Am 15.05.2018 um 14:33 schrieb Erick Erickson:
You might find this useful:
https://lucidworks.com/2015/10/05/really-batch-updates-solr-2/
I have seen that already and can confirm it.
From my observations about a 3x3 cluster with 3 server and my hardware:
- have at least 6 CPUs on each server to keep search performance during NRT
indexing
- I tried with batch-/queue-size between 100 and 10000.
--- with a batch size of 100 and nearly even distribution accross 3
shards I get about 33 docs per update per shard.
--- with a batch size of 1000 I get about 333 docs per update per shard
--- with a batch size of 10000 it can go up to 3333 docs per shard
Yes, the last is "it can go up to" because the size is obviuosly to high
and I get lots of smaler updates "FROMLEADER". So somewhere between
1000 and 10000 is the best size for my 3x3 cluster with my hardware.
Another observation in a 3x3 cluster, a multi-node (3 JVM 4G instances per
server [3 nodes]) outperforms a multi-core (1 JVM 12G instance per
server [3 cores]) due to JAVA GC impact at multi-core.
A multi-node at 60qps has nearly the performance as a multi-core at 30qps.
One tricky bit: Assuming docs have a random distribution amongst
shards, you should batch so at least 100 docs go to each _shard_. You
can see from the link that the speedup is mostly going from 1 to 100.
So if you have 5 shards, I'd create batches of at least 500. That was
a fairly simple test with stupid-simple docs. Large complicated
documents wouldn't show the same curve.
Setup for PULL and TLOG isn't hard, just specify the number of TLOG or
PULL replicas you want at collection creation time. NOTE: this is only
on Solr 7x. See:
https://lucene.apache.org/solr/guide/7_3/shards-and-indexing-data-in-solrcloud.html#types-of-replicas
Unfortunately I'm still at solr 6.4.2 and therefore have to stay with NRT.
About creating your own queue, mine usually look like
List<SolrInputDocument> list...
while (more docs) {
list.add(new_doc);
if (list.size > X) {
client.add(list);
list.clear();
}
}
Yes, mine looks similar, a recursive file traverser with for-loop over files.
But don't forget a final client.add(list) after the while-loop ;-)
Not exactly a sophisticated queue ;).....
On Tue, May 15, 2018 at 8:15 AM, Bernd Fehling
<bernd.fehl...@uni-bielefeld.de> wrote:
Hi Erik,
yes indeed, batching solved it.
I used ConcurrentUpdateSolrClient with queue size of 10000 but
CloudSolrClient doesn't have this feature.
I build my own queue now.
Ah!!! So I obviously use default NRT but actually don't need it because
I don't have any NRT data to index. A latency of several hours is OK for me.
Currently I'm testing with a 3x3 core-cluster (3 server, 3 cores per
server).
I also tested with 3x3 node-cluster (3 server, 3 nodes per server) which
performed
better, less influence of GarbageCollection.
I have to read more about PULL or TLOG replicas, how to set this up and so
on.
If it is to complex I will go with NRT and indexing is anyway during the
night.
Thanks for pointing this out.
Regards,
Bernd
Am 15.05.2018 um 13:28 schrieb Erick Erickson:
What did you do to solve your performance problem?
Batching updates is one thing that helps performance.
bq. I thought that only the leaders are under load
until any commit and then replicate to the other replicas.
True if (and only if) you're using PULL or TLOG replicas.
When using the default NRT replicas, every replica indexes
the docs, it doesn't matter whether they are the leader or replica.
That's required for NRT. Using CloudSolrClient has no bearing
on that functionality.
Best,
Erick
On Tue, May 15, 2018 at 6:53 AM, Bernd Fehling
<bernd.fehl...@uni-bielefeld.de> wrote:
Thanks, solved, performance is good now.
Regards,
Bernd
Am 15.05.2018 um 08:12 schrieb Bernd Fehling:
OK, I have the CloudSolrClient with SolrJ now running but it seams
a bit slower compared to ConcurrentUpdateSolrClient.
This was not expected.
The logs show that CloudSolrClient send the docs only to the leaders.
So the only advantage of CloudSolrClient is that it is "Cloud aware"?
With ConcurrentUpdateSolrClient I get about 1600 docs/sec for loading.
With CloudSolrClient I get only about 1200 docs/sec.
The system monitoring shows that with CloudSolrClient all nodes and
cores
are under heavy load. I thought that only the leaders are under load
until any commit and then replicate to the other replicas.
And that the replicas which are no leader have capacity to answer search
requests.
I think I still don't get the advantage of CloudSolrClient?
Regards,
Bernd
Am 09.05.2018 um 19:15 schrieb Erick Erickson:
You may not need to deal with any of this.
The default CloudSolrClient call creates a new LBHttpSolrClient for
you. So unless you're doing something custom with any LBHttpSolrClient
you create, you don't need to create one yourself.
Second, the default for CloudSolrClient.add() is to take the list of
documents you provide into sub-lists that consist of the docs destined
for a particular shard and sends those to the leader.
Do the default not work for you?
Best,
Erick
On Wed, May 9, 2018 at 2:54 AM, Bernd Fehling
<bernd.fehl...@uni-bielefeld.de> wrote:
Hi list,
while going from single core master/slave to cloud multi core/node
with leader/replica I want to change my SolrJ loading, because
ConcurrentUpdateSolrClient isn't cloud aware and has performance
impacts.
I want to use CloudSolrClient with LBHttpSolrClient and updates
should only go to shard leaders.
Question, what is the difference between sendUpdatesOnlyToShardLeaders
and sendDirectUpdatesToShardLeadersOnly?
Regards,
Bernd
--
*************************************************************
Bernd Fehling Bielefeld University Library
Dipl.-Inform. (FH) LibTec - Library Technology
Universitätsstr. 25 and Knowledge Management
33615 Bielefeld
Tel. +49 521 106-4060 bernd.fehling(at)uni-bielefeld.de
https://www.ub.uni-bielefeld.de/~befehl/
BASE - Bielefeld Academic Search Engine - www.base-search.net
*************************************************************
