chore(tests) include kong*.rockspec to cache hash calculation on travis

[bungle]

Summary

If it is not included in cache hash calculation, dependencies that are updated in .rockspec might be leaked to other tests / builds.

[dndx]

I would like to hear more opinions from others, for the following reasons:

1. Every time before the tests are ran, make dev is executed, which will correct all the dependency issue, if any.
2. Adding rockspec makes the cache granularity unnecessarily large. This means even if we are just adding new Kong lua files to the rockspec, a full rebuild is triggered, which is very slow and reduces our cache hit rate.

Basically, everything that can not be corrected by make dev are inside the .requirments file, which is part of the hash as f today. Everything else should be correctable by make dev much quicker than a full rebuild.

Obviously, being defensive is a good thing, for the CI runs the most defensive thing we could have done is to not use cache at all. We choose to not do that because it creates more harm without solving any known issues for us, the same principle should apply here as well.

[bungle]

Basically, everything that can not be corrected by make dev are inside the .requirements

We tried couple of things. But didn't find the solution that is better than this. Could we go with this solution as it makes cache correct, and then try to optimize it later? If you have better way to solve this, I am open to see it (make dev should not just go and purge developers luarocks, so it is a bit hard to implement correctly). Here is a sister PR to this that is for next and it is green on travis: #6383 (so this fix really works). The Github Actions fix can be found on kong-build-tools.

If cache gets correct, everything will suddenly turn green, so you never know do you have this: #6361 screwing up the cache. It depends that does cache have that or almost any other branch based on next.

I don't agree this creates more harm. It does similar thing to editing .requirements. So editing Makefile or .rockpec does use different cache (just like editing .requirements). But those are edited rarely.

[kikito]

We could only hash the rockspec dependencies, but not when adding a new file to Kong (or when adding a blank line to the rockspec).

resty -e "setmetatable(_G,nil); assert(loadfile('kong-2.1.3-0.rockspec'))(); table.sort(dependencies); for _,d in ipairs(dependencies) do print(d) end"

Output looks like:

inspect == 3.1.1
kong-lapis == 1.8.1.2
kong-plugin-acme ~> 0.2
...

Could we hash that 👆instead?

[bungle]

We could only hash the rockspec dependencies, but not when adding a new file to Kong (or when adding a blank line to the rockspec).

resty -e "setmetatable(_G,nil); assert(loadfile('kong-2.1.3-0.rockspec'))(); table.sort(dependencies); for _,d in ipairs(dependencies) do print(d) end"

Output looks like:

inspect == 3.1.1
kong-lapis == 1.8.1.2
kong-plugin-acme ~> 0.2
...

Could we hash that 👆instead?

Or just cat kong-2.1.4-0.rockspec | awk '/dependencies/,/}/ { print }'

[bungle]

Something like:

DEPS_HASH=$({ cat .ci/setup_env.sh .travis.yml .requirements Makefile; cat kong-*.rockspec | awk '/dependencies/,/}/'; } | md5sum | awk '{ print $1 }')

[bungle]

Something like:

DEPS_HASH=$({ cat .ci/setup_env.sh .travis.yml .requirements Makefile; cat kong-*.rockspec | awk '/dependencies/,/}/'; } | md5sum | awk '{ print $1 }')

I updated the PR with this change. Btw. we are also soon dropping travis... so... just to get things green until that.

[bungle]

I'll close this, let's just merge this to next:
#6383