在ArangoDb群集上进行了两个FULLTEXT搜索：涉及到V8

Question

I am investigating ArangoDb cluster and found out that in case of usage two FULLTEXT() searches one of them involves V8 engine. My data:

[
{
"TITL": "Attacks induced by bromocryptin in Parkinson patients",
"WORD": [
        "hascites",
        "Six patients with Parkinson's disease"
         ],
"ID":1,
},
{
"TITL": "Linear modeling of possible mechanisms for Parkinson tremor generation",
"WORD": [
        "hascites",
        "jsubsetIM"
         ],
"ID":2,
},
{
"TITL": "Drug-induced parkinsonism in the rat- a model for biochemical ...",
"WORD": [
        "hascites",
        "Following treatment with reserpine or alternatively with ...",
        "hasabstract"
        ],
"ID":3,
}
]

Simplest query:

FOR title IN FULLTEXT(pmshort,"TITL","parkinson")
    FOR word IN FULLTEXT(pmshort,"WORD","hascites")
        FILTER title.ID==word.ID
    RETURN title

In other words, I am trying to find all documents that have parkinson in TITL and hascites in WORD . This example is seriously simplified, so the usage of something like

FILTER word.WORD=='hascites'

is not possible. Two or more FULLTEXT searches are required for providing the necessary functionality. Collection includes about 520,000 documents. FullText indexes are set up on each field.

I found out that each of FULLTEXT queries, being run separately, involves index:

Execution plan:
 Id   NodeType        Site         Est.   Comment
  1   SingletonNode   DBS             1   * ROOT
  5   IndexNode       DBS        526577     - FOR title IN pmshort   /* fulltext index scan */
  8   RemoteNode      COOR       526577       - REMOTE
  9   GatherNode      COOR       526577       - GATHER 
  4   ReturnNode      COOR       526577       - RETURN title

But in case of usage both FOR first one is being processed by V8 (JavaScript) and runs on coordinator, not DBS:

Execution plan:
 Id   NodeType            Site           Est.   Comment
  1   SingletonNode       COOR              1   * ROOT
  2   CalculationNode     COOR              1     - LET #2 = FULLTEXT(pmshort   /* all collection documents */, "TITL", "parkinson")   /* v8 expression */
  3   EnumerateListNode   COOR            100     - FOR title IN #2   /* list iteration */
 10   ScatterNode         COOR            100       - SCATTER
 11   RemoteNode          DBS             100       - REMOTE
  9   IndexNode           DBS        52657700       - FOR word IN pmshort   /* fulltext index scan */
  6   CalculationNode     DBS        52657700         - LET #6 = (title.`ID` == word.`ID`)   /* simple expression */   /* collections used: word : pmshort */
  7   FilterNode          DBS        52657700         - FILTER #6
 12   RemoteNode          COOR       52657700         - REMOTE
 13   GatherNode          COOR       52657700         - GATHER 
  8   ReturnNode          COOR       52657700         - RETURN title

Of course, this slows down system a lot. So my questions are: 1. Why ArangoDb cluster can't process both conditions on DBS, not on coordinator (COOR)? 2. How to avoid such situation since performance drops 300-500 times? 3. May be somebody can point on some additional materials to read about this.

Any help is appreciated. Thanks!

Answer 1

It looks like the query optimizer stops looking for further fulltext improvements after having applied one fulltext transformation in each query/subquery.

A potential fix for this can be found in this pull request (which targets 3.3.10).

Answer 2

Thanks a lot! It should be available in 3.3.10 and future 3.4, right?