investigate expected cost of cloud deployment, and as well as possible approaches for measuring cost

[jmartin-sul]

from sul-dlss/speech-to-text#3

see if back of the envelope estimates could rule out some of our possible implementations, or if they all look close enough in cost to keep pursuing all paths.

this involves two orthogonal sets of choices: how to deploy, and how to measure cost of deployed services.

deployment choices (not all mutually exclusive, but some are):

• ECS, possibly with some rules around scaling, unclear whether we can scale from 0 to 1 or more GPU EC2 instances in the ECS cluster. may need to have at least one EC2 instance always up, but it might be ok for it to be non-GPU? is it possible to have heterogeneous mix of VM types in one ECS cluster?
• always on EC2 VM
• thin vs heavy whisper docker container (i.e. is the container built such that it contains the models it needs, or does it download them on startup? whisper models total out to about 13 GB, says @edsu)
• SageMaker vs docker container running whisper (concerns about SageMaker configurability and training data policy captured in [investigate/prototype] speech_to_text_generation_service approach 2: Explore AWS SageMaker speech-to-text#4)

cost estimate approach:

• cost calculator
• deploy and test w/ tagged resources (@edsu seemed to be leaning in this direction, using Add initial Docker container speech-to-text#9 as a starting point for investigation)

[jmartin-sul]

@edsu feel free to claim this and put it in "in progress" if you're already working on it. anyone else, feel free to also assign yourself simultaneously and coordinate, if you're exploring paths besides tagged deployment of a docker container (or @edsu can say if he'd like to claim this whole issue for now, to prevent duplicated effort)

[jmartin-sul]

we might need a bit more of a handle on what ECS deployment and scaling look like, before we can really measure cost.

but we can also get a sense of how much it costs to caption a certain duration of media in ECS, and same in EC2. might be able to get a rough sense of cost by measuring that in the absence of ability to scale.

first few months might see the service on all the time anyway, since we probably have a substantial backlog to caption.

@andrewjbtw is interested in unit cost, and using that to determine how much we need to throttle e.g. bulk actions or preassembly batches for captioning. throttling may come naturally from our ECS or EC2 setup? e.g. set a limit on the number of STT containers, which are essentially single threaded (poll, work, poll, work, ...).