Re: Crash report for some ICU-52 (debian8) COLLATE and work_mem values

Peter Eisentraut <peter(dot)eisentraut(at)2ndquadrant(dot)com> writes:
> On 8/7/17 11:17, Peter Eisentraut wrote:
>> I can confirm that that avoids the crash in ICU 52, so it's something to
>> consider. We would need more testing about which other ICU versions are
>> affected, however.

> In my testing, there are no crashes in ICU 51 and ICU 53, so we could
> restrict this workaround to version 52 specifically. Perhaps someone
> else can confirm my test results.

I did some desultory comparison of this source code in ICU 51.2, 52.1,
and 53.1. The crash-related code fragments are not significantly
different in 51.2 and 52.1. It may be that 52's problem is triggered by
broken collation data that wasn't broken in previous versions ... but
I have to say that I don't find 51 worthy of trust either. It seems
quite likely that pre-52 can still be made to crash, we just don't yet
have a test case that triggers it.

In 53, they have thrown away ucol.cpp lock stock and barrel, and rewritten
the whole thing from scratch in C++. So it's not surprising that these
bugs are gone. (Whether it's got other bugs is hard to tell. I do not
get the impression, just from eyeballing it, that the code quality went
up significantly :-(.) So that explains why upstream is not back-patching
fixes for these problems from newer versions: there are no newer versions
that they could possibly back-patch from.

I'm on board with Peter G's proposal: let's distrust ucol_strcollUTF8
in ICU < 53. If those folk felt a need to rewrite it completely after
only three release cycles, that should tell us something.

regards, tom lane