In fact, I am not using a character sequence, but a custom COMPOSITE TYPE; part of which is a hstore:
CREATE TYPE SSTORE AS (scope TEXT, kvp hstore);
I created some functions and operators around it -- sort-of costly operations for comparing SSTOREs. Well: more costly than it would be to compare INTEGERs, I guess.
The sets of Key/Values held by SSTORE will show extensive similarities, that is: common pairs of key/value pairs.
So, essentially it looks like this:
CREATE TABLE textseqs(the_key SSTORE)
CREATE UNIQUE INDEX ON textseqs USING BTREE( hash_sstore(the_key), the_key )
Does this make sense? I am seeking some sensible trade-off between performance and flexibility...
Regards,
Rob
2011/5/14 David Johnston <polobo@xxxxxxxxx>
Could you index the reverse of the string so the unique part appears first?
On May 14, 2011, at 11:20, InterRob <rob.marjot@xxxxxxxxx> wrote:
> Dear list,
>
> I would be pleased if you could share some thoughts with me on the following: say I wish to maintain a table with all distinct character sequences (variable length) showing series with strong similarities. Example:
> "abbbabacccdef"
> "abbbabaccdcdf"
> "abbbabaccdcgf"
> ...
> "qwtrhdffdd"
> ...
> "qwtrhdffdds"
> ...
> "qwtrhdffddsspp"
> "qwtrhdffddsspf"
> "qwtrhdffddssph"
> "qwtrhdffddsspL"
> etc.
>
> Think of them as ordered values (array-like), as a set having many values in common, in the same "elements" (that is: positions; in my application, each position has some particular meaning -- the sequence represents a set of particular settings)
>
> CREATE TABLE textseqs(txtseq TEXT)
>
> What would be an efficient approach in enforcing a UNIQUE constraint?
>
> I was thinking of using hashbuckets in a b-tree:
> CREATE UNIQUE INDEX ON textseqs USING BTREE( hashtext(txtseq), txtseq )
>
> This index would "cache" hashes for each row. Upon inserting of a new row, traversing the index involves the comparison of two single integers for each node in the b-tree, until the actual hash value (if it exists) was reached. Then, only within that bucket (the hashes won't be unique), the more expensive string comparing is required; involving a sequencial comparison of (potentially maaaany) characters. Yet, within that bucket, the character series may be expected to show stronger differences than a plain sorted list of all values in the table would have, indexed by a b-tree index. Wouldn't traversing such a plain (non-composite, single column, on: "txtseq") b-tree index involve a sequencial comparison of (potentially many) characters *at EACH NODE* of the tree ? Or am I mistaken that each node is filled with actual values from the txtseq column?
>
> Thank you for your input!
>
> Cheers,
> Rob
