The CI failures indicate that we use different code paths for actual computations with and without views (the differences are roudoff errors). This is surprising. What is the reason?

In general I am OK to close my PR and finalize this. Thank you!

This is surprising. What is the reason?

I agree. I expected them to take the same code path and therefore be exactly equal. I'll try figure out what is going on.

@andreasnoack - what is still needed in this PR to merge it?

There is some strange test failure on Julia 1.0 (I would assume that the failing == test should pass). @andreasnoack Do you know the reason?

That failure is indeed weird. I wonder whether it happens at random or whether it's something specific to Julia 1.0 on Mac.

Codecov Report

Merging #446 (fdd62fc) into master (be96833) will increase coverage by 0.11%.
The diff coverage is 95.00%.

@@            Coverage Diff             @@
##           master     #446      +/-   ##
==========================================
+ Coverage   84.01%   84.12%   +0.11%     
==========================================
  Files           7        7              
  Lines         807      819      +12     
==========================================
+ Hits          678      689      +11     
- Misses        129      130       +1     

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update be96833...fdd62fc. Read the comment docs.

I can reproduce this locally with Julia 1.0

julia> coef(glm(X, y, Normal(), IdentityLink()))
2-element Array{Float64,1}:
 0.9999999999999999
 1.0

julia> coef(glm(view(X, 1:9, :), view(y, 1:9), Normal(), IdentityLink()))
2-element Array{Float64,1}:
 1.0
 1.0

I don't know why the two versions are taking a different code path somewhere. If somebody has any idea then I'd like to hear it.

This also fails on Julia 1.7 so I'll try to figure out where the divergence is happening.

There is the same problem for lm as for glm (tested under Julia 1.0):

julia> coef(lm(X, y))
2-element Array{Float64,1}:
 1.0
 1.0000000000000004

julia> coef(lm(view(X, 1:9, :), view(y, 1:9)))
2-element Array{Float64,1}:
 0.9999999999999999
 1.0

julia> coef(glm(X, y, Normal(), IdentityLink()))
2-element Array{Float64,1}:
 0.9999999999999999
 0.9999999999999999

julia> coef(glm(view(X, 1:9, :), view(y, 1:9), Normal(), IdentityLink()))
2-element Array{Float64,1}:
 0.9999999999999999
 1.0

So this turned out to be pretty dumb. The view was 1:9 but he length was 10 so the two problems weren't identical. I just hadn't added any noise to the dependent variable so it was a perfect fit except for rounding differences. I've now added some noise and made the views the same lengths as the original data.

Good! Though it looks like my suggestions got reverted.

Though it looks like my suggestions got reverted.

Hm. Thanks for catching that. I tried to squash the intermediate commits while rebasing on master and I must somehow have ended up reverting those changes while doing that. I've reintroduced your suggestions.

I've now also included the tests from #466 and bumped the version.