Thread: TOAST issue on custom index access method
Hi *,
I have a strange issue using a custom built index structure. My index access method support document type composed of words (as tsvector) and points (1-dimensional arrays of them). For internal reasons, I have to save the documents as a whole inside my structure (for proper reorganisations).
So, I form the tuples using index_form_tuple with the proper description. Everything works fine, as long as the documents are quite small. However, if the tsvector becomes too large, I run into a problem of not being able to store the documents, because (obviously) the tsvector is too large for one page.
Everything works fine, as long as I do insert ...(select...) statement, i.e. I have some table that feeds my table, where the index is set up, with data using, e.g.:
insert into documents(words, points) (select words, points from docs);
This query works in every case, because the words that I receive in my index are properly toasted in advance and the size of the documents is quite limited then. The document is constructed as CREATE TYPE document AS (words tsvector, points _point) .
However, I tried to insert a document containing a large words vector directly (the tsvector is of length 770, actually the points don't matter here too much).
I tried this using:
INSERT INTO documents (words, points) VALUES (to_tsvector('english', '<extremely long string constant.....>'), '<points array>');
If I perform the statement, mentioned above (I can also extract data of my feeding table 'docs' and use them as input), i.e. I insert the large vector directly, without pulling it out of some feeding table, I get a problem, when I call index_form_tuple, resulting in the following output:
ERROR: index row requires 12320 bytes, maximum size is 8191
So, creating my tuple using index_form_tuple fails because of the tsvector not being toasted in prior to inserting it to my index.
What I tried to solve this issue here, is to extract the words from the document (in my index) and calling 'Datum toast_compress_datum(Datum value)'in order to compress the tsvector into a proper toast table. Unfortunately, this function always returned NULL, which meant that the vector could not be compressed, resulting in the same error.
Actually, currently, I don't know, how to proceed (of course, there would be the possibilty of stopping the direct insertion, creating a table which triggers the insertion into the real data table, containing the index and inserting the row there after the insertion, which I don't assume to be a good solution).
Does anybody know about this issue of getting untoasted values if the values are inserted directly and toasted values, if copying the data from another table?
I hope that somebody could help me, here in order that my access method is able to run, properly.
Thanks in advance
Best regards
Carsten Kropf
P.S.: The version I use to program is 8.4.2
Carsten Kropf <ckropf2@fh-hof.de> writes:
> I have a strange issue using a custom built index structure. My index access method support document type composed of
words(as tsvector) and points (1-dimensional arrays of them). For internal reasons, I have to save the documents as a
wholeinside my structure (for proper reorganisations).
> So, I form the tuples using index_form_tuple with the proper description. Everything works fine, as long as the
documentsare quite small. However, if the tsvector becomes too large, I run into a problem of not being able to store
thedocuments, because (obviously) the tsvector is too large for one page.
Well, of course. I think this is a fundamentally bad index design. You
didn't say exactly what sort of searches you want this index type to
accelerate, but perhaps you need a design closer to GIN, in which you'd
make index entries for individual words not whole documents.
> What I tried to solve this issue here, is to extract the words from the document (in my index) and calling 'Datum
toast_compress_datum(Datumvalue)'in order to compress the tsvector into a proper toast table.
Indexes don't have toast tables.
regards, tom lane
Oh, thanks. I didn't know about index tables not having access to associated toast values. The index access method is (at leastin the textual part) similar to GIN, however, I needed to do some changes to it. Saving the whole document is actuallyonly important for vacuum. I think, I will find some workaround to solve this issue. However, it is a little bit strange, that I get toasted values (when inserting from another table) and untoasted values,if I insert items directly. Could anybody please explain this to me? Best regardsCarsten Kropf Am 24.06.2010 um 16:20 schrieb Tom Lane: > Carsten Kropf <ckropf2@fh-hof.de> writes: >> I have a strange issue using a custom built index structure. My index access method support document type composed ofwords (as tsvector) and points (1-dimensional arrays of them). For internal reasons, I have to save the documents as awhole inside my structure (for proper reorganisations). >> So, I form the tuples using index_form_tuple with the proper description. Everything works fine, as long as the documentsare quite small. However, if the tsvector becomes too large, I run into a problem of not being able to store thedocuments, because (obviously) the tsvector is too large for one page. > > Well, of course. I think this is a fundamentally bad index design. You > didn't say exactly what sort of searches you want this index type to > accelerate, but perhaps you need a design closer to GIN, in which you'd > make index entries for individual words not whole documents. > >> What I tried to solve this issue here, is to extract the words from the document (in my index) and calling 'Datum toast_compress_datum(Datumvalue)'in order to compress the tsvector into a proper toast table. > > Indexes don't have toast tables. > > regards, tom lane > > -- > Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) > To make changes to your subscription: > http://www.postgresql.org/mailpref/pgsql-hackers
Carsten Kropf <ckropf2@fh-hof.de> writes:
> However, it is a little bit strange, that I get toasted values (when inserting from another table) and untoasted
values,if I insert items directly. Could anybody please explain this to me?
Huh? An index will never ever get passed an externally-toasted value.
See the TOAST_INDEX_HACK code in indextuple.c.
regards, tom lane
Hmm, k, then I don't know exactly what happens. I have some values in my composite type. The size of the index tuple (if comingfrom one table and being copied to the target table) is about 15x bytes. Otherwise, if I insert the same tuple directly,it is about 12kb. So, I don't know exactly what happens here, unfortunately. Best regardsCarsten Kropf Am 24.06.2010 um 17:08 schrieb Tom Lane: > Carsten Kropf <ckropf2@fh-hof.de> writes: >> However, it is a little bit strange, that I get toasted values (when inserting from another table) and untoasted values,if I insert items directly. Could anybody please explain this to me? > > Huh? An index will never ever get passed an externally-toasted value. > See the TOAST_INDEX_HACK code in indextuple.c. > > regards, tom lane > > -- > Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) > To make changes to your subscription: > http://www.postgresql.org/mailpref/pgsql-hackers