Two-stage prcomp cannot be aborted

Description

With h2o.prcomp(d, k=605, pca_method="Randomized") (d is a matrix of 1000 rows, 784 columns, of which 605 are non-const columns, i.e. it is the first 1000 rows of MNIST data.):

After being frustrated with it taking so long, and not knowing how long it might still run for (see bug #3322) I clicked stop from within RStudio. According to R it is stopped. However the water meter tells me 1 core is still maxed out, and `top` confirms that h2o is still running.

My guess is that it is because h2o.prcomp, with "Randomized" (or "Power" or "GLRM") first makes a PCA model, then makes an SVD model, and that the abort kills the PCA model, but it doesn't kill the SVD model?

I've marked this as a critical bug because the only way to stop that thread working was to kill the whole java process , this killing the cluster and losing all my models and data.

Assignee

New H2O Bugs

Fix versions

None

Reporter

Darren Cook

Support ticket URL

None

Labels

None

Affected Spark version

None

Customer Request Type

None

Task progress

None

CustomerVisible

No

Affects versions

Priority

Critical
Configure