I think this is an issue with the colab code. The data for the trendpoints use dataframes which have filtered out projects for each different build which failed for some of the available toolchains.

For instance, if there is one toolchain that at build X had a failure for a test, that test gets filtered out to prevent having a non-fair comparison with the other toolchains for the same build X.

I'll switch to use the non-filtered data for the trendpoints, and those spikes should get smoothed out.