Re: Deprecating Hash Indexes

On Sunday, October 14, 2012 03:45:49 PM Simon Riggs wrote:
> As discussed on other threads, Hash Indexes are currently a broken
> feature and bring us into disrepute.
>
> I describe them as broken because they are neither crash safe, nor
> could they properly be called unlogged indexes (or if so, why just
> hash?). And also because if somebody suggested implementing them the
> way they are now, they would be told they were insane because silent
> data corruption is not something we tolerate anymore. We know why they
> are like that, but its time to remove the rest of the historical
> research legacy. It's hard to say "We respect your data [except if you
> press here]" and be taken seriously.
>
> Suggested actions are
>
> * Put WARNINGs in the docs against the use of hash indexes, backpatch
> to 8.3. CREATE INDEX gives no warning currently, though Index Types
> does mention a caution.
>
> * Mention in the current docs that hash indexes are likely to be
> deprecated completely in future releases. Should anybody ever make
> them work, we can change that advice quickly but I don't think we're
> going to.
>
> Personally, I would like to see them removed into a contrib module to
> allow people to re-add them if they understand the risks. ISTM better
> to confiscate all foot-guns before they go off and then re-issue them
> to marksmen, at the risk of annoying the people that use them with
> full knowledge but that's likely a contentious issue.
>
> Alternate views?

I vote for at least logging a wal record when a hash index is modified which
uses incomplete actions to set 'indisready = false' in case its replayed. That
should only use a rather minor amount of code and should help users to find
problems faster.

Greetings,

Andres
--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services