Is it possible that all your requests are routed to that single shard?
I.e. you are not using the smart client that round-robins requests? I
think that could cause all of the merging of results to be done on a
single node.
Also - is it possible you have a "bad" document in that shard? Like one
that has a GB stored field or something?
-Mike
On 5/31/2014 5:54 PM, Joe Gresock wrote:
Interesting thought about the routing. Our document ids are in 3 parts:
<10-digit identifier>!<epoch timestamp>!<format>
e.g., 5/12345678!130000025603!TEXT
Each object has an identifier, and there may be multiple versions of the
object, hence the timestamp. We like to be able to pull back all of the
versions of an object at once, hence the routing scheme.
The nature of the identifier is that a great many of them begin with a
certain number. I'd be interested to know more about the hashing scheme
used for the document routing. Perhaps the first character gives it more
weight as to which shard it lands in?
It seems strange that certain of the most highly-searched documents would
happen to fall on this shard, but you may be onto something. We'll scrape
through some non-distributed queries and see what we can find.
On Sat, May 31, 2014 at 1:47 PM, Erick Erickson <erickerick...@gmail.com>
wrote:
This is very weird.
Are you sure that all the Java versions are identical? And all the JVM
parameters are the same? Grasping at straws here.
More grasping at straws: I'm a little suspicious that you are using
routing. You say that the indexes are about the same size, but is it is
possible that your routing is somehow loading the problem shard abnormally?
By that I mean somehow the documents on that shard are different, or have a
drastically higher number of hits than the other shards?
You can fire queries at shards with &distrib=false and NOT have it go to
other shards, perhaps if you can isolate the problem queries that might
shed some light on the problem.
Best
er...@baffled.com
On Sat, May 31, 2014 at 8:33 AM, Joe Gresock <jgres...@gmail.com> wrote:
It has taken as little as 2 minutes to happen the last time we tried. It
basically happens upon high query load (peak user hours during the day).
When we reduce functionality by disabling most searches, it stabilizes.
So it really is only on high query load. Our ingest rate is fairly low.
It happens no matter how many nodes in the shard are up.
Joe
On Sat, May 31, 2014 at 11:04 AM, Jack Krupansky <
j...@basetechnology.com>
wrote:
When you restart, how long does it take it hit the problem? And how
much
query or update activity is happening in that time? Is there any other
activity showing up in the log?
If you bring up only a single node in that problematic shard, do you
still
see the problem?
-- Jack Krupansky
-----Original Message----- From: Joe Gresock
Sent: Saturday, May 31, 2014 9:34 AM
To: solr-user@lucene.apache.org
Subject: Uneven shard heap usage
Hi folks,
I'm trying to figure out why one shard of an evenly-distributed 3-shard
cluster would suddenly start running out of heap space, after 9+ months
of
stable performance. We're using the "!" delimiter in our ids to
distribute
the documents, and indeed the disk size of our shards are very similar
(31-32GB on disk per replica).
Our setup is:
9 VMs with 16GB RAM, 8 vcpus (with a 4:1 oversubscription ratio, so
basically 2 physical CPUs), 24GB disk
3 shards, 3 replicas per shard (1 leader, 2 replicas, whatever). We
reserve 10g heap for each solr instance.
Also 3 zookeeper VMs, which are very stable
Since the troubles started, we've been monitoring all 9 with jvisualvm,
and
shards 2 and 3 keep a steady amount of heap space reserved, always
having
horizontal lines (with some minor gc). They're using 4-5GB heap, and
when
we force gc using jvisualvm, they drop to 1GB usage. Shard 1, however,
quickly has a steep slope, and eventually has concurrent mode failures
in
the gc logs, requiring us to restart the instances when they can no
longer
do anything but gc.
We've tried ruling out physical host problems by moving all 3 Shard 1
replicas to different hosts that are underutilized, however we still
get
the same problem. We'll still be working on ruling out infrastructure
issues, but I wanted to ask the questions here in case it makes sense:
* Does it make sense that all the replicas on one shard of a cluster
would
have heap problems, when the other shard replicas do not, assuming a
fairly
even data distribution?
* One thing we changed recently was to make all of our fields stored,
instead of only half of them. This was to support atomic updates. Can
stored fields, even though lazily loaded, cause problems like this?
Thanks for any input,
Joe
--
I know what it is to be in need, and I know what it is to have plenty.
I
have learned the secret of being content in any and every situation,
whether well fed or hungry, whether living in plenty or in want. I can
do
all this through him who gives me strength. *-Philippians 4:12-13*
--
I know what it is to be in need, and I know what it is to have plenty. I
have learned the secret of being content in any and every situation,
whether well fed or hungry, whether living in plenty or in want. I can
do
all this through him who gives me strength. *-Philippians 4:12-13*
