I.e. when trying to reproduce other bug I tried creating ddoc with just view and running it on about 700k docs and my node just died. It is perfectly reproducible, especially on i386 where address space is more limited.
beam.smp dies by abort in V8 and message looks like this:
- Fatal error in v8::ObjectTemplate::New()
- V8 is no longer usable
And by googling it appears that this may happen when running out of memory. We've also seen lots of threads in core dumps and our suspicion that mapreduce contexts are leaked was confirmed by grepping geocouch code.