Free-tier servers

[sgbaird]

Workflow orchestration

Prefect

Prefect, specifically managed execution, with limit of 1 workspace, max of 1 concurrent flow run per workspace, and max of 1 managed execution pool (i.e., you can only have one Prefect instance running at a time across the whole free-tier account, though you can still integrate external runners/services). For both free-tier and pro, 2GB RAM, 24 hour max job runtime. Free-tier limited to 10 compute hrs/month, pro is 250 hrs/month. See below for snippets from docs:

Concurrency & work pools: Free tier accounts are limited to: Maximum of 1 concurrent flow run per workspace across all prefect:managed pools [and] Maximum of 1 managed execution work pool per workspace.

Pro tier and above accounts are limited to: Maximum of 10 concurrent flow runs per workspace across all prefect:managed pools [and] Maximum of 5 managed execution work pools per workspace.

Resources: Memory is limited to 2GB of RAM, which includes all operations such as dependency installation. Maximum job run time is 24 hours.

Usage limits: Free tier accounts are limited to ten compute hours per workspace per month. Pro tier and above accounts are limited to 250 hours per workspace per month. you can view your compute hours quota usage on the work pools page.

What about a deployment? Does this take up compute hours?

Prefect Python dependencies

Also, another big consideration is how it deals with Python dependencies. In the case of Prefect:

You can install Python package dependencies at runtime by configuring job_variables={"pip_packages": ["pandas", "prefect-aws"]} like this:

from prefect import flow

if __name__ == "__main__":
    flow.from_source(
    source="https://github.com/desertaxle/demo.git",
    entrypoint="flow.py:my_flow",
    ).deploy(
        name="test-managed-flow",
        work_pool_name="my-managed-pool",
        job_variables={"pip_packages": ["pandas", "prefect-aws"]}
    )

Alternatively, you can specify a requirements.txt file and reference it in your prefect.yaml pull_step.

General-purpose

• Oracle, for example several tutorials out there about hosting a Node-RED server on Oracle
  • Two AMD Compute VMs - would still likely work fine with Ax I think (though might potentially run out of memory if the models get too large). I think it's unlikely Ax would work well on an ARM processor, at least based on my running it from a Raspberry Pi 4B.
• AWS, though I believe you need to attach a credit card, and once you go past limits you start getting charged.
  • https://www.techtarget.com/searchcloudcomputing/tip/An-overview-of-Amazon-EC2-vs-AWS-Lambda
  • Non-free-tier
    • Likely somewhere between $25-$100/year, depending on the performance of the virtual machine
    • https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/burstable-performance-instances-unlimited-mode.html
    • https://aws.amazon.com/ec2/pricing/reserved-instances/pricing/
    • https://aws.amazon.com/ec2/pricing/on-demand/