Server load alone doesn't always indicate the server's ability to serve
queries. Memory and cache state are important too, and they're not as
easy to monitor. Additionally, server load at any single point in time
or a short term average is not indicative of the server's ability to
handle search requests if indexing happens in short but intense bursts.
It can also complicate things if there are more than one Solr instance
running on a single server.
I'm definitely not against intelligent routing. In many cases it makes
perfect sense, and I'd still like to use it, just limited to the pull
replicas.
--Ere
Erick Erickson kirjoitti 5.1.2018 klo 19.03:
Actually, I think a much better option is to route queries to server load.
The theory of preferring pull replicas to leaders would be that the leader
will be doing the indexing work and the pull replicas would be doing less
work therefore serving queries faster. But that's a fragile assumption.
Let's say indexing stops totally. Now your leader is sitting there idle
when it could be serving queries.
The autoscaling work will allow for more intelligent routing, you can
monitor the CPU load on your servers and if the leader has some spare
cycles use them .vs. crudely routing all queries to pull replicas (or tlog
replicas for that matter). NOTE: I don't know whether this is being
actively worked on or not, but seems a logical extension of the increased
monitoring capabilities being put in place for autoscaling, but I'd rather
see effort put in there than support routing based solely on a node's type.
Best,
Erick
On Fri, Jan 5, 2018 at 7:51 AM, Emir Arnautović <
emir.arnauto...@sematext.com> wrote:
It is interesting that ES had similar feature to prefer primary/replica
but it deprecating that and will remove it - could not find explanation why.
Emir
--
Monitoring - Log Management - Alerting - Anomaly Detection
Solr & Elasticsearch Consulting Support Training - http://sematext.com/
On 5 Jan 2018, at 15:22, Ere Maijala <ere.maij...@helsinki.fi> wrote:
Hi,
It would be really nice to have a server-side option, though. Not
everyone uses Solrj, and a typical fairly dummy client just queries the
server without any understanding about shards etc. Solr could be clever
enough to not forward the query to NRT shards when configured to prefer
PULL shards and they're available. Maybe it could be something similar to
the preferLocalShards parameter, like "preferShardTypes=TLOG,PULL".
--Ere
Emir Arnautović kirjoitti 14.12.2017 klo 11.41:
Hi Stanislav,
I don’t think that there is a built in feature to do this, but that
sounds like nice feature of Solrj - maybe you should check if available.
You can implement it outside of Solrj - check cluster state to see which
shards are available and send queries only to pull replicas.
HTH,
Emir
--
Monitoring - Log Management - Alerting - Anomaly Detection
Solr & Elasticsearch Consulting Support Training - http://sematext.com/
On 14 Dec 2017, at 09:58, Stanislav Sandalnikov <
s.sandalni...@gmail.com> wrote:
Hi,
We have a Solr 7.1 setup with SolrCloud where we have multiple shards
on one server (for indexing) each shard has a pull replica on other servers.
What are the possible ways to limit search request only to pull type
replicase?
At the moment the only solution I found is to append shards parameter
to each query, but if new shards added later it requires to change
solrconfig. Is it the only way to do this?
Thank you
Regards
Stanislav
--
Ere Maijala
Kansalliskirjasto / The National Library of Finland
--
Ere Maijala
Kansalliskirjasto / The National Library of Finland
