Peter Duffy wrote:
> Loading the dump just consisted of "gunzip -c <file> | mysql
> ...
> This time, the mysql client kept choking, with complaints
> that the input contained binary zeroes.
> ...
> I tried gunzipping the file on the VM, and on a separate physical
> box with the same version of gunzip, then comparing the md5sums of
> the files: original - md5sum matches; unzipped file - md5sums didn't
> match.

Just to summarize then this really has nothing to do with mysql or
mariadb but can be reduced to a corruption problem either with data
going into gzip or being produced from gzip.

> At the moment, what seems to be inescapable is that gunzipping files in
> linux on virtualbox VMs can in some cases cause file corruption. Given
> that the identical procedure works on a physical box, it seems to
> exclude gzip itself, the shell (bash), the file system, and the kernel.
> The only thing which seems to be left is virtualbox itself.

The first thing I worried about would be if the gunzip has enough
memory to run and complete the action. And also if there is enough
disk space in the mysql backend storage. There is little error
detection in gzip and GIGO garbage-in-garbage-out applies to it. I
would also double check that the input files are identical between the
two cases too.

Since I would be worried about having enough memory I would check the
system log /var/log/syslog looking for OOM events. The Out Of Memory
Killer might have gotten involved. That would be bad.

> Has anyone else ever experienced anything else even remotely similar?
> I've not been able to find anything relevant on google.

Over the years have seen things as simple as a bad SATA cable produce
data corruption reading data off of disks. Have seen bad RAM dimms
corrupt blocks of memory, since ECC RAM is rare these days.

You are asking if something crazy is possible all I can respond with
is that yes it is possible for crazy things to happen and produce
bizarre behavior. That can make debugging difficult. But so far
there doesn't seem to be quite enough evidence to show that something
that deep is happening. Not yet anyway.

But it seems you have a reproducible test case. That's good. That
the gunzip < file | md5sum produces the wrong result is much simpler.

> Also - has anyone any preferences for virtualisation under linux:
> what seems to be the most reliable from the range of available tools
> to do it?

I thought Linux KVM has pretty much taken over the virtualization
world on Linux. Mostly using libvirt and virt-manager.

> Thoughts would be most welcome!

Good luck!
Bob