13.0.10 doesn't work on 10+ year old processors

[btm]

Jun 21 02:27:53 chef-server kernel: [548563.317770] traps: nginx[16830] trap invalid opcode ip:7f9957b85803 sp:7fff6f65f150 error:0 in libluajit-5.1.so.2.1.0[7f9957b7d000+9a000]

The version of openresty we're using now requires SSE4.2 support and my Xeon E5450 processors only support SSE4.1. 12.17.33 worked fine.

We can disable this with --with-luajit-xcflags='-mno-sse4.2', but @markan is currently investigating the repercussions of that.

See openresty/docker-openresty#39 for similar discussion.

Aha! Link: https://chef.aha.io/features/SH-3042

[markan]

Proposal: Document this in the release notes and hardware requirements page, and accept that older CPUs will not be supported.

The impact of disabling this is 'substantially slower string creation'; (this hash function is used both for the 'table' functionality and internally for code lookup)
Some accounts have a 2x change in execution time. However lua is a relatively small portion of our request time, so this probably isn't anywhere near that bad in practice.

The other cost of this is either we have to inflict this on every user of omnibus-software or maintain our own definition in chef-server.

Who are we protecting?
Nehalem (2009) and later processors support SSE4.2
Every AWS instance back to the c1/m1 series supports this.
Every Google Cloud instance I could find supports this.

So I'm leaning towards not supporting CPUs older than that.

[PrajaktaPurohit]

Update requirements to list minimum processor version.

[danielcbright]

I just had a customer run into this while upgrading their Chef 12.x installation - it's an older AMD Opteron setup but still very beefy (8 core, 32gb RAM). I think we should at least get this into the docs as a warning, or possibly make this an option in chef-server.rb. Just my 0.02

[markan]

Strongly agree about the documentation requirement.

Could you expand on what you mean by "possibly make this an option in chef-server.rb"?

[ch166]

The software should discover this and report the error as part of the setup process - rather than having random opcode crashes with no indication of what's going wrong.