if I
set the queue_size to any value (deprecated now?) and boosting the threads
does not seem to help since even at 20 we're an order of magnitude off.
Suggestions? Comments?
On Sun, Oct 26, 2014 at 2:26 AM, DuyHai Doan <doanduy...@gmail.com> wrote:
> Hello Maxime
>
> Can you put the complete logs and config somewhere ? It would be
> interesting to know what is the cause of the OOM.
>
> On Sun, Oct 26, 2014 at 3:15 AM, Maxime <maxim...@gmail.com> wrote:
>
>> Thanks a lot that is comforting. We are also small at the moment so I
>> definitely can relate with the idea of keeping small and simple at a level
>> where it just works.
>>
>> I see the new Apache version has a lot of fixes so I will try to upgrade
>> before I look into downgrading.
>>
>>
>> On Saturday, October 25, 2014, Laing, Michael <michael.la...@nytimes.com>
>> wrote:
>>
>>> Since no one else has stepped in...
>>>
>>> We have run clusters with ridiculously small nodes - I have a production
>>> cluster in AWS with 4GB nodes each with 1 CPU and disk-based instance
>>> storage. It works fine but you can see those little puppies struggle...
>>>
>>> And I ran into problems such as you observe...
>>>
>>> Upgrading Java to the latest 1.7 and - most importantly - *reverting to
>>> the default configuration, esp. for heap*, seemed to settle things down
>>> completely. Also make sure that you are using the 'recommended production
>>> settings' from the docs on your boxen.
>>>
>>> However we are running 2.0.x not 2.1.0 so YMMV.
>>>
>>> And we are switching to 15GB nodes w 2 heftier CPUs each and SSD storage
>>> - still a 'small' machine, but much more reasonable for C*.
>>>
>>> However I can't say I am an expert, since I deliberately keep things so
>>> simple that we do not encounter problems - it just works so I dig into
>>> other stuff.
>>>
>>> ml
>>>
>>>
>>> On Sat, Oct 25, 2014 at 5:22 PM, Maxime <maxim...@gmail.com> wrote:
>>>
>>>> Hello, I've been trying to add a new node to my cluster ( 4 nodes ) for
>>>> a few days now.
>>>>
>>>> I started by adding a node similar to my current configuration, 4 GB or
>>>> RAM + 2 Cores on DigitalOcean. However every time, I would end up getting
>>>> OOM errors after many log entries of the type:
>>>>
>>>> INFO [SlabPoolCleaner] 2014-10-25 13:44:57,240
>>>> ColumnFamilyStore.java:856 - Enqueuing flush of mycf: 5383 (0%) on-heap, 0
>>>> (0%) off-heap
>>>>
>>>> leading to:
>>>>
>>>> ka-120-Data.db (39291 bytes) for commitlog position
>>>> ReplayPosition(segmentId=1414243978538, position=23699418)
>>>> WARN [SharedPool-Worker-13] 2014-10-25 13:48:18,032
>>>> AbstractTracingAwareExecutorService.java:167 - Uncaught exception on thread
>>>> Thread[SharedPool-Worker-13,5,main]: {}
>>>> java.lang.OutOfMemoryError: Java heap space
>>>>
>>>> Thinking it had to do with either compaction somehow or streaming, 2
>>>> activities I've had tremendous issues with in the past; I tried to slow
>>>> down the setstreamthroughput to extremely low values all the way to 5. I
>>>> also tried setting setcompactionthoughput to 0, and then reading that in
>>>> some cases it might be too fast, down to 8. Nothing worked, it merely
>>>> vaguely changed the mean time to OOM but not in a way indicating either was
>>>> anywhere a solution.
>>>>
>>>> The nodes were configured with 2 GB of Heap initially, I tried to crank
>>>> it up to 3 GB, stressing the host memory to its limit.
>>>>
>>>> After doing some exploration (I am considering writing a Cassandra Ops
>>>> documentation with lessons learned since there seems to be little of it in
>>>> organized fashions), I read that some people had strange issues on
>>>> lower-end boxes like that, so I bit the bullet and upgraded my new node to
>>>> a 8GB + 4 Core instance, which was anecdotally better.
>>>>
>>>> To my complete shock, exact same issues are present, even raising the
>>>> Heap memory to 6 GB. I figure it can't be a "normal" situation anymore, but
>>>> must be a bug somehow.
>>>>
>>>> My cluster is 4 nodes, RF of 2, about 160 GB of data across all nodes.
>>>> About 10 CF of varying sizes. Runtime writes are between 300 to 900 /
>>>> second. Cassandra 2.1.0, nothing too wild.
>>>>
>>>> Has anyone encountered these kinds of issues before? I would really
>>>> enjoy hearing about the experiences of people trying to run small-sized
>>>> clusters like mine. From everything I read, Cassandra operations go very
>>>> well on large (16 GB + 8 Cores) machines, but I'm sad to report I've had
>>>> nothing but trouble trying to run on smaller machines, perhaps I can learn
>>>> from other's experience?
>>>>
>>>> Full logs can be provided to anyone interested.
>>>>
>>>> Cheers
>>>>
>>>
>>>
>
