Daniel González Arribas

After today's 8.9.0 release, the outcome is pretty similar; however, the error triggers earlier at the `promote_to_concrete` stage, when building the ODEProblem.

On KDE neon (basically Ubuntu LTS), julia 1.7.3: ```$ uname -a Linux ----- 5.13.0-51-generic #58~20.04.1-Ubuntu SMP Tue Jun 14 11:29:12 UTC 2022 x86_64 x86_64 x86_64 GNU/Linux ``` ``` warning: didn't...

This did indeed pop up in the discussion we had with Oscar, I think we just did not act on this specific point (probably because I was not sure if...

Linearly equispaced data coordinates:    ```julia using BenchmarkTools, DataInterpolations # Equispaced begin exec_times = map(3:8) do x N = 10^x u = rand(N) tmax = exp(rand() * 10)...

It's clear that the log-distributed case is quite bad for the PR with large amounts of data (or, potentially, just a smaller cache; I'm on a Ryzen 7800 X3D, which...

Sounds good (I was similarly thinking on using a threshold based on testing with randomly-spaced abscissae, though not based on variance). I have some doubts regarding implementation: do you think...

Alright, will try to reopen this soon

Revisiting this topic, I added a simple mechanism along the lines of what we discussed. In particular: - At interpolant creation time, we check for "quasilinearity" by the simple and...

Yeah, there are indeed false positives with this criterion; I felt that having minimal overhead with a cheap criterion is worth some false positives (particularly when you can manually opt-out...

Alright, I made a new proposal, it's similar to the variance one but I take the different w.r.t. the straight line, which seems more directly relevant (you could build a...