"Holy Grail" Serverless

[beriberikix]

On Reddit you shared,

I think the holy grail would be a serverless platform that has persistent disk. You just push to a Heroku-like system and write to a local SQLite database that's automatically replicated that would be awesome...I've been skeptical that it's possible/practical but I had a good chat yesterday that gave me some ideas. I'm going to see if I can hack something together.

I agree that would be awesome! I envision you start from a serverless runtime (Lambda/OF/Kn/KEDA) and provide the app developer a SQL interface. The added benefit of this model is local development/testing becomes trivial even with litestream not available (I think?)

What are some of the challenges you foresee? I imagine #1 & #8 are part of the answer but are there others?

[benbjohnson]

The problem with SQL & serverless is that SQLite requires a single writer while serverless has no concept of that. It could work if you essentially treated writes as a network connection (like ODBC) to a central service but all reads could happen locally on the serverless function. You'd also have to replicate your database down to all instances of the serverless functions so they could get fast reads but that could be problematic with a large database.

IIRC, functions go cold after just a few minutes so you'd need to keep them warm to have any possibility of persisting the local DB copy.

[beriberikix]

Indeed, that's some of the thorniest parts of the problem. There's a good post about how Segment uses SQLite with shared volumes and tons of folks exploring "Stateful Serverless" and StatefulSets (if you're world is K8s.) And in that model, you don't need to keep functions warm, just the volumes (I think?)

[benbjohnson]

Thanks for the links. StatefulSets seems pretty straightforward to get working—especially since it numbers the pods sequentially. The first pod could be the writer node and the rest are replicas.

Someone else mentioned fly.io (@superfly) which is Heroku-like but has a persistent storage option. I played around with a bit and I think that would be pretty easy to get a distributed SQLite database working on. Their smallest node size plus 1GB storage is about $2.10/mo and they have 20 regions. I'm going to see if I can get a global SQLite database running for $42/month lol :)

[taeold]

Google Cloud Run recently released a feature to mount network drives!

https://cloud.google.com/run/docs/using-network-file-systems

Wondering if that capability can fit into the idea discussed here.

[terem42]

The biggest drawback of serverless architecture is the weight of context and costs of saving/restoring it
I.e. the whole purpose of database is to structurize the context, save it and quickly restore/filter on some standartised clauses like SQL, i.e. ease aforementioned operations.
I fail to see what's the profit in serverless architecture here ?

[benbjohnson]

@terem42 I think @gedw99 nails a lot of the points. The two biggest benefits of serverless that I see are low maintenance and low cost. Something like fly.io does a great job of being low maintenance and even does a pretty good job of being low cost. However, if you had a bunch of small, self-hosted, infrequently used applications then the cost can add up compared to a true serverless architecture. After their free tier, a single node would be $2/month—not huge, but if you run 10 apps then it can start adding up.