Hi Dimitris,
statestore_frequency_update_ms not change during the update, is 2000ms
by default. And not more frequent REFRESH statement.
Actually we have 2 problem here:
1. the size of the serialized metadata of some tables exceed the maximum
allowed size of a byte array (2GB). The potential reason are :
a) Before the upgrading, the number of partitions and metadata of per
table are added day by day , but once we upgrade and restart the catalog, the
entire metadata of per table is transferred to catalog, so this transferring
exceed the maximum allowed size of a byte array, does it make sense?
2. The memory usage of catalog server increase continuously , and before
the upgrade, catalog server only use 20G memory but now it reach our limit
whether the limit is 32G or 96G. So I doubt there is a memory leak or
inefficient GC
And I see the issue https://issues.cloudera.org/browse/IMPALA-3910, seems that
it is related to my problem, but why the memory usage of catalog server
increase continuously after upgrading.
Thanks
Songbo
发件人: Dimitris Tsirogiannis [mailto:dtsirogian...@cloudera.com]
发送时间: 2016年9月23日 11:33
收件人: user@impala.incubator.apache.org
抄送: Matthew Jacobs
主题: Re: 答复: Catalog Server Out of memory when upgrading from CDH 5.5 -> 5.7
That error message indicates that the size of the serialized metadata of some
tables exceed the maximum allowed size of a byte array (2GB) and doesn't have
anything to do with the size of the heap. The logs indicate a large number of
refresh tables that could trigger this kind of errors, especially if these
tables have large metadata and/or incremental stats. You may want to see if
there are any changes in the workload (e.g. more frequent REFRESH statements
than before) and if the value of the statestore_frequency_update_ms was changed
after the upgrade.
On Thu, Sep 22, 2016 at 7:24 PM, 廖松博
<liaoson...@gridsum.com<mailto:liaoson...@gridsum.com>> wrote:
Hi,
The event is the upgrading from CDH 5.5 -> 5.7. And the Operation is
the regular operations with Impala 2.3 which we work well with for a long time
, but after the upgrading, it consume so much memory.
It seems a memory leak, because as follow chart:
[cid:image001.jpg@01D21582.157D2320]
when we use impala 2.3, it use 20G memory more or less, but when we upgrade to
2.5 and set Java Heap Size of Catalog Server in Bytes = 32 G, but the catalog
server consume the 32G rapidly, and we reset Java Heap Size of Catalog Server
in Bytes to 64G , it also increase to 64G in a not long time.
So I double if it is a memory leak and I see the issue
https://issues.cloudera.org/browse/IMPALA-3568 , is it the root cause ?
And the attached is the catalog server log and it contains the :
java.lang.OutOfMemoryError: Requested array size exceeds VM limit.
1. Is the root cause of the memory leak?
2. Any way to find the table which cause the exception?
Thanks.
Songbo
发件人: Dimitris Tsirogiannis
[mailto:dtsirogian...@cloudera.com<mailto:dtsirogian...@cloudera.com>]
发送时间: 2016年9月23日 8:49
收件人: user@impala.incubator.apache.org<mailto:user@impala.incubator.apache.org>
抄送: Matthew Jacobs
主题: Re: Catalog Server Out of memory when upgrading from CDH 5.5 -> 5.7
There seem to be some events after sept 21 that trigger a spike in memory
usage. Can you identify from CM what these events might be? For example,
refreshing tables with lots of metadata. Looking at the catalog logs might also
give you some hints on what the system is doing during those memory spikes.
Dimitris
On Thursday, September 22, 2016, 廖松博
<liaoson...@gridsum.com<mailto:liaoson...@gridsum.com>> wrote:
Hi guys,
We upgrade out CDH cluster from 5.5 to 5.7 at Sep 20, and from the
following figure:
[cid:image001.jpg@01D21582.157D2320]
You can see the catalog server in impala 2.5 use much more memory than
what in impala 2.3 and causes the out of memory error.
We set the Load Catalog In Background = false and restart the service
but not help.
We set Java Heap Size of Catalog Server in Bytes from 32G to 64G ,but
the memory usage also increase to 64G.
Is there any bugs in catalog server for impala 2.5? Any suggestion for
us?
Thanks.
Songbo
