Questions on Sprite placement and general feedback on Sprites

Hey all,

I’ve been using Sprites in anger the past few months. Big admirer of fly.io machines for years, and Sprites are similarly a great abstraction that I’m enjoying using. I think the design of the product is brilliant for the kinds of AI work I’m increasingly doing as a dev.

The biggest one is placement and regions. I can’t find any documentation about how sprites get placed, except a few mentions on some forum posts that it’s within the user’s region. However this doesn’t appear to be the case for me - my sprites are always in Los Angeles which isn’t ideal for latency reasons. Is there a a way I can set them to be closer to me? Or more guidance on how this is figured out?

On a more general note, I’ve found myself frustrated a few times by being unable to find specifics on how Sprites work in the docs. It seems to me a lot of the Sprite docs are written by an LLM. I am far from an AI hater, obviously I use Sprites for AI, but I would really prefer if the docs were written by hand. You can feel the difference when someone has put deliberate effort into thinking about how to communicate your product and how to use it, but the Sprite docs are full of LLMisms like “Size memory-hungry workloads against headroom you’ve measured, not a fixed number.” which are just visual noise I need to parse to find what I’m looking for.

And finally reliability: I’ve had lots of Sprite failures mostly from memory usage or from disks failing. For memory usage, I’d love some specifics on how memory grows and how to account for that. I have some really memory hungry workloads that just explode the sprite instantly before memory can grow. The other issue I hit a lot is disks failing. Often my sprites will die a few minutes in because their disk just disconnects. I’ve grown accustomed to just recreating them frequently because I expect them to die maybe ~50% of the time before they can complete the workload I want to use them for.

Hm… Los Angeles does seem to be a Sprite magnet lately, :magnet:…

What would your own preferred region be? (Your time zone looks Australia-ish, but the forum software doesn’t always guess that correctly.)

My guess is that the allocation algorithm incorporates a significant weighting for spare capacity. (That’s how I would design it, anyway.)

Also, APAC might simply not have any Sprite pools at present. Underlying costs are significantly greater there, whereas Fly.io is charging a flat rate for Sprites, at least for now.

Last I heard, they were actively considering adding such a feature, but it was not yet slated for inclusion:

Possibly people would have to opt in to more complexity when it comes to costs. For example, you might have to bid against an increasingly steep gradient as syd nears capacity.

Yeah, the continued reports of storage problems are a little concerning, just as an outside observer… (As a reminder, people should use the Questions / Help forum category when they’re mainly hoping that Fly.io staff will take part in the thread.) Are you using their new SBD I/O layer?

This post introduces its replacement, the Sprites Block Device (SBD). It’s the bag of spaghetti you grab off the shelf, and the only difference you’ll notice is better reliability.