lucene-solr-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Erick Erickson <erickerick...@gmail.com>
Subject Re: SolrCloud query results order master vs replica
Date Mon, 03 Feb 2014 19:19:17 GMT
This should only be happening if the scores are _exactly_ the same,
which is actually
quite rare. In that case, the tied scores are broken by the internal
Lucene document
ID, and the relative order of the docs on the two machines isn't
guaranteed to be the
same, the internal ID can change during segment merging, which is NOT the same
on both machines.

But this should be relatively rare. If you're doing *:* queries or
other such, then they
aren't scored (see ConstantScoreQuery). So in practical terms, I suspect you're
seeing some kind of test artifact. Try adding &debug=all to the query
and you'll see
how documents are scored.

Best,
Erick

On Mon, Feb 3, 2014 at 6:57 AM, M. Flatterie <nicflatterie@yahoo.com> wrote:
> Greetings,
>
> My setup is:
> - SolrCloud V4.3
> - On collection
> - one shard
> - 1 master, 1 replica
>
> so each instance contains the entire index.  The index is rather small and the replica
is used for robustness.  There is no need (IMHO) to split shard the index (yet, until the
index gets bigger).
>
> My question:
> - if I do a query on a product name (that is what the index is about) on the master I
get a certain number of results and the documents.
> - if I do the same query on the replica, I get the same number of results but the docs
are in a different order.
> - I do not specify a sort parameter in my query, simply a q=<product name>.
> - obviously if I force a sort order, everything is ok, same results, same order from
both instances.
> - am I wrong in expecting the same results, in the SAME order?
>
> Follow up question if the order is not guaranteed:
> - should I force the dev. to use an explicit sort order?
> - if we force the sort, we then bypass the ranking / score order do we not?
> - should I force all queries to go to the master and fall back on the replica only in
the context of a total loss of the master?
>
> Other useful information:
>   - the admin page shows same number of documents in both instances.
>   - logs are clean, load and replication and queries worked ok.
>   - the web application that queries SOLR round robins between the two instances, so
getting results in a different order is bad for consistency.
>
> Thank you for your help!
>
> Nic
>

Mime
View raw message