Use hipercow_resources when running tasks #63
Conversation
| dat <- hipercow_driver_prepare(driver, root, environment()) | ||
|
|
||
| if (!is.null(resources$hold_until$computed)) { |
There was a problem hiding this comment.
I feel this could be done at validation, unless someone sits on their resources a long time.
There was a problem hiding this comment.
I am in two minds about this. I wonder if people might create a couple of resource objects for their "test_job" and their "real_job" and their "massive_job_that_always_waits_til_midnight", and then have those as kind of ready-made resources to launch their jobs in the same way as last time, without having to think about it.
I am not sure whether this is terribly likely, and might not be totally advisable - once in a while, available resources might change (but then they will need a new package version and will have to rerun their resource things anyway).
But also, I'm not sure doing the translation during submission causes any real problem does it?
There was a problem hiding this comment.
the only "problem" caused by doing it at submission is that there's two places to do the validation really. In particular, the validation here must never fail or it will be quite annoying for the user. We can see how this goes in practice if you want - I expect this flow will need some adjustment when it comes in contact with users, no matter what we pick.
We might sketch out a flow of validation etc as we complicate this too (environment variables and parallel configuration will all interact here too)
| @@ -6,6 +6,7 @@ windows_cluster_info <- function(config, path_root) { | |||
| max_cores = 32, | |||
| max_ram = 512, | |||
| queues = c("AllNodes", "Training"), | |||
| default_queue = "AllNodes", | |||
There was a problem hiding this comment.
or, have the first element of queues be default?
There was a problem hiding this comment.
Hmm - I think I maybe prefer saying explicitly default_queue than queues[1] but not strongly so...
| @@ -281,7 +294,7 @@ hipercow_resources_validate <- function(resources, | |||
| validate_cluster_requested_nodes( | |||
| resources$requested_nodes$computed, cluster_info$nodes) | |||
|
|
|||
| TRUE | |||
| resources | |||
There was a problem hiding this comment.
Hmm. By doing this (with the assignment of resources$queue$computed) we do open the door to the users sneaking a different queue through here if they are sufficiently motivated. Perhaps that's ok?
There was a problem hiding this comment.
So the only thing that gets changed here in the return is the queue gets default-ised if it was null - which would be good to do somewhere - I prefer that to having the user to explicitly talk about the queue.
I wonder if along with the discussion on when to turn "tonight" into a real time etc.... perhaps we really do have two stages of valdation here, the first one for all the syntactic stuff, and theoretical checks, and the second at last minute when we translate the timings, fill in the queue with a default if it was left NULL, and flag the sort of failures we'd only see when we consider the actual running of the job on an actual cluster...
For conversation later!
| dat <- hipercow_driver_prepare(driver, root, environment()) | ||
|
|
||
| if (!is.null(resources$hold_until$computed)) { |
There was a problem hiding this comment.
the only "problem" caused by doing it at submission is that there's two places to do the validation really. In particular, the validation here must never fail or it will be quite annoying for the user. We can see how this goes in practice if you want - I expect this flow will need some adjustment when it comes in contact with users, no matter what we pick.
We might sketch out a flow of validation etc as we complicate this too (environment variables and parallel configuration will all interact here too)
tests/testthat/test-interface.R
Outdated
| resources = res)), | ||
| "Submitted task '[[:xdigit:]]{32}' using 'elsewhere'") | ||
|
|
||
| # TODO HERE - Something is wrong with my mockery - mock_special_time |
There was a problem hiding this comment.
I'll talk you through this in person, it's an unfortunate consequence of how mockery works.
Usually some contortions are required here to make good tests
Co-authored-by: Rich FitzJohn <[email protected]>
A hipercow_resource can now be passed into all the task_create functions, and ultimately task_submit - OR - it can be left NULL and a working default will be picked up.
task_retry is a bit interesting - I haven't done anthing special.
Also - I have one mockery I couldn't get working in test-interface - I wanted to check that special_time got called once within task_submit - and I'm sure it does from manual browser()-ing - but I've done something wrong with the mock and it doesn't count the call...