TPI website basic example is fragile

[tasdomas]

I found this while testing out basic example in the README

If a user were to set spot to anything other than 0, the task will fail to start:

TPI [INFO] LOG 0 >> 2022-08-05 11:20:29 Started tpi-task.service.                           
TPI [INFO] LOG 0 >> 2022-08-05 11:20:29 (re)starting training loop from 1337 up to 1337 epochs
TPI [INFO] LOG 0 >> 2022-08-05 11:20:30 1337                                                
TPI [INFO] LOG 0 >> 2022-08-05 11:20:50 Error: invalid argument "0.05" for "--spot" flag: strconv.ParseBool: parsing "0.05": invalid syntax
TPI [INFO] LOG 0 >> 2022-08-05 11:20:50 tpi-task.service: Control process exited, code=exited, status=1/FAILURE
TPI [INFO] LOG 0 >> 2022-08-05 11:20:50 tpi-task.service: Failed with result 'exit-code'.   

An example main.tf that reproduces this (the only line changed is spot = 0.05):

terraform {
  required_providers { iterative = { source = "github.com/iterative/iterative" } }
}
provider "iterative" {}

resource "iterative_task" "example" {
  cloud      = "az" # or any of: gcp, az, k8s
  machine    = "m"   # medium. Or any of: l, xl, m+k80, xl+v100, ...
  spot       = 0.05     # auto-price. Default -1 to disable, or >0 for hourly USD limit
  disk_size  = -1    # GB. Default -1 for automatic
  permission_set = "/subscriptions/cee76754-ef49-49f7-b371-f6841fa82182/resourceGroups/domas-test-rg/providers/Microsoft.ManagedIdentity/userAssignedIdentities/domas-test-managed-identity"
  storage {
    workdir = "."       # default blank (don't upload)
    output  = "results" # default blank (don't download). Relative to workdir
  }
  script = <<-END
    #!/bin/bash

    # create output directory if needed
    mkdir -p results
    # read last result (in case of spot/preemptible instance recovery)
    if test -f results/epoch.txt; then EPOCH="$(cat results/epoch.txt)"; fi
    EPOCH=$${EPOCH:-1}  # start from 1 if last result not found

    echo "(re)starting training loop from $EPOCH up to 1337 epochs"
    for epoch in $(seq $EPOCH 1337); do
      sleep 1
      echo "$epoch" | tee results/epoch.txt
    done
  END
}

This is caused by multiple decisions at multiple levels:

• the tpi binary attempts to interpret a main.tf file for parameters
• the tpi binary expects spot to parse as a boolean

There are several ways to address this, including updating the basic terraform configuration example, but I think the way tpi parses the spot field needs to be changed.

[0x2b3bfa0]

We could set spot to Float64Var instead of BoolVar so it's consistent with the legacy Terraform schema.

https://github.com/iterative/terraform-provider-iterative/blob/142cd5b8b1aba9b0deb746b9fc5cfe3b6fecdd39/cmd/create/create.go#L53

[0x2b3bfa0]

However, asking users to pass --spot=0 in order to enable automatic spot pricing seems rather counterintuitive to me. I'd even ask myself whether allowing specifying a fixed spot price is a good idea. 🤔

[casperdcl]

whether allowing specifying a fixed spot price is a good idea

Do all clouds guarantee default(auto) spot pricing < on-demand?

[tasdomas]

I think the root cause of this is using the same binary for the CLI tool, for instance shutdown in provisioned machines and as a terraform provider plugin. I don't think there would be any harm in separating the three. The only blocking issue would be finding proper names for the three..

[0x2b3bfa0]

Do all clouds guarantee default(auto) spot pricing < on-demand?

✅ aws

No maximum price: Your Spot Instance will launch at the current Spot price. The price will never exceed the On-Demand price. (Recommended)

✅ az

The VM will not be evicted for pricing reasons. The max price will be the current price, up to the price for standard VMs. You will never be charged above the standard price.

❓ gcp

Not applicable

[0x2b3bfa0]

I think the root cause[^1] of this is using the same binary for the CLI tool, for instance shutdown in provisioned machines and as a terraform provider plugin.

Using the same binary as a Terraform provider and as a command line tool is definitely a bizarre choice. Still, it's not directly related to the float64 versus bool conundrum; API should look the same everywhere if it makes sense.

I don't think there would be any harm in separating the three.

The devil is in the details, but yes, task can be safely extracted as a module without changing a single byte of the code apart from the name.

The only blocking issue would be finding proper names for the three.

Not sure if it's “the only” issue, but it's probably the biggest one.

[^1]: Note that root causes are rarely the root causes you're looking for. 😅

[casperdcl]

so something like this?

• long term (major relase)
  • everywhere: on_demand: bool=True
• intermediate
  • task: spot: float=-1
  • standalone binary: on_demand: bool=True

[0x2b3bfa0]

Given that spot is a widely used term, I'd consider using it instead of more creative alternatives like !on_demand 🤔

[casperdcl]

I thought GCP uses s/spot/preemptible/ and s/on.demand/standard/ :)

[0x2b3bfa0]

Google Cloud followed the spot naming trend recently. 😅

• https://cloud.google.com/spot-vms
• https://cloud.google.com/compute/docs/instances/spot

[casperdcl]

Anyway the point of renaming is mostly to retain some backward-compatibility? The alternative -- spot: bool=False -- will silently break existing user workflows that have spot: float (presumably the float will get cast to a bool in the most undesirable way?)

[0x2b3bfa0]

the point of renaming is mostly to retain some backward-compatibility

Isn't it a bit too early to get started with “perfect backwards compatibility” on this project?

[casperdcl]

added to https://github.com/iterative/terraform-provider-iterative/issues/641 :)

[DavidGOrtega]

However, asking users to pass --spot=0 in order to enable automatic spot pricing seems rather counterintuitive to me.

Was your idea 😁 machine and runner have that term as spot and spot_price @0x2b3bfa0

[0x2b3bfa0]

Guilty as charged! 😅 Exposing two separate options didn't seem to me a good option either.

I wonder if we should just expose --spot=<boolean> and leave pricing to cloud providers. Who wants to set it manually anyway?[^1]

[^1]: Not a haphazard assumption anymore, let's check telemetry data.