Re: KVP table vs. hstore - hstore performance (Was: Postgres NoSQL emulation)

From: Jim Nasby <jim(at)nasby(dot)net>
To: Merlin Moncure <mmoncure(at)gmail(dot)com>
Cc: Stefan Keller <sfkeller(at)gmail(dot)com>, pgsql-performance(at)postgresql(dot)org
Subject: Re: KVP table vs. hstore - hstore performance (Was: Postgres NoSQL emulation)
Date: 2011-05-17 19:30:46
Views: Raw Message | Whole Thread | Download mbox | Resend email
Lists: pgsql-performance

On May 16, 2011, at 8:47 AM, Merlin Moncure wrote:
> On Sat, May 14, 2011 at 5:10 AM, Stefan Keller <sfkeller(at)gmail(dot)com> wrote:
>> Hi,
>> I am conducting a benchmark to compare KVP table vs. hstore and got
>> bad hstore performance results when the no. of records is greater than
>> about 500'000.
>> CREATE TABLE kvp ( id SERIAL PRIMARY KEY, key text NOT NULL, value text );
>> -- with index on key
>> CREATE TABLE myhstore ( id SERIAL PRIMARY KEY, obj hstore NOT NULL );
>> -- with GIST index on obj
>> Does anyone have experience with that?
> hstore is not really designed for large-ish sets like that.

And KVP is? ;)

IIRC hstore ends up just storing everything as text, with pointers to know where things start and end. There's no real indexing inside hstore, so basically the only thing it can do is scan the entire hstore.

That said, I would strongly reconsider using KVP for anything except the most trivial of data sets. It is *extremely* inefficient. Do you really have absolutely no idea what *any* of your keys will be? Even if you need to support a certain amount of non-deterministic stuff, I would put everything you possibly can into real fields and only use KVP or hstore for things that you really didn't anticipate.

Keep in mind that for every *value*, your overhead is 24 bytes for the heap header, 2+ varlena bytes in the heap, plus the length of the key. In the index you're looking at 6+ bytes of overhead, 1+ byte for varlena, plus the length of the key. The PK will cost you an additional 16-24 bytes, depending on alignment. So that's a *minimum* of ~50 bytes per value, and realistically the overhead will be closer to 65-70 bytes, *per value*. Unless your values are decent-sized strings, the overhead is going to be many times larger than the actual data!
Jim C. Nasby, Database Architect jim(at)nasby(dot)net
512.569.9461 (cell)

In response to


Browse pgsql-performance by date

  From Date Subject
Next Message Stefan Keller 2011-05-17 22:07:05 Re: KVP table vs. hstore - hstore performance (Was: Postgres NoSQL emulation)
Previous Message Jim Nasby 2011-05-17 19:19:02 Re: reducing random_page_cost from 4 to 2 to force index scan