The other thing I thought about.. what's the distribution of Key-Values
that you're writing? Specifically, do many of the Keys sort "near" each
other. Similarly, do you notice excessive load on some tservers, but not
all (the "Tablet Servers" page on the Monitor is a good check)?
Consider the following: you have 10 tservers and you have 10 proxy
servers. The first thought is that 10 tservers should be plenty to
balance the load of those 10 proxy servers. However, a problem arises
when if the data that each of those proxy servers is writing happens to
reside on a _small number of tablet servers_. Thus, your 10 proxy
servers might only be writing to one or two tabletservers.
If you notice that you're getting skew like this (or even just know that
you're apt to have a situation where multiple clients might write data
that sorts close to one another), it would be a good idea to add splits
to your table before starting your workload.
e.g. if you consider that your Key-space is the numbers from 1 to 10,
and you have ten tservers, it would be a good idea to add splits 1, 2,
... 10, so that each tservers hosts at least one tablet (e.g. [1,2),
[2,3)... [10,+inf)). Having at least 5 or 10 tablets per tserver per
table (split according to the distribution of your data) might help ease
the load.
On 2/11/14, 10:47 AM, Diego Woitasen wrote:
Same results with 2G tserver.memory.maps.max.
May be we just reached the limit :)
On Mon, Feb 10, 2014 at 7:08 PM, Diego Woitasen
<[email protected]> wrote:
On Mon, Feb 10, 2014 at 6:21 PM, Josh Elser <[email protected]> wrote:
I assume you're running a datanode along side the tserver on that node? That
may be stretching the capabilities of that node (not to mention ec2 nodes
tend to be a little flakey in general). 2G for the tserver.memory.maps.max
might be a little safer.
You got an error in a tserver log about that IOException in internalReader.
After that, the tserver was still alive? And the proxy client was dead -
quit normally?
Yes, everything is still alive.
If that's the case, the proxy might just be disconnecting in a noisy manner?
Right!
I'll try with 2G tserver.memory.maps.max.
On 2/10/14, 3:38 PM, Diego Woitasen wrote:
Hi,
I tried increasing the tserver.memory.maps.max to 3G and failed
again, but with other error. I have a heap size of 3G and 7.5 GB of
total ram.
The error that I've found in the crashed tserver is:
2014-02-08 03:37:35,497 [util.TServerUtils$THsHaServer] WARN : Got an
IOException in internalRead!
The tserver haven't crashed, but the client was disconnected during the
test.
Another hint is welcome :)
On Mon, Feb 3, 2014 at 3:58 PM, Josh Elser <[email protected]> wrote:
Oh, ok. So that isn't quite as bad as it seems.
The "commits are held" exception is thrown when the tserver is running
low
on memory. The tserver will block new mutations coming in until it can
process the ones it already has and free up some memory. This makes sense
that you would see this more often when you have more proxy servers as
the
total amount of Mutations you can send to your Accumulo instance is
increased. With one proxy server, your tserver had enough memory to
process
the incoming data. With many proxy servers, your tservers would likely
fall
over eventually because they'll get bogged down in JVM garbage
collection.
If you have more memory that you can give the tservers, that would help.
Also, you should make sure that you're using the Accumulo native maps as
this will use off-JVM-heap space instead of JVM heap which should help
tremendously with your ingest rates.
Native maps should be on by default unless you turned them off using the
property 'tserver.memory.maps.native.enabled' in accumulo-site.xml.
Additionally, you can try increasing the size of the native maps using
'tserver.memory.maps.max' in accumulo-site.xml. Just be aware that with
the
native maps, you need to ensure that total_ram > JVM_heap +
tserver.memory.maps.max
- Josh
On 2/3/14, 1:33 PM, Diego Woitasen wrote:
I've launched the cluster again and I was able to reproduce the error:
In the proxy I had the same error that I mention in one of my previous
messages, about a failure in a table server. I checked the log of that
tablet server and I found:
2014-02-03 18:02:24,065 [thrift.ProcessFunction] ERROR: Internal error
processing update
org.apache.accumulo.server.tabletserver.HoldTimeoutException: Commits
are
held
A lot of times.
Full log if someone want to have a look:
http://www.vhgroup.net/diegows/tserver_matrix-slave-07.accumulo-ec2-test.com.debug.log
Regards,
Diego
On Mon, Feb 3, 2014 at 12:11 PM, Josh Elser <[email protected]>
wrote:
I would assume that that proxy service would become a bottleneck fairly
quickly and your throughput would benefit from running multiple
proxies,
but I don't have substantive numbers to back up that assertion.
I'll put this on my list and see if I can reproduce something.
On 2/3/14, 7:42 AM, Diego Woitasen wrote:
I have to run the tests again because they were ec2 instances and I've
destroyed. It's easy to reproduce BTW.
My question is, does it makes sense to run multiple proxies? Are there
a limit? Right now I'm trying with 10 nodes and 10 proxies (running on
every node). May be that doesn't make sense or it's a buggy
configuration.
On Fri, Jan 31, 2014 at 7:29 PM, Josh Elser <[email protected]>
wrote:
When you had multiple proxies, what were the failures on that tablet
server
(10.202.6.46:9997).
I'm curious why using one proxy didn't cause errors but multiple did.
On 1/31/14, 4:44 PM, Diego Woitasen wrote:
I've reproduced the error and I've found this in the proxy logs:
2014-01-31 19:47:50,430 [server.THsHaServer] WARN : Got an
IOException in internalRead!
java.io.IOException: Connection reset by peer
at sun.nio.ch.FileDispatcherImpl.read0(Native Method)
at
sun.nio.ch.SocketDispatcher.read(SocketDispatcher.java:39)
at
sun.nio.ch.IOUtil.readIntoNativeBuffer(IOUtil.java:223)
at sun.nio.ch.IOUtil.read(IOUtil.java:197)
at
sun.nio.ch.SocketChannelImpl.read(SocketChannelImpl.java:379)
at
org.apache.thrift.transport.TNonblockingSocket.read(TNonblockingSocket.java:141)
at
org.apache.thrift.server.AbstractNonblockingServer$FrameBuffer.internalRead(AbstractNonblockingServer.java:515)
at
org.apache.thrift.server.AbstractNonblockingServer$FrameBuffer.read(AbstractNonblockingServer.java:305)
at
org.apache.thrift.server.AbstractNonblockingServer$AbstractSelectThread.handleRead(AbstractNonblockingServer.java:202)
at
org.apache.thrift.server.TNonblockingServer$SelectAcceptThread.select(TNonblockingServer.java:198)
at
org.apache.thrift.server.TNonblockingServer$SelectAcceptThread.run(TNonblockingServer.java:154)
2014-01-31 19:51:13,185 [impl.ThriftTransportPool] WARN :
Server
10.202.6.46:9997:9997 (30000) had 20 failures in a short time
period,
will not complain anymore
A lot of this messages appear in all the proxies.
I tried the same stress tests agaisnt one proxy and I was able to
increase the load without getting any error.
Regards,
Diego
On Thu, Jan 30, 2014 at 2:47 PM, Keith Turner <[email protected]>
wrote:
Do you see more information in the proxy logs? "# exceptions 1"
indicates
an unexpected exception occured in the batch writer client code.
The
proxy
uses this client code, so maybe there will be a more detailed stack
trace
in
its logs.
On Thu, Jan 30, 2014 at 9:46 AM, Diego Woitasen
<[email protected]>
wrote:
Hi,
I'm testing with a ten node cluster with the proxy enabled in
all
the
nodes. I'm doing a stress test balancing the connection between
the
proxies using round robin. When I increase the load (400 workers
writting) I get this error:
AccumuloSecurityException:
AccumuloSecurityException(msg='org.apache.accumulo.core.client.MutationsRejectedException:
# constraint violations : 0 security codes: [] # server errors 0
#
exceptions 1')
The complete message is:
AccumuloSecurityException:
AccumuloSecurityException(msg='org.apache.accumulo.core.client.MutationsRejectedException:
# constraint violations : 0 security codes: [] # server errors 0
#
exceptions 1')
kvlayer-test client failed!
Traceback (most recent call last):
File "tests/kvlayer/test_accumulo_throughput.py", line 64,
in
__call__
self.client.put('t1', ((u,), self.one_mb))
File
"/home/ubuntu/kvlayer-env/local/lib/python2.7/site-packages/kvlayer-0.2.7-py2.7.egg/kvlayer/_decorators.py",
line 26, in wrapper
return method(*args, **kwargs)
File
"/home/ubuntu/kvlayer-env/local/lib/python2.7/site-packages/kvlayer-0.2.7-py2.7.egg/kvlayer/_accumulo.py",
line 154, in put
batch_writer.close()
File
"/home/ubuntu/kvlayer-env/local/lib/python2.7/site-packages/pyaccumulo_dev-1.5.0.2-py2.7.egg/pyaccumulo/__init__.py",
line 126, in close
self._conn.client.closeWriter(self._writer)
File
"/home/ubuntu/kvlayer-env/local/lib/python2.7/site-packages/pyaccumulo_dev-1.5.0.2-py2.7.egg/pyaccumulo/proxy/AccumuloProxy.py",
line 3149, in closeWriter
self.recv_closeWriter()
File
"/home/ubuntu/kvlayer-env/local/lib/python2.7/site-packages/pyaccumulo_dev-1.5.0.2-py2.7.egg/pyaccumulo/proxy/AccumuloProxy.py",
line 3172, in recv_closeWriter
raise result.ouch2
I'm not sure if the errror is produced by the way I'm using the
cluster with multiple proxies, may be I should use one.
Ideas are welcome.
Regards,
Diego
--
Diego Woitasen
VHGroup - Linux and Open Source solutions architect
www.vhgroup.net
--
Diego Woitasen
VHGroup - Linux and Open Source solutions architect
www.vhgroup.net
