Re: Collations and Replication; Next Steps

Let me double check that assertion before we go too far with it.

Most of the problems I've seen are across 5 and 6 boundaries.  I thought I had case where it was within a minor release but I can't find it right now.  I'm going to dig.

That being said the sort order changes whether you statically or dynamically link (demonstrated on 4+ machines running different linux flavors), so at the point I have no reason to trust the stability of the sort across any build.  I legitimately question whether strcoll is buggy.  Ex. I have cases where for three strings a, b and c:  a > b, but  (a || c) < (b || c).  That's right postfixing doesn't hold.  It actually calls into question the index scan optimization that occurs when you do LIKE 'test%' even on a single machine, but I don't want to bite that off at the moment.

My mentality has switched to 'don't trust any change until shown otherwise', so that may have bled into my last email.

- Matt K.




On Sep 17, 2014, at 8:17 AM, Robert Haas <robertmhaas@gmail.com>
 wrote:

> On Wed, Sep 17, 2014 at 9:07 AM, Matthew Kelly <mkelly@tripadvisor.com> wrote:
>> Here is where I think the timezone and PostGIS cases are fundamentally different:
>> I can pretty easily make sure that all my servers run in the same timezone.  That's just good practice.  I'm also going to install the same version of PostGIS everywhere in a cluster.  I'll build PostGIS and its dependencies from the exact same source files, regardless of when I build the machine.
>> 
>> Timezone is a user level setting; PostGIS is a user level library used by a subset.
>> 
>> glibc is a system level library, and text is a core data type, however.  Changing versions to something that doesn't match the kernel can lead to system level instability, broken linkers, etc.  (I know because I tried).  Here are some subtle other problems that fall out:
>> 
>> * Upgrading glibc, the kernel, and linker through the package manager in order to get security updates can cause the corruption.
>> * A basebackup that is taken in production and placed on a backup server might not be valid on that server, or your desktop machine, or on the spare you keep to do PITR when someone screws up.
>> * Unless you keep _all_ of your clusters on the same OS, machines from your database spare pool probably won't be the right OS when you add them to the cluster because a member failed.
>> 
>> Keep in mind here, by OS I mean CentOS versions.  (we're running a mix of late 5.x and 6.x, because of our numerous issues with the 6.x kernel)
>> 
>> The problem with LC_IDENTIFICATION is that every machine I have seen reports revision "1.0", date "2000-06-24".  It doesn't seem like the versioning is being actively maintained.
>> 
>> I'm with Martjin here, lets go ICU, if only because it moves sorting to a user level library, instead of a system level.  Martjin do you have a link to the out of tree patch?  If not I'll find it.  I'd like to apply it to a branch and start playing with it.
> 
> What I find astonishing is that whoever maintains glibc (or the Red
> Hat packaging for it) thinks it's OK to change the collation order in
> a minor release.  I'd understand changing it between, say, RHEL 6 and
> RHEL 7.  But the idea that minor release, supposedly safe updates
> think they can whack this around without breaking applications really
> kind of blows my mind.
> 
> -- 
> Robert Haas
> EnterpriseDB: http://www.enterprisedb.com
> The Enterprise PostgreSQL Company