elasticsearch custom_score multiplication is inaccurate - elasticsearch

Question

I've inserted some documents which are all identical except for one floating-point field, called a.
When script of a custom_score query is set to just _score, the resulting score is 0.40464813 for a particular query matching some fields. When script is then changed to _score * a (mvel) for the same query, where a is 9.908349251612433, the final score becomes 4.0619955.
Now, if I run this calculation via Chrome's JS console, I get 4.009394996051871.
4.0619955 (elasticsearch)
4.009394996051871 (Chrome)
This is quite a difference and produces an incorrect ordering of results. Why could it be, and is there a way to correct it?

Answer 1

If I run a simple calculation using the numbers you provided, then I get the result that you expect.
curl -XPOST 'http://127.0.0.1:9200/test/test?pretty=1' -d '
{
"a" : 9.90834925161243
}
'
curl -XGET 'http://127.0.0.1:9200/test/test/_search?pretty=1' -d '
{
"query" : {
"custom_score" : {
"script" : "0.40464813 *doc[\u0027a\u0027].value",
"query" : {
"match_all" : {}
}
}
}
}
'
# {
# "hits" : {
# "hits" : [
# {
# "_source" : {
# "a" : 9.90834925161243
# },
# "_score" : 4.009395,
# "_index" : "test",
# "_id" : "lPesz0j6RT-Xt76aATcFOw",
# "_type" : "test"
# }
# ],
# "max_score" : 4.009395,
# "total" : 1
# },
# "timed_out" : false,
# "_shards" : {
# "failed" : 0,
# "successful" : 5,
# "total" : 5
# },
# "took" : 1
# }
I think what you are running into here is testing too little data across multiple shards.
Doc frequencies are calculated per shard by default. So if you have two identical docs on shard_1 and one doc on shard_2, then the docs on shard_1 will score lower than the docs on shard_2.
With more data, the document frequencies tend to even out over shards. But when testing small amounts of data you either want to create an index with only one shard, or to add search_type=dfs_query_then_fetch to the query string params.
This calculates global doc frequencies across all involved shards before calculating the scores.
If you set explain to true in your query, then you can see exactly how your scores are being calculated