Re: Proposal: Implement failover on libpq connect level.

On Thu, Sep 3, 2015 at 12:57 PM, Shulgin, Oleksandr
<oleksandr(dot)shulgin(at)zalando(dot)de> wrote:
> On Thu, Sep 3, 2015 at 6:02 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>> Maybe someday we should have all that, but I think for right now
>> that's complicating things unnecessarily. I think the best proposal
>> so far is to allow the host=X option to be repeated multiple times.
>> If you repeat the host=X option N times, you can also repeat the
>> port=X option exactly N times, or else you can specify it just once.
>> Done.
>
> But this already breaks backwards-compatibility with any clients who belief
> that whatever value specified the latest takes precedence. I'm not arguing
> that there are such use cases in the wild or that it's entirely sane thing
> to do, but still.

Yep. If we care about backward compatibility, there can be a new
option that must be specified to get the new behavior. We can also
decide not to care about this case.

> More importantly, this will break any code that tries to parse the conninfo
> string and produce a hashmap from it for modification.

That is true, but I am not sure I agree that it is important. Switch
to a hashmap whose values are arrays.

>> Alternatively, leave the host=X option alone and add a new option
>> hostlist=X, allowing a comma-separated list of names or IPs, with each
>> hostname or IP allowed an optional :port suffix. If host=X parameter
>> is omitted or the connection to that machine fails, try everybody in
>> the hostlist concurrently, or with some configurable (and presumably
>> short) delay between one and then next. Again, done.
>
> The exact behavior in case of both host/port and hostlist are specified
> becomes really tricky then. It's already tricky enough, if you recall the
> service files -- how are they going to come into play here?

It doesn't seem that tricky to me, but maybe I'm biased by having just
invented it 5 minutes ago.

> I believe the less there are implicit workings in the way libpq connects,
> the better.

I don't disagree with that as a general rule - only when it keeps us
from implementing useful features.

>>> Alternatively, change the rules for parsing the existing host=X
>> parameter so that we split it on some separator that isn't a valid
>> hostname character, and then strip off an optional :port syntax from
>> each entry; that value, if present, overrides port=X for that entry.
>
> It's tempting to use ':' as the separator here, but it's still valid for
> directory names and host can be one in case of UN*X sockets.

The directory name is only likely to contain : on Windows, and Windows
doesn't support UNIX sockets.

All of these objections seem pretty thin to me. I'd accept any of
them as a reason for preferring one alternative over another, but I
don't accept that the presence of a few problems of this magnitude
means we should give up on the feature. It's a good enough feature
that it is worth the possibility of slightly inconveniencing someone
running in an unusual configuration.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company