Обсуждение: tsvector field length limitation

Hi folks —

We utilize Postgres’ full text search system pretty heavily in our team’s operations and often index tens of millions
ofrecords with varying lengths of text. In most cases, the text we need to index is pretty short (no more than.
hundredsof words) but in rare cases a single record is very very long (high hundreds of thousands of words or longer).
Withthose records, we run into the max tsvector length requirement "The length of a tsvector (lexemes + positions) must
beless than 1 megabyte” 

I understand the performance implications of having very long tsvectors (our gin index updates are pretty terrible in
somecases) but would really appreciate it if the max tsvector length were larger (even 5MB would make a huge
difference)or if that error were a stern warning rather than a hard error. 

Is there any way to disable or increase that limit in Postgres 10.3? Perhaps in a future version?

Thank you!
Jonathan

Jonathan Marks <jonathanaverymarks@gmail.com> writes:
> ... we run into the max tsvector length requirement "The length of a tsvector (lexemes + positions) must be less than
1megabyte” 

> Is there any way to disable or increase that limit in Postgres 10.3?

No; it's forced by the representation used for tsvector, which stores
lexeme offsets in 20-bit fields (cf WordEntry in
src/include/tsearch/ts_type.h).  Perhaps that was short-sighted but
I don't foresee it changing anytime soon.  You'd more or less need
a whole new datatype ("bigtsvector"?) to make it happen.

            regards, tom lane

What if we just didn’t use positional arguments at all? I.e. we just populate the tsvector with lexemes only?

> On Jun 20, 2018, at 10:49 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Jonathan Marks <jonathanaverymarks@gmail.com> writes:
>> ... we run into the max tsvector length requirement "The length of a tsvector (lexemes + positions) must be less
than1 megabyte” 
>
>> Is there any way to disable or increase that limit in Postgres 10.3?
>
> No; it's forced by the representation used for tsvector, which stores
> lexeme offsets in 20-bit fields (cf WordEntry in
> src/include/tsearch/ts_type.h).  Perhaps that was short-sighted but
> I don't foresee it changing anytime soon.  You'd more or less need
> a whole new datatype ("bigtsvector"?) to make it happen.
>
>             regards, tom lane

Hi Jonathan,

Check out this potential fix/extension

https://github.com/postgrespro/tsvector2





--
Sent from: http://www.postgresql-archive.org/PostgreSQL-general-f1843780.html