Hey all! Today we released the final step in a fairly large rearchitecture of Hyperdrive. We'll be

Hey all! Today we released the final step in a fairly large rearchitecture of Hyperdrive. We'll be getting into the details in a full length blog post a little down the road, but for now here are the takeaways.

All newly created Hyperdrives will perform substantially more efficiently. This goes double for users with the origin database in APAC.
This should primarily manifest as lower latencies and fewer disconnects.
This will affect all Hyperdrives with or without caching.
We will be backfilling this solution to existing Hyperdrives progressively, over the course of the next couple weeks, starting next Monday. If you want yours handled sooner, please feel free to message me or respond in a thread here.

This one took the whole team quite a bit of work, it's been fun to ship. I'm excited to finally have it going out the door.

dan•2/4/25, 8:34 PM

I've done some search on this channel and it seems like cache busting/purging is still WIP? I ran into a bug today that took me awhile to figure out it was being caused by hyperdrive - user creates a comment, the query client then invalidates the comments query, it refetches the stale data causing my optimistic comment to disappear. This makes sense given hyperdrive caching. I'm just curious as to what my workaround should be given that cache purging isn't a thing (yet).

dan•2/4/25, 8:35 PM

It'd be a little contrived for me to add some volatile function to the query given that it'd have to originate from the client and then drill through the api router and modify queries in a weird way

Ddan I've done some search on this channel and it seems like cache busting/purging is...

AJROP•2/4/25, 8:44 PM

The alternative is likely just a second hyperdrive config with caching disabled, and direct your queries to whichever is appropriate at the application level.

Not a great alternative, we're aware.

Ddan I've done some search on this channel and it seems like cache busting/purging is...

dave•2/5/25, 3:27 PM

iirc ORDER BY random()ORDER BY random() will always cause a cache miss, though @AJR can confirm if that's correct

dan•2/5/25, 3:28 PM

it'd also cause strange query results

AAJR Hey all! Today we released the final step in a fairly large rearchitecture of H...

dave•2/5/25, 3:28 PM

Still no defined list of IP addresses that Hyperdrive connects from, right?

Ddan it'd also cause strange query results 😛

dave•2/5/25, 3:29 PM

Totally depends on what you're doing. If you are sorting by a date column, but are returning the date column anyway, you can sort it in JavaScript after.

dan•2/5/25, 3:29 PM

Not if you're paginating

dan•2/5/25, 3:30 PM

I just turned off caching for now, I'll wait until cache purging is shipped, I don't love modifying the application layer in weird ways to get expected query behavior

Ddan Not if you're paginating

dave•2/5/25, 3:32 PM

hmm, there's probably another trick you could use with random(). Like WHERE random() >= 0;WHERE random() >= 0;.

dan•2/5/25, 3:33 PM

I think at that point it'd make more sense to just use a scond hyderdrive config with caching disabled for queries that you're doing that to

Ddan I just turned off caching for now, I'll wait until cache purging is shipped, I d...

dave•2/5/25, 3:33 PM

Is there a reason why you aren't returning the comment in the same query where the user creates it?

Ddave Is there a reason why you aren't returning the comment in the same query where t...

dan•2/5/25, 3:33 PM

I am, but think about the user that gets a push notification and looks at their phone immediately, that comment won't be loaded in

Ddan I am, but think about the user that gets a push notification and looks at their ...

dave•2/5/25, 3:34 PM

delay the notification by the cache duration?

dan•2/5/25, 3:35 PM

I think the two hyperdrive configs is a pretty good solution, since in my case it's really there's a set of resources that frequently are mutated and should never be cached, and sets of resources that can be cached

dan•2/5/25, 3:35 PM

Cache purging would just allow optimization of this

Ddave Still no defined list of IP addresses that Hyperdrive connects from, right?

AJROP•2/5/25, 3:35 PM

Nope. Soon(tm), though.

Part of the reason that this is taking so long is that a bunch of folks went ahead and tried to allowlist based on the shared/untrusted ip pool. We don't want to just break them (though at some point they'll get broken anyway when the pool shifts enough to clip them, but at least it won't be because we did it). So the migration over is kind of a process.

dave•2/5/25, 3:36 PM

@AJR oh, another reason why exposing Hyperdrive as a regular PostgreSQL pooler would be great: we could do stacked Hyperdrive configs! e.g. one "main" Hyperdrive with no caching connected to the real backend directly, and then have another Hyperdrive config with caching enabled that connects to the main Hyperdrive's pooler.

dan•2/5/25, 3:37 PM

I think it's still worth using hyperdrive without caching though right?

Ddan I think it's still worth using hyperdrive without caching though right?

dave•2/5/25, 3:37 PM

yes.

Ddan I think it's still worth using hyperdrive without caching though right?

AJROP•2/5/25, 3:38 PM

Oh yeah. Measure how long it takes to create a connection to a pooler/origin further away than a couple hundred miles and you'll have a clear picture of why only ~half of our configs even having caching enabled.

AJROP•2/5/25, 3:38 PM

I really need to write that "how hyperdrive works" blog post soon. I'm finally allowed to.

Ddave @AJR oh, another reason why exposing Hyperdrive as a regular PostgreSQL pooler w...

AJROP•2/5/25, 3:39 PM

I'm hopeful we can get per-query cache bypass/adjustment done sooner than that would take.

AAJR Nope. Soon(tm), though. Part of the reason that this is taking so long is th...

dave•2/5/25, 3:40 PM

I'm pretty sure the answer to this is "heck no", but would it possible to have a unique outgoing port for each Hyperdrive config? (And I don't think it would be too insane, since it's only 25 ports per paid Workers account that would need to be "reserved".) Or even 1 unique port per account would be nice.

That would be really nice to firewall based off of.

Ddave I'm pretty sure the answer to this is "heck no", but would it possible to have a...

AJROP•2/5/25, 3:40 PM

Hmmm. I'll ask the team, it's not something we've talked about afaik

AAJR I really need to write that "how hyperdrive works" blog post soon. I'm finally ...

dave•2/5/25, 3:42 PM

I would be curious if using Hyperdrive reduces bandwidth pricing (and by how much) for Google Cloud customers, since Cloudflare has a reduced bandwidth price iirc.

Ddave I would be curious if using Hyperdrive reduces bandwidth pricing (and by how muc...

AJROP•2/5/25, 3:43 PM

The goal is that it should, yes. Still needs some work to get all the way there, I believe. I don't work with that project a ton and there is some nuance IIRC.

AAJR The goal is that it should, yes. Still needs some work to get all the way there...

dave•2/5/25, 3:44 PM

iirc it's already reduced pricing, no?

dave•2/5/25, 3:45 PM

https://cloud.google.com/network-connectivity/docs/cdn-interconnect#pricing

AJROP•2/5/25, 3:46 PM

It gets complicated. It goes GCP->Workers->Hyperdrive for ingress, so I believe that's already reduced. Egress from ??->Workers->Hyperdrive->GCS is the one I'm not certain about.

dave•2/5/25, 3:46 PM

Workers->Hyperdrive->GCS would be ingress from Google's standpoint (which is free)

AAJR Nope. Soon(tm), though. Part of the reason that this is taking so long is th...

dave•2/5/25, 3:55 PM

You should be able to automatically tell who did the allowlist by trying to connect from a new IP. Then whoever fails, email them giving them a warning that they need to use the new proper official list of IP ranges.

AJROP•2/5/25, 3:57 PM

We landed with just never migrating them, plus courtesy email. Same difference, though.

AAJR We landed with just never migrating them, plus courtesy email. Same difference,...

dave•2/5/25, 4:03 PM

oh, is there a list already somewhere secret?

AJROP•2/5/25, 4:04 PM

Of the shared address pool? Not publicly, but folks are clever. Hazard of making tools for other devs, I guess. In some regions things are stable enough that folks noticed that all Hyperdrive connections were coming in from a small number of addresses, and eventually just coded that into their firewall rules.

AJROP•2/5/25, 4:05 PM

(that's my theory, based on the available evidence. maybe someone dug up a list I don't know about)

oritwoen•2/6/25, 12:07 PM

Is there any eta for mysql support?

Ooritwoen Is there any eta for mysql support?

AJROP•2/6/25, 12:14 PM

We generally don't put out dates in advance of releases. That tends to make nobody happy.

It's being actively worked on, though.

AAJR We generally don't put out dates in advance of releases. That tends to make nob...

oritwoen•2/6/25, 12:15 PM

Okay, good for me, thanks

Sam•2/8/25, 8:45 PM

Can you connect hyperdrive to a digital ocean managed postgres db or is the IP range still not allowed?

SSam Can you connect hyperdrive to a digital ocean managed postgres db or is the IP r...

AJROP•2/9/25, 1:33 PM

Sorry, not sure I understand the question.

Hyperdrive doesn't refuse to egress to DO ranges. Is there an issue with ingressing to DO from the shared pool?

AAJR Sorry, not sure I understand the question. Hyperdrive doesn't refuse to egress ...

Sam•2/9/25, 1:37 PM

Hey thanks for reply on this. Yes when I try to setup hyperdrive + a managed DO database in cf dashboard it errors

I have setup connections with supabase in the past successfully and I searched this channel for issues with DO leading me to believe it’s with DO ip range?

AJROP•2/9/25, 1:52 PM

Interesting. I searched and you're right, TCP connect from workers to DO seems to have some issues.

I'm not sure if it's due to the shared address range or not, though.

I'll check Monday with my teammate who's working on the migration to the CF IP ranges and see if we have a way to test if DO cooperates with those. I'd be curious to know, honestly

AAJR Interesting. I searched and you're right, TCP connect from workers to DO seems t...

Sam•2/9/25, 2:02 PM

I have digital ocean and azure credits to burn and the latter seem to be priced crazily so would be ideal!

Maybe if there is no solution with public there is one with private connection?

I’d be open to using d1 instead also but I have about 50gb of data already so I don’t know if limits can be increased to say 100gb (for future scale)

Also happy to discuss using r2 as this is mostly a primary store of library like data with reads being carried out from my typesense instance.

I am on CF startups program also

AJROP•2/9/25, 2:16 PM

FYI @thomasgauvin ^

Might be in a better spot to address what will fit your use case best.

Stilau•2/9/25, 2:23 PM

Hey,
I am not sure what Hyperdrive is, but does this make it possible to have a database being hidden behind the Cloudflare Network?

SStilau Hey, I am not sure what Hyperdrive is, but does this make it possible to have a ...

AJROP•2/9/25, 2:43 PM

Yes. Our tunnel integration does exactly that.

AAJR Yes. Our tunnel integration does exactly that.

Stilau•2/9/25, 2:45 PM

I want to hide my database behind Cloudflare's Network without needing to have the CloudflaredCloudflared application on all devices..

Is that what Hyperdrive fixes?

AJROP•2/9/25, 2:52 PM

Hyperdrive is a way to talk from Workers to a centralized database. We integrate with tunnels too so you'd only need have cloudflared running once somewhere within your vpc to egress the traffic to your DB from.

Based on your questions I'm assuming you're not running your application on Workers?

AAJR Hyperdrive is a way to talk from Workers to a centralized database. We integrat...

Stilau•2/9/25, 2:56 PM

No I am not..

Hey all! Today we released the final step in a fairly large rearchitecture of Hyperdrive. We'll be

Similar Threads

Similar Threads

Similar Threads