I've made a bunch of comments - I think they're all usability related rather than anything to do with functionality.

I'm not set up for azure so I can't run this code so I won't click approve.

I have asked questions on a couple of them - will fix the rest later today. Thanks for the review !

I have fiddled with azure enough to have this start up a VM for me, using the below configuration. However it doesn't seem to successfully start up the htex workers - from dmesg on the started VM, it looks like maybe I'm running out of memory so I'll try a larger one.

Killing parsl with ctrl-C doesn't kill the VMs, which is worrying as that leaves a drain on real money behind at the end of the run.

config = Config(
    executors=[
        HighThroughputExecutor(
            label='azure_htex_single_node',
            address="fitzroy.cqx.ltd.uk",
            provider=AzureProvider(
                vm_reference=
                  {
                    'publisher': "Canonical",
                    'offer': "UbuntuServer",
                    'sku': "19.04",
                    'version': "latest",

                    'vm_size': "Standard_B1ls",

                    'disk_size_gb': 10,
                    "admin_username": "benc",
                    "password": "ligo123"
                },

                key_name='aws-us-east1',
                key_file='./azurecreds.pem',
                init_blocks=1,
                max_blocks=1,
            ),
        )
    ],
)

Using the below config I am able to run some simple test @python_apps:

config = Config(
    executors=[
        HighThroughputExecutor(
            label='azure_htex_single_node',
            address="fitzroy.cqx.ltd.uk",
            provider=AzureProvider(
                vm_reference=
                  {
                    'publisher': "Canonical",
                    'offer': "UbuntuServer",
                    'sku': "19.04",
                    'version': "latest",

                    'vm_size': "Standard_B4ms",

                    'disk_size_gb': 10,
                    "admin_username": "benc",
                    "password": "ligo123"
                },

                key_name='aws-us-east1',
                key_file='./azurecreds.pem',
                init_blocks=1,
                max_blocks=1,
            ),
        )
    ],
)

Killing parsl with ctrl-C doesn't kill the VMs, which is worrying as that leaves a drain on real money behind at the end of the run.

Killing parsl with ctrl-C should send cancel() requests and thus should close all of the VMs. In my testing, this worked as it should. Perhaps your VMs running out of memory caused this to not occur. Did the VMs go away with the updated, working config?

With regards to the rest of these changes, I will make them this evening.

I've done some more testing on the issue of not deleting on ctrl-C. If a submit() call is not still running, the VMs will indeed delete themselves. Unfortunately, Azure's API for running scripts on the VM (required for connecting back to Parsl controller) is both synchronous and very slow (30 sec min), so if you ctrl-C within 30 sec or so of a submit() request, Parsl doesn't know that an instance has been started, and so can't clean it up on shutdown. Any ideas on how to deal with this?

There's enough information stored at this point, before the script is run, to be able to kill the VM, even though parsl doesn't know about it.

submit, line 255-256

        vm_info = async_vm_creation.result()
        self.instances.append(vm_info.name)

Once you have that info, maybe you could put a try block around the whole of the rest of the code in submit that deals with killing the VM if an exception occurs. I think you might be able to see a KeyboardInterrupt exception there when ctrl-C is pressed (I'm not sure).

The idea being that until submit returns, it is the responsibility of that try block to tidy up; and then as soon as it ends and passes the ID back to the rest of parsl, it is the responsibility of the rest of parsl to call cancel.

There's enough information stored at this point, before the script is run, to be able to kill the VM, even though parsl doesn't know about it.

submit, line 255-256

        vm_info = async_vm_creation.result()
        self.instances.append(vm_info.name)

Once you have that info, maybe you could put a try block around the whole of the rest of the code in submit that deals with killing the VM if an exception occurs. I think you might be able to see a KeyboardInterrupt exception there when ctrl-C is pressed (I'm not sure).

The idea being that until submit returns, it is the responsibility of that try block to tidy up; and then as soon as it ends and passes the ID back to the rest of parsl, it is the responsibility of the rest of parsl to call cancel.

This makes sense to me. I will make those changes later today.

Done

ok this is looking good. I will give it another try out probably tomorrow and hopefully will merge it then