Jesper Krogh <jesper@krogh.cc> writes:
> Conceptually searching for the "full dataset" would always be fastest
> solved by a seq-scan. The query planner enforces this so much, so not
> even "enable_seqscan=off" can convince it to to something else.
> ...
> Would it be possible to implement the "Filtering" using the gin-index and
> a subsequent visibillity-check as on the index-scan?
You're failing to make any sense whatsoever. If you're reading the full
dataset, there is no filter condition. If there is a potentially
indexable filter condition, the planner will certainly consider that.
Personally I think the issue here has got more to do with the
non-immutability of the single-argument form of to_tsquery, which means
it gets re-evaluated at every row during a seqscan. Do your results
change if you work with to_tsquery('english', ...) (or whatever your
default TS config is)?
regards, tom lane