unsafe caching of .fmt files

[Karl Berry]

Hi Han-Wen - neither fmt generation nor any other runtime generation has
ever been parallel-safe. Even if I added locking to mktexfmt (I guess
you are making use of the dynamic fmt creation), something I am not at
all anxious to do, the failed simultaneous runs that would result
doesn't sound like a good outcome for you.

Instead, I think you need to generate the formats you need before
starting the parallelism, either with multiple mktexfmt foo.fmt calls or
fmtutil-user <something> (or fmtutil-sys, but if you want ~/.texlive2019
to be used instead of /usr/local/texlive/... or whatever, that sounds
like fmtutil-user), depending on your needs.

Indeed, in this situation, I also suggest that you explicitly disable
any mktex* creation.
MKTEXFMT=0; export MKTEXFMT;
MKTEXTFM=0; export MKTEXTFM;
MKTEXMF=0; export MKTEXMF;
MKTEXPK=0; export MKTEXPK

Alternative: you could set TEXMFVAR or TEXMFSYSVAR to a per-process
directory. But then every TeX run will need to create its own .fmt,
which sounds like an awful waste of time. --best, karl.