is it an expensive operation? and why should that be the developer's concern and not the binding/pla

Aarbfay is it an expensive operation? and why should that be the developer's concern and...

L

lambrospetrou•1/23/25, 12:34 PM

In reality most real-world applications do not expose directly the raw DO IDs as their own application IDs, so you anyway will have to do some conversion/validation/parsing before passing it to the

get(...)

get(...)

get(...)

get(...) method. So, it's 1 extra line among tens/hundreds.

I do usually end up creating a helper function though that wraps my applications logic and achieves what you suggest above. 1 line to create DO stubs, but based on my apps parsing logic.

K

KWAC•1/23/25, 1:53 PM

why it closes connection when i send message to server?

// Global list to store all active WebSocket connections
const clients = new Set();

addEventListener('fetch', event => {
  event.respondWith(handleRequest(event.request));
});

async function handleRequest(request) {
  const upgradeHeader = request.headers.get('Upgrade');
  if (!upgradeHeader || upgradeHeader !== 'websocket') {
    return new Response('Expected Upgrade: websocket', { status: 426 });
  }

  const webSocketPair = new WebSocketPair();
  const [client, server] = Object.values(webSocketPair);

  // Add the new connection to the global list
  clients.add(server);
  server.accept();

  // Handle incoming messages
  server.addEventListener('message', event => {
    console.log('Message received from client:', event.data);

    // Broadcast the message to all connected clients
    clients.forEach(client => {
      if (client !== server && client.readyState === WebSocket.OPEN) {
        client.send(`Broadcast: ${event.data}`);
      }
    });
  });

  // Handle connection close
  server.addEventListener('close', () => {
    console.log('Client disconnected');
    // Remove the connection from the global list
    clients.delete(server);
  });

  return new Response(null, {
    status: 101,
    webSocket: client,
  });
}

// Global list to store all active WebSocket connections
const clients = new Set();

addEventListener('fetch', event => {
  event.respondWith(handleRequest(event.request));
});

async function handleRequest(request) {
  const upgradeHeader = request.headers.get('Upgrade');
  if (!upgradeHeader || upgradeHeader !== 'websocket') {
    return new Response('Expected Upgrade: websocket', { status: 426 });
  }

  const webSocketPair = new WebSocketPair();
  const [client, server] = Object.values(webSocketPair);

  // Add the new connection to the global list
  clients.add(server);
  server.accept();

  // Handle incoming messages
  server.addEventListener('message', event => {
    console.log('Message received from client:', event.data);

    // Broadcast the message to all connected clients
    clients.forEach(client => {
      if (client !== server && client.readyState === WebSocket.OPEN) {
        client.send(`Broadcast: ${event.data}`);
      }
    });
  });

  // Handle connection close
  server.addEventListener('close', () => {
    console.log('Client disconnected');
    // Remove the connection from the global list
    clients.delete(server);
  });

  return new Response(null, {
    status: 101,
    webSocket: client,
  });
}

// Global list to store all active WebSocket connections
const clients = new Set();

addEventListener('fetch', event => {
  event.respondWith(handleRequest(event.request));
});

async function handleRequest(request) {
  const upgradeHeader = request.headers.get('Upgrade');
  if (!upgradeHeader || upgradeHeader !== 'websocket') {
    return new Response('Expected Upgrade: websocket', { status: 426 });
  }

  const webSocketPair = new WebSocketPair();
  const [client, server] = Object.values(webSocketPair);

  // Add the new connection to the global list
  clients.add(server);
  server.accept();

  // Handle incoming messages
  server.addEventListener('message', event => {
    console.log('Message received from client:', event.data);

    // Broadcast the message to all connected clients
    clients.forEach(client => {
      if (client !== server && client.readyState === WebSocket.OPEN) {
        client.send(`Broadcast: ${event.data}`);
      }
    });
  });

  // Handle connection close
  server.addEventListener('close', () => {
    console.log('Client disconnected');
    // Remove the connection from the global list
    clients.delete(server);
  });

  return new Response(null, {
    status: 101,
    webSocket: client,
  });
}

// Global list to store all active WebSocket connections
const clients = new Set();

addEventListener('fetch', event => {
  event.respondWith(handleRequest(event.request));
});

async function handleRequest(request) {
  const upgradeHeader = request.headers.get('Upgrade');
  if (!upgradeHeader || upgradeHeader !== 'websocket') {
    return new Response('Expected Upgrade: websocket', { status: 426 });
  }

  const webSocketPair = new WebSocketPair();
  const [client, server] = Object.values(webSocketPair);

  // Add the new connection to the global list
  clients.add(server);
  server.accept();

  // Handle incoming messages
  server.addEventListener('message', event => {
    console.log('Message received from client:', event.data);

    // Broadcast the message to all connected clients
    clients.forEach(client => {
      if (client !== server && client.readyState === WebSocket.OPEN) {
        client.send(`Broadcast: ${event.data}`);
      }
    });
  });

  // Handle connection close
  server.addEventListener('close', () => {
    console.log('Client disconnected');
    // Remove the connection from the global list
    clients.delete(server);
  });

  return new Response(null, {
    status: 101,
    webSocket: client,
  });
}

Aalgads Pleasantly surprised this evening. I wanted to see if the outbox pattern could ...

J

James•1/23/25, 3:02 PM

Id work to ensure heavy compute takes place away from the DO due to cost. You pay for wall time on a DO but not a worker, so api calls are much cheaper outside a DO. You could adjust your logic to still use triggers on the DO but offload to a worker once the alarm kicks off.

Aalgads Here is sample code that should give a good idea of what I am trying to do: http...

H

HowlOftheSun•1/23/25, 3:32 PM

Thanks!

J

James•1/23/25, 4:08 PM

@algads , caveat is that you lose the synchronicity of a single threaded DO once you offload to a worker, events may push to your API out of order

Aalgads @Ahz thinking about this some more though, since the DO is single threaded and t...

J

James•1/23/25, 5:00 PM

not if downstream workers are delayed or blocked, req1 to API hangs, req2 5 seconds later finishes first.

J

James•1/23/25, 5:14 PM

it could fire again, but the DO will be busy on its current thread, but if you await anywhere in the DO, then you're open to losing synch

and if you fire the alarm and offload to a worker, well thats all async and anything goes

JJames it could fire again, but the DO will be busy on its current thread, but if you a...

L

lambrospetrou•1/23/25, 5:53 PM

That's not how it would work. Each time you set the alarm you schedule one execution. And you setup a new alarm in that invocation, so until you set a new alarm time, there won't be a new alarm invocation.

J

James•1/23/25, 6:09 PM

are you sure we're talking about the same thing?

alarm 1 fires, has async tasks and takes 10 seconds with some async sprinkled throughout
during the execution context switching, other things can happen on the DO

I know only one alarm can be set at a time, but in this case the alarm already fired, and its up to the developer to manage the data flow from there, especially if these tasks run async

JJames are you sure we're talking about the same thing? alarm 1 fires, has async tasks...

L

lambrospetrou•1/23/25, 6:10 PM

For multiple async requests sure, it's the developer, but previous messages were arguing about having two alarm invocations at the same time.

Llambrospetrou For multiple async requests sure, it's the developer, but previous messages were...

J

James•1/23/25, 6:14 PM

and they could be running at the "same time" if async and the alarm execution runs past whats expected, my understanding is the alarm can only have one scheduled at a time, but once fired, one could schedule it again inside the same execution, or another execution depending on how the async calls are interleaved

Aalgads OK, let me make sure I understand this. If I set a DO to invoke an alarm, the a...

森

森@VerseLink (RW)•1/23/25, 8:10 PM

You are overthinking it, it's simple, alarm like other functions are simply like a I/O event like fetch, but triggered by Cloudflare's internal system once the time is to invoke the alarm function.

森

森@VerseLink (RW)•1/23/25, 8:12 PM

so it's just like calling multiple fetch, except in this case it's just alarm, and also it automatically retries to call the alarm function just in case it threw an error

森

森@VerseLink (RW)•1/23/25, 8:15 PM

so no, they technically don't get "queued", they just get called whenever the time is up, regardless if any alaram or fetch is being invoked or not; this makes them possible to run concurrently

森

森@VerseLink (RW)•1/23/25, 8:16 PM

so if you really want to have a "queued" up behaviour, you should simply

async alarm() {
  doSomething();
  await this.ctx.storage.setAlarm(); // set the next alarm 
}

async alarm() {
  doSomething();
  await this.ctx.storage.setAlarm(); // set the next alarm 
}

async alarm() {
  doSomething();
  await this.ctx.storage.setAlarm(); // set the next alarm 
}

async alarm() {
  doSomething();
  await this.ctx.storage.setAlarm(); // set the next alarm 
}

this way the alarm will behave as if they were in queue, and have this nice added bonus that only successful alarm calls will proceed to the next alarm call

森

森@VerseLink (RW)•1/23/25, 8:18 PM

before workflow was a thing, you would implement a queue using DO's KV API, and deque one action out, execute it via alarm, and queue for the next action (pretty sure this is also what is happening underneath the hood, just not in JS land)

P

peachneo•1/24/25, 7:45 AM

I have this function:

async function exhaustAlarms(stub: DurableObjectStub) {
    let count = 0;
    while (true) {
        let ran = await runDurableObjectAlarm(stub)    
        // await new Promise(resolve => setTimeout(resolve, 100))
        if (!ran) {
            break
        };
        count++;
    }
    return count;
}

async function exhaustAlarms(stub: DurableObjectStub) {
    let count = 0;
    while (true) {
        let ran = await runDurableObjectAlarm(stub)    
        // await new Promise(resolve => setTimeout(resolve, 100))
        if (!ran) {
            break
        };
        count++;
    }
    return count;
}

async function exhaustAlarms(stub: DurableObjectStub) {
    let count = 0;
    while (true) {
        let ran = await runDurableObjectAlarm(stub)    
        // await new Promise(resolve => setTimeout(resolve, 100))
        if (!ran) {
            break
        };
        count++;
    }
    return count;
}

async function exhaustAlarms(stub: DurableObjectStub) {
    let count = 0;
    while (true) {
        let ran = await runDurableObjectAlarm(stub)    
        // await new Promise(resolve => setTimeout(resolve, 100))
        if (!ran) {
            break
        };
        count++;
    }
    return count;
}

If I don't add some small delay (see the setTimeout), I get this issue:

workerd/io/worker.c++:2082: info: uncaught exception; source = Uncaught (in promise); stack = Error: Durable Object storage is no longer accessible.

P

peachneo•1/24/25, 8:03 AM

overall, testing of DOs seems really fragile...I get inconsistent errors, different errors on live reload than hard restart etc

Ppeachneo I have this function: ``` async function exhaustAlarms(stub: DurableObjectStub)...

森

森@VerseLink (RW)•1/24/25, 9:01 AM

What is runDurableObjectAlarm

Ppeachneo overall, testing of DOs seems really fragile...I get inconsistent errors, differ...

森

森@VerseLink (RW)•1/24/25, 9:03 AM

I realized this ago, not fragile, but just the error thrown usually doesn’t align with the actual error

L

LearnNMake•1/25/25, 2:46 AM

I am researching migrating to a D1 Database. I need one source of truth for the relational data. I will be updating the database with a backend application. The frontend will be performing reads. I need everyone in the world to get the most up to date data. I read that D1 performs replication geographically so this doesn't seem like a good fit. I would rather there be a delay than a client get the wrong or outdated data. Is it possible to somehow force one persistent storage to be used world wide? Are durable objects a solution for this dilemma? Can I enforce everyone in the United States to somehow use one specific D1 Database?

LLearnNMake I am researching migrating to a D1 Database. I need one source of truth for the ...

B

boop•1/25/25, 8:24 AM

There aren't replicas on D1 yet

LLearnNMake I am researching migrating to a D1 Database. I need one source of truth for the ...

L

lambrospetrou•1/25/25, 10:39 AM

Can I enforce everyone in the United States to somehow use one specific D1 Database?

This applies both to D1 and DO. If you want to use different databases per region, you can create a few of them, and then use your worker request location (in the request headers) to decide where to route your queries (choosing the database/DO that you want). Needs a bit of code, but doable.

I read that D1 performs replication geographically so this doesn't seem like a good fit. I would rather there be a delay than a client get the wrong or outdated data.

Once D1 launches read replication you will still be able to get the latest data if needed, we will not break the default consistency. This would mean though that all your requests would need to reach the primary database (transparent to your code, we would do this for you probably), so you would basically use D1 as its current form forgoing the replication features.

More on this once it's ready and out though, replication is not released yet, so now any query you make to D1 goes to the same database, hence you always get the latest data.

森

森@VerseLink (RW)•1/25/25, 11:41 AM

is it faster to store a user's durable id (calculated from the user id) in D1, then query D1 to get the DO object faster; or is it faster to just use the user's id to get the DO?

森

森@VerseLink (RW)•1/25/25, 11:42 AM

and it is a good idea the cache the DO's id (store on client via JWT) or just get the DO via the user id everytime instead?

森森@VerseLink (RW)is it faster to store a user's durable id (calculated from the user id) in D1, t...

L

lambrospetrou•1/25/25, 1:36 PM

Definitely directly to the DO. If you go to D1 first, you are already reaching a remote database, so you might as well go to the DO directly.
Read https://developers.cloudflare.com/reference-architecture/diagrams/storage/durable-object-control-data-plane-pattern/ to get some more insight.

森森@VerseLink (RW)and it is a good idea the cache the DO's id (store on client via JWT) or just ge...

L

lambrospetrou•1/25/25, 1:38 PM

There shouldn't be a noticeable difference between these once you have regular traffic. Using

idFromName

idFromName

idFromName

idFromName has one extra hop the first time you query it but then it's cached in cloudflare's datacenters, so assuming you have regular requests, there shoudn't be a performance benefit caching the DO ID yourself.
Always benchmark though, and it always depends on your traffic. Don't optimize prematurely.

森

森@VerseLink (RW)•1/25/25, 1:42 PM

Thank you very much!

Let's say we allow users to look up any user's public metadata (for example

/user/{userid}

/user/{userid}

/user/{userid}

/user/{userid}, would it be better if we first look up D1 to see if the user id exists, or directly get the user's DO via

idFromName()

idFromName()

idFromName()

idFromName(), and see if it was previously initialized, destroy it (via

deleteAll()

deleteAll()

deleteAll()

deleteAll()) if not?

I would of course benchmark this, but would like some pointers.

森森@VerseLink (RW)Thank you very much! Let's say we allow users to look up any user's public meta...

L

lambrospetrou•1/25/25, 1:50 PM

and see if it was previously initialized, destroy it (via deleteAll()) if not?

Why would you need to delete all if nothing has been initialized?
In this case you need to make your constructor readonly, not writing anything to storage unless needed.
If you do all your table initialization in the constructor, then this won't work nicely.

Anyway, it depends on your usecase. There is a similar discussion a few weeks ago in this group about initialization done in constructor or not for publicly accessible links.

森

森@VerseLink (RW)•1/25/25, 2:19 PM

Got it, looking thru the documentation you provided, I have better understanding of what's happening behind the scene for DO.
Readonly as in it's acceptable to do

SELECT * FROM 'tablename'

SELECT * FROM 'tablename'

SELECT * FROM 'tablename'

SELECT * FROM 'tablename' and we know the DO is not initialized if the table name doesn't exist? (or querying

sqlite_master

sqlite_master

sqlite_master

sqlite_master to check if table exists)

森森@VerseLink (RW)Got it, looking thru the documentation you provided, I have better understanding...

J

James•1/25/25, 4:14 PM

DO object / stub initialization and table initialization are completely independent concepts. The constructor talk is just a convenience some developers take on.

You’ll always get errors querying tables don’t exist until you create them. How you manage migrations including table init is your call.

J

James•1/25/25, 4:16 PM

I personally find it computationally wasteful to constantly re-init or look for tables which may or may not be there on init. As in my app lifecycle, I may run millions or billions of requests on a DO between migrations I choose to create.

JJames DO object / stub initialization and table initialization are completely independ...

森

森@VerseLink (RW)•1/25/25, 4:23 PM

yeh i understand they are different concepts, i'm just concerned if i were to invoke any sql commands, it may not be readonly as I imagined

JJames I personally find it computationally wasteful to constantly re-init or look for ...

森

森@VerseLink (RW)•1/25/25, 4:23 PM

so your approach is to have a db lookup first then to invoke a DO?

J

James•1/25/25, 5:47 PM

The db is just a set of calls in the DO. You have to be in the DO to access it.

Aalgads When you say after the first hop, its cached in the local datacenter, is it cach...

L

lambrospetrou•1/26/25, 12:36 PM

There are too many variables in what you describe. How you instantiate DOs, how long that takes, is it serially or paralle, and many other things.
In general, any work you can parallelize from your worker the better. You should not pre-optimize DO stub creation until you actually measure it and see that it's the bottleneck in your code. Performance optimizations just based on intuition doesn't work. Measure and then optimize accordingly.
Start with simple code doing what you need and take it from there.

森森@VerseLink (RW)so your approach is to have a db lookup first then to invoke a DO?

L

lambrospetrou•1/26/25, 12:39 PM

My mental model for Durable Objects constructor is different than most.
As the step 3 in https://developers.cloudflare.com/reference-architecture/diagrams/storage/durable-object-control-data-plane-pattern/ describes, I prefer doing the expensive DO initialization in a dedicated

init()

init()

init()

init() call that is part of a resource creation flow. Including any initial SQL table creation and setup. After the init() completes, any other SQL migration is done only when necessary (at the point of access, not the constructor, and it's very fast/noop majority of the time when done properly: https://github.com/lambrospetrou/durable-utils?tab=readme-ov-file#sqlite-schema-migrations ).

I don't consider the constructor to be the place to do expensive initialization. The DO constructor as of today should be fast, and I only use it to initialize the "in-memory state of the DO". Loading anything from the storage into memory if necessary or just not do anything expensive.
It's different from how most folks think about the DO constructor, but in practice this works much better, especially if your DOs are not receiving enough traffic to stay warm and you have many cold starts, so you need to keep your constructor lightweight.

Llambrospetrou My mental model for Durable Objects constructor is different than most. As the s...

森

森@VerseLink (RW)•1/26/25, 2:26 PM

Thanks, yeh, I'm using Drizzle to manage migration of our database structure (just that, still write raw SQL most of the time), and the migration code it provided was less ideal so I had to incorporate your suggestion into it.

I guess premature optimization is evil, but I'm still curious about how I can approach DO init a bit better. I sort of understood how DO creates objects that were previously created (say via idFromName), but I'm still uncertain how DO creates when it's a new name id that it had never seen before. Does it ask every data center if that name id has existed before creating the object? or does it create it first anyways, and only sync at storage level?

If it was the latter, if two people decides to create the same name id DO at the same time on the opposite side of the earth, they would not see the same in memory status as the other? (or is there more memory syncing magic involved)

森森@VerseLink (RW)Thanks, yeh, I'm using Drizzle to manage migration of our database structure (ju...

L

lambrospetrou•1/27/25, 10:37 AM

Does it ask every data center if that name id has existed before creating the object? or does it create it first anyways, and only sync at storage level?

This is my personal suggestion, but you shouldn't design your applications based on internal implementation.
If DOs internally ask every datacenter, or some of them, or use a central database, should not impact your designs. The end result for you and your application is how long would this process take. So, do that, measure it, and only if it's slow for your use-case start thinking about optimizations. Internal implementations can and will change over time.

If it was the latter, if two people decides to create the same name id DO at the same time on the opposite side of the earth, they would not see the same in memory status as the other? (or is there more memory syncing magic involved)

Read https://developers.cloudflare.com/durable-objects/platform/known-issues/
The uniqueness guarantee is strictly enforced when you access storage.

L

lemmih•1/27/25, 10:39 AM

Can I emit log messages from a Durable Object?

G

geg•1/28/25, 8:11 AM

Hi, can it be that runtime updates break hibernated websockets? My DO can still send messages out but incoming messages stop arriving at random times. I noticed that it doesn't happen over the weekends and as soon as the weekend is over it starts happening almost every day, sometimes every couple of hours on a given day. My DO code couldn't be simpler, message in -> message out. I strongly believe it's runtime updates at this point. Reconnecting the "writing" websocket fixes it.

Ggeg Hi, can it be that runtime updates break hibernated websockets? My DO can still ...

L

lambrospetrou•1/28/25, 8:26 AM

Not sure about this specific issue, but you should be doing reconnection anyway, since servers can go down for many reasons, therefore your clients should attempt to reconnect (with usual jitter+backoff) if the connection fails.

Llambrospetrou Not sure about this specific issue, but you should be doing reconnection anyway,...

G

geg•1/28/25, 8:29 AM

What I meant with "messages stop arriving" is that there is a connection with no errors, yet messages don't arrive.

is it an expensive operation? and why should that be the developer's concern and not the binding/pla

Similar Threads

Similar Threads

Similar Threads