On Fri, May 03, 2019 at 09:42:17PM +0200, Tomas Vondra wrote:
>On Fri, May 03, 2019 at 12:21:36PM -0400, Tom Lane wrote:
>>Alvaro Herrera <alvherre@2ndquadrant.com> writes:
>>>Error reporting in extended statistics is inconsistent -- many messages
>>>that are ereport() in mvdistinct.c are elog() in the other modules.
>>>...
>>>I think this should be cleaned up, while at the same time not giving too
>>>much hassle for translators; for example, this message
>>>dependencies.c: elog(ERROR, "invalid MVDependencies size %zd (expected at least %zd)",
>>>should not only be turned into an ereport(), but also the MVDependencies
>>>part turned into a %s. (Alternatively, we could decide I was wrong and
>>>turn them all back into elogs, but I obviously vote against that.)
>>
>>FWIW, I'd vote the other way: that seems like a clear "internal error",
>>so making translators deal with it is just make-work. It should be an
>>elog. If there's a reasonably plausible way for a user to trigger an
>>error condition, then yes ereport, but if we're reporting situations
>>that couldn't happen without a server bug then elog seems fine.
>>
>
>Yeah, I agree. Most of (peshaps all) those errors are internal errors,
>and thus should be elog. I'll take care of clening this up a bit.
>
OK, so here is a patch, using elog() for all places except for the
input function, where we simply report we don't accept those values.
I agree those are internal errors, usually meaning the statistics object
was either corrupted or there's a bug in how it's built/serialized.
Users should not be able to trigger those cases (the only thing I can
think of is sending a bogus value through send/recv functions, that
simply do byteaout/byteain).
Now, what about backpatch? It's a small tweak, but it makes the life a
bit easier for translators ...
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
