I explained why the cache api doesn't work in condition 2

Well KV is just the cache api under the hood basically so its the same thing. As you currently have

I'm not sure where you get that assumption from? The cache api is colo-specific (meaning my writes in one colo don't take place elsewhere) whereas KV is global (beyond the normal 60s that kv caches)

What I mean is it would basically be the same if you just used the cache api instead because of the fact on read kv caches data for 60s. It does have a central store yes but your main central store is D1 which means cache api + D1 is basically the equal of KV.

My reasoning for using KV at all is it's a lot more long-lasting than the cache api will be

yah, but when it's not is really slow ... not much faster than D1

From my testing I'm getting around 1/3rd of the response time from KV cold to D1 - which is significant enough a difference

I agree with the global store of KV being important here. If you had a lot of requests and a cache miss and they all went to D1 it might be a bad time. But KV can handle it

But D1 is new and will get better there

Why not cache D1 with a DO?

Oh for sure! I've seen it improve massively since the early alpha

That's the entire conversation we're having 😅 Debating whether that gives enough of a perf change from KV

Where’s the nearest core to IAD? I would assume that the difference would probably be in the 10s of ms then?

looks like 2000 miles to PDX (pretty sure that's the KV one?) 😅

D1 is on DOs so it sounds more like your problem is that your D1 isn't in IAD? I've lost the context tbh

So it should be at least a little bit of a performance benefit, as long as the Worker-to-DO connection isn’t super slow

Basically, use DOs as cache for D1, since a vanilla DO will almost always respond faster

There's no controls for D1 location either atm

Or DOs really either

but you can create them from other DOs in specific locations

so create your D1 from something like iad.colo.do

If a single DO (a single thread) can handle the request load then it seems like it'd eliminate your cold KV read issue

This is what I'm thinking -- will have to play around with it

You could also hash among N DOs or whatever

But yeah Cache API -> DO shards -> D1 doesn't seem like a terrible plan to me 🤷‍♂️

What’s the backing for the DO DB?

Can't say

And subject to change

Aw darn…

Thanks for the thoughts all! Will do some playing around