Workers allow a limit of 6 concurrent outgoing connections. If you run into that it can cause some h

Workers allow a limit of 6 concurrent outgoing connections. If you run into that it can cause some hard-to-debug error scenarios. That's the main concern for that setting, which was my focus when adding that note. If you don't want multiple concurrent queries,

max: 1

max: 1

max: 1

max: 1 is probably better as it's fewer moving parts to debug if you do run into issues.

AAJR Workers allow a limit of 6 concurrent outgoing connections. If you run into that...

AJROP•3/9/25, 4:15 PM

There's an open issue to set the default to something safer for Workers, but I'm not sure how that's going to play out: https://github.com/porsager/postgres/issues/1023

GitHub

Default connections to less than 10 in Cloudflare Workers environme...

With the following Cloudflare Worker, errors can be observed when deployed to production: import postgres from 'postgres'; export default { async fetch(request, env, ctx): Promise<Respon...

AAJR Workers allow a limit of 6 concurrent outgoing connections. If you run into that...

dave•3/10/25, 8:22 AM

Is there likely any noticeable difference between 1 and 5?

AAJR Hyperdrive is a transaction mode pooler, so those 25 go a lot further than you m...

dave•3/10/25, 8:25 AM

Interesting, why are you limited to just 60? (Not an issue for me yet in my use-cases.)

I thought Hyperdrive would be able to handle a near infinite number of connections from Workers.

Kkchro3 are there any recommendations on how to update hyperdrive if you want to rotate ...

dave•3/10/25, 8:25 AM

https://www.ai.moda/en/blog/automate-hyperdrive-iam-auth-with-workers

ai.moda

Automate Hyperdrive IAM Authentication with Workers

Learn how to automatically manage rotating credentials and AWS IAM tokens for PostgreSQL databases using Cloudflare Workers and Hyperdrive.

AAJR Plans is a strong word. It's something we've discussed internally, but no firm ...

dave•3/10/25, 8:33 AM

Monthly ping to let the team know I would be willing to pay for this. (Which is saying a lot, since I’m famously a freeloader when it comes to cloud services.)

AAJR Hyperdrive is a transaction mode pooler, so those 25 go a lot further than you m...

dave•3/10/25, 8:56 AM

Oh, I think I have run into this actually.. does this translate to about 60 concurrent Workers then? Or more?

Ddave Oh, I think I have run into this actually.. does this translate to about 60 conc...

AJROP•3/10/25, 11:05 AM

Substantially more. We see folks up past 1k QPS with this right now, glancing at logs at the moment.

Ddave Interesting, why are you limited to just 60? (Not an issue for me yet in my use-...

AJROP•3/10/25, 11:11 AM

Couple reasons.

Connections do use memory and CPU and those aren't free, even for us.

Given that it's not self service to change it, if we pick a value that's too high Hyperdrive will just knock over free tier offerings from most providers. When we do make this self service we'll allow a higher maximum value to be set.

If folks hit this limit (or expect to) and reach out to us, we can set it higher manually for them for now. Many customers have. It's a good forcing function for us not to put off the self-service work for too long.

At some point we'll probably allow an absurdly high value to be set with some sort of billing implication, such that we can recoup the cost of resources if people really want to blow the doors off something. We've got some internal teams using Hyperdrive for their products set as high as 1500.

Ddave Is there likely any noticeable difference between 1 and 5?

AJROP•3/10/25, 11:15 AM

This one is hard for me to answer, as someone who spends more time in C or Rust than in JavaScript.

Each of those really is a separate connection, so if you do actually want to have multiple queries going at the same time than I think probably there is a difference, yes.

But on Workers I think the general design philosophy would be to do more Workers, not more stuff within a Worker, so I'm not sure how often that would come up in practice.

LLaPh Where can I request Hyperdrive for MySQL early access?

AJROP•3/10/25, 1:03 PM

@thomasgauvin one for your list. Not sure what info you need.

Beta isn't running yet, but should be pretty soon.

sam•3/10/25, 5:22 PM

anyone know off the top of their head what the best way to timeout a hyperdrive query using postgres.js?

we setup the client with these opts { connect_timeout: 1, max: 5 }{ connect_timeout: 1, max: 5 } and also have some setTimeout(setTimeout( helpers wrapping the whole function issuing the query -- but still see some successful queries record latencies of multiple minutes sometimes

my understanding is that the wall clock in cf workers will only advance when there is external io performed, so setTimeout will often but not always enforce the timeout? (our metrics confirm this, but just making sure i understand why)

maybe just need to add a max_lifetime or idle_timeout to the connection options instead?

(pics are in US-mountain tz, with the cf dashboard showing avg query latency spiking around the same time as we were seeing internal metrics for max latencies higher than all the timeouts we have set in code)

Original message was deleted

AJROP•3/10/25, 6:47 PM

Sorry, not sure I'm following.

I would describe the use case of Hyperdrive as: "I, as a user, want to make a fullstack application on Workers, with my database hosted on Postgres somewhere. I will accomplish this by writing a Worker script that uses a standard Javascript Postgres driver/ORM to query my database".

Websockets can certainly play into some architectures depending on your use case, but I'd imagine most of our users do not use them.

TTyveris Regarding the hyperdrive limitations you mentioned, if my plan is to connect to ...

AJROP•3/10/25, 7:05 PM

It will be handled automatically. There are two different "types" of connections at play here. The Hyperdrive->Origin connections are limited, and automatically managed. You shouldn't need to do anything to manage those except not leave a transaction open forever (which breaks Postgres anyway so probably don't do that). Whenever a query completes that connection is put back into the pool for other queries to use. The phrase to search for to learn more would be "transaction-mode connection pooler"

The Worker->Hyperdrive connections can be left open, and are not limited. That's what you'd be leaving sitting open with the architecture you're describing. That should work fine. There's a 15-second timeout on those if they're idle, so if you're going to have low traffic you may want to have a "re-open my connection" control flow path, I guess?

AJROP•3/10/25, 7:07 PM

The only caveat I'd add is that a websocket connection does count as 1 of the 6 a Worker is allotted, so you'll be down 1 out of the gate and should plan your settings/query paralellism accordingly.

AAJR The only caveat I'd add is that a websocket connection does count as 1 of the 6 ...

Isaac McFadyen•3/10/25, 7:09 PM

Is this also the case for incoming websockets?

Isaac McFadyen•3/10/25, 7:10 PM

Figured that was the case for outgoing concurrent subrequest limit but didn't know it also applied to incoming

AJROP•3/10/25, 7:11 PM

Ahhhh. I might be wrong on that. I had always assumed it was both for tcp/websockets. Better to confirm in a Workers channel . Good call, sorry about that.

AAJR Ahhhh. I might be wrong on that. I had always assumed it was both for tcp/websoc...

Isaac McFadyen•3/10/25, 7:28 PM

I assumed it was too, but I'm not sure if it applies to incoming websockets as well as outgoing (which I think the user was referring to? Or at least I read it that way)

Ssam anyone know off the top of their head what the best way to timeout a hyperdrive ...

AJROP•3/11/25, 2:52 AM

Sorry, missed this earlier. If it were me, I'd probably solve this on the DB side with statement_timeoutstatement_timeout . Cut out all the interactions across the stack and just have queries get kiboshed if they run too long.

https://www.postgresql.org/docs/9.6/runtime-config-client.html

PostgreSQL Documentation

Client Connection Defaults

luiseok•3/11/25, 2:30 PM

Hi! I discovered that Hyperdrive has not been functioning properly since around 04:00 UTC on March 10.

It seems that the connection pool is insufficient, so I increased the pool size in Supabase. However, I am still encountering a "Connection terminated unexpectedly" error. No changes have been made to the code, and I have confirmed that connecting directly to the Supabase database using the connection string from a local environment works without any issues.

Would it be possible to get some assistance regarding this issue?

Brayden•3/11/25, 7:51 PM

Connected two different databases to Hyperdrive and when I try to query against them through a Worker I see it oscillate between the following two errors:

write CONNECTION_CLOSED 725fa92510fc468cd02c82f3759d365f.hyperdrive.local:5432

write CONNECTION_CLOSED 725fa92510fc468cd02c82f3759d365f.hyperdrive.local:5432

Internal error.

Internal error.

Any known issues happening at the moment or is this more indicative that it's something wrong on my end?

AJROP•3/11/25, 7:52 PM

We're releasing at the moment, but that shouldn't cause either of those. Send me the IDs and I'll take a look?

AJROP•3/11/25, 7:53 PM

Aside - those error messages should not be user-visible. This release includes some improvements to that, but if it's not better after today we'll do another pass on that. Apologies for the lack of visibility there.

kunal•3/11/25, 10:34 PM

I just created a new hyperdrive connection (replacing one regional replica rds instance with another, same database spec but moving regions slightly). I’m seeing much higher/variable latencies than the previous connection (in the latency tab in metrics). I’m wondering if something about my database is slow, or if it takes hyperdrive a while to figure out where to put its pools close to my database?

Kkunal I just created a new hyperdrive connection (replacing one regional replica rds i...

AJROP•3/11/25, 11:50 PM

if you DM me the IDs of new/old I can take a look in the morning

AAJR Sorry, missed this earlier. If it were me, I'd probably solve this on the DB si...

sam•3/12/25, 7:59 PM

huh it actually ended up not being our usage of hyperdrive, nuance in the timing apis, nor our pg statement_timeout

i forgot that we were also utilizing the cf cache api tucked into this codepath, not covered by our setTimeout(setTimeout( wrapper

turns out it was the cache.get(cache.get( call that was taking a butt load of time -- so just gotta make sure we wrap all those with timeouts

thanks for the help regardless!

AAJR Aside - those error messages should not be user-visible. This release includes ...

dave•3/13/25, 6:27 AM

I’ve definitely seen them before many times.

Ddave I’ve definitely seen them before many times.

AJROP•3/13/25, 11:05 AM

Okay. Well, if you see them after this week please toss me a heads up. That is not what we want to see.

Nick•3/13/25, 4:34 PM

I'm running into the error Error: write CONNECTION_CLOSED [some value].hyperdrive.local:5432Error: write CONNECTION_CLOSED [some value].hyperdrive.local:5432 on 100% of my requests to a Hyperdrive private database. I think it might be coming from a database error because I've just modified the schemas, but don't see anywhere to view a more specific error message. Is there somewhere where that will be logged?

Nick•3/13/25, 4:56 PM

I may have actually tracked this down to just being a poor network connection between my EC2 instance and the argotunnel endpoints. I ran cloudflared tunnel diagcloudflared tunnel diag and that resulted in errors

NNick I may have actually tracked this down to just being a poor network connection be...

dave•3/14/25, 10:58 AM

Which region (and you using IPv6 or IPv4)?Just curious so I don't run into it

Ddave Which region (and you using IPv6 or IPv4)?Just curious so I don't run into it 😛

Nick•3/14/25, 3:07 PM

I got the error message fixed, but the diag command still seems to fail. This VPC is IPv4-only, so maybe that's why it fails? The region is us-west-2b. The other thing worth noting is that it's a dev instance, so I've tried to keep the instance size as small as will function. Ultimately, it's working reliably now by correcting an SQL query that previously had a bad cast.

NNick I'm running into the error `Error: write CONNECTION_CLOSED [some value].hyperdri...

AJROP•3/14/25, 3:33 PM

Mind DM'ing me your Hyperdrive ID so I can dig into why there wasn't more info in that error message?

Chaika•3/14/25, 3:57 PM

Hyperdrive is so cool, the private tunnel connection setup process is very slick and guided, and just works.

I saw someone mention it before, so in case anyone else comes across this: Clickhouse's Postgres interface does kind of work with Hyperdrive, I wouldn't recommend it though.
You have to enable SSL/TLS in Clickhouse config, only use Simple Queries, not extended (so no parameters, and .simple() on sqlsql template or .unsafe() for postgres.js). It then works for most queries, but on repeated ones it'll often throw a connection error. The best way to get it to work reliably is to throw an error after your query (like by running ()()), and Clickhouse Postgres resets the entire connection on all errors, so you'll get a new conn on the next attempt, rip connection pooling though. It works fine for my limited use of small data querying though, and connecting securely through a tunnel.

Clickhouse's MySQL Interface seems to be much more fleshed out and workable, so when Hyperdrive supports that, will probably be much nicer...

vrballer•3/14/25, 11:05 PM

Hello our product currently uses Cloudflare Workers + Neon DB + Drizzle. It's been going great but recently we're hitting scaling issues over 1M requests per day. Specifically around not enough connections.

Using Hyperdrive appeared to be the right choice and the switch seemed relatively simple.

However we've run up against a wall with this error. Queries do seem to work but i guess not all of them? We tried using fetch_typesfetch_types but no dice

vrballer•3/14/25, 11:13 PM

^ Also not sure if it's related but we do use jsonbjsonb in some cases

Vvrballer ^ Also not sure if it's related but we do use `jsonb` in some cases

KingMaker•3/14/25, 11:18 PM

Hi, I’m in @VRBaller’s team

Adding onto this, we are doing nested jsonb queries:
for example:

value->>childValue

Vvrballer Hello our product currently uses Cloudflare Workers + Neon DB + Drizzle. It's be...

AJROP•3/15/25, 12:19 AM

We'll go ahead and bump your connection allotment in our next release. Please DM me your Hyperdrive and account IDs.

We'll be planning to cut a release starting early next week.

AAJR We'll go ahead and bump your connection allotment in our next release. Please DM...

AJROP•3/15/25, 12:22 AM

FYI also @KingMaker

Vvrballer Hello our product currently uses Cloudflare Workers + Neon DB + Drizzle. It's be...

AJROP•3/15/25, 12:33 AM

Also, to make sure I'm reading this right, are you having two separate issues? One is scaling, the other is a specific query?

AAJR Also, to make sure I'm reading this right, are you having two separate issues? O...

vrballer•3/15/25, 12:45 AM

We were having connection issues to our Neon Postgres. It seemed like HyperDrive was the right solution for large scale due to it's strong pooling mechanism. That might be working correctly however it seems like fetching data via Hyperdrive is behaving differently

vrballer•3/15/25, 12:46 AM

Essentially we're trying to go from:

export function createDbConnection(databaseUrl: string) {
  const sql = neon(databaseUrl);
  return drizzle(sql, { schema });
}

export function createDbConnection(databaseUrl: string) {
  const sql = neon(databaseUrl);
  return drizzle(sql, { schema });
}

to this to use Hyperdrive (it's possible this is wrong – we didn't make that many changes)

export function createDbConnection(hyperDrivedatabaseUrl: string) {
  client = postgres(hyperDrivedatabaseUrl,{
    max: 1,
    fetch_types: false,
  },);

  return drizzle(client, { schema });
}

export function createDbConnection(hyperDrivedatabaseUrl: string) {
  client = postgres(hyperDrivedatabaseUrl,{
    max: 1,
    fetch_types: false,
  },);

  return drizzle(client, { schema });
}

AJROP•3/15/25, 12:48 AM

That is odd. That error is a type error from elsewhere in your code, but ultimately it shouldn't happen if your script had been working before and that's all you changed.

AJROP•3/15/25, 12:49 AM

Honestly I'm not overly familiar with neon's driver or drizzle, so I'm afraid I may not be the best positioned here.

Is the error getting thrown here when you try to connect, or later on when you go to run queries?

AJROP•3/15/25, 12:50 AM

Followup question, are you able to run any queries successfully at all?

vrballer•3/15/25, 12:58 AM

queries are showing up in hyperdrive

vrballer•3/15/25, 12:59 AM

so it's connected correctly

Workers allow a limit of 6 concurrent outgoing connections. If you run into that it can cause some h

Similar Threads

Similar Threads

Similar Threads