Re: performance problems with bulk inserts/updates on tsrange with gist-based exclude constrains

On 16/09/2016 12:00, John R Pierce wrote:
> On 9/16/2016 3:46 AM, Chris Withers wrote:
>>>
>>> when you do updates, are you changing any of the indexed fields, or
>>> just "value" ?
>> Yeah, it's a temporal table, so "updates" involve modifying the period
>> column for a row to set its end ts, and then inserting a new row with
>> a start ts running on from that.
>
> thats expensive, as it has to reindex that row. and range indexes are
> more expensive than timestamp indexes
>
> modifiyng the primary key is kind of a violation of one of the basic
> rules of relational databases as it means the row can't be referenced by
> another table.

Right, but these rows have no natural primary key. Would it help if I
just added an auto-incrementing integer key? Would that make a positive
difference or would it just be a wasted column?

> I expect the expensive one is the constraint that ensures no periods
> overlap for the given key. I'm not sure how that can be done short of
> a full scan for each update/insert.

Indeed, I wonder if making the constraint deferrable might help for the
bulk case?

> it might actually perform better
> if you write the index with the key first as presumably the key is
> invariant ?

You mean:

PRIMARY KEY, btree (period, key) as opposed to
>
>
>
> --
> john r pierce, recycling bits in santa cruz
>