CI improvements

[andyundso]

Background

I read through @aharpervc feedback on my PR for the Ruby 3.0 support (#524). He asked to implement three features for the CI noted here.

• Cache the compiled ports on Windows
• Publish an artifact of the built gem
• Publish test results to Circle CI

The issue I see with the current CI is that each Windows build step compiles the gem for itself. I assume this means that the compiled version works exactly for one Ruby version. So in order to test it on Windows, you would need to download the artifact using your Ruby version and architecture that you want to test. However, in the final fat gem, the needed binaries are provided for all the supported Ruby versions.

Second issue I see is the build time on Windows. Although the time would go down when using a cache, the gem was only build on Windows, but not tested.

So the main idea behind those pipeline changes are:

• Compile the ports and tiny_tds code on Linux for Windows as it's faster.
• Build the regular gem and the fat gem for Windows on Windows using the precompiled ports and code and store them in a workspace.
• Run the tests on Windows using the precompiled ports and code to have confidence that they work.

I originally had another step in mind that runs after all the tests passed and will build the actual gem, but this proofed to not work very well with cache, as it always recompiled the tiny_tds code, which is the one thing I didn't wanted to happen.

Preparation

There were a couple of changes necessary to enable this workflow. Some are the same as in #524.

• Lock Bundler down to v2.3.26 when installing it on Windows, as it's the last release to support Ruby 2.4 and 2.5.
• Remove a workaround for rake-compiler-dock on Windows. There are more details about this in 7b33fe3b. But TL;DR Docker without WSL2 on Windows isn't support anymore and WSL2 is Linux again, so fix is obsolete.
• Bump OpenSSL to v1.1.1s. I got an error when compiling the 32bit version of the ports for Windows. I went over the changelog and didn't see anything where I would assume that it breaks the gem. However, I'm not really knowledgable how this gem and openssl interact, so chances are that I overlooked something.

Windows setup

I would say the Windows setup is quite simple:

• Install the correct Ruby version as done before.
• Install an MSSQL Express 2017. The PowerShell script is adapted from the MSSQL Windows containers.
• Install toxiproxy-server
• Run the tests

There is one test named raises TinyTds exception with dead connection network failure that didn't work. I was able to reproduce this on my Windows machine, but didn't really see how I can fix that, so I skip it on Windows for now as it seems to work fine on Linux.

Performance

Previously, the Windows builds had about 30min to complete. I checked several runs with this pipeline to see how it performs now.

• The cross-compile takes about 6 minutes without cache, 90 seconds if the cache is present.
• Ruby installation clocks between 5 and 6 minutes.
• MSSQL installation takes about 5 minutes without cache, 3 minutes with cache.
• toxiproxy-server is another 90 seconds.
• tests takes about 90 seconds.
• Building the final gems takes another 90 seconds.

So totalling at between 14 and 22 minutes, depending if the cache is present.

... speaking of caches

I added caches for the following things:

• The precompiled ports.
• The precompiled code in tmp, rebuilt on every commit but then shared between jobs.
• The MSSQL installation file.

Potentially we can also cache:

• The built ports on Linux
• The downloaded Ruby versions for Windows.

Other changes

• I added builds against Ruby 2.4 just to see if it works. @aharpervc asked if it makes sense to drop support for older Ruby version. on one hand I think it would be good to keep it for now as #523 could introduce a few breaking changes. But on the other hand, the gemspec specifies that the gem supports Ruby 2.0 and newer and for the versions 2.0 to 2.3, tests don't run on CI and the code is not precompiled on Windows. So support for those versions is already somewhat not given.
• I noted that the Linux steps always ran against Ruby 2.7, so I fixed it.
• Publish test results to CircleCI.

Sorry for the very long PR description, I wanted to be explicit what I built and thought as there are quite some changes.

[andyundso]

ping @aharpervc :)

[andyundso]

glad to hear that the changes generally look good to you.

I updated the PR description now to better reflect what's actually in the PR, as some thing changed with your initial review.

[andyundso]

since it's been two weeks, I would like to ask if you could review the PR so can get it merged (@aharpervc @bvogelzang @larskanis)?

[andyundso]

sorry to annoy again, but Ruby 2.7 goes end of life in about a month and it would be nice if I could start working on the support for Ruby 3, but those changes need to get merged first (@aharpervc @larskanis @bvogelzang) (not sure why Github removed the review requests , i didn't do it intentionally ..)

[aharpervc]

@andyundso I've approved this PR & added you as a collaborator, so you should be able to merge when you're ready

[andyundso]

@andyundso I've approved this PR & added you as a collaborator, so you should be able to merge when you're ready

great, thanks for the trust! will probably merge on Sunday so I can get back to #524 afterwards.