Cloudflare Developers•12mo ago

rpc has limit of i think 1mb arguments/response over the wire

BBeefp rpc has limit of i think 1mb arguments/response over the wire

lambrospetrou•1/9/25, 7:15 PM

You can pass and return streams to overcome the 1mb limit. These streams can even be inside an object.

rob•1/9/25, 7:37 PM

if im looking to do a real time chat with rooms is partyserver the bst bet to get going? Is there any starter examples for that

marko.333•1/9/25, 9:11 PM

I'm a little uncertain of the concurrency rules for Durable Objects RPC calls?
If i have a DO("id1").doStuff() and many concurrent clients are connected to same DO instance ("id1"), then (roughly) all callers to doStuff() are serialized when a storage call is in-progress.
But i wanted to confirm that serialized execution is only triggered for storage calls - correct?
For example, what if the doStuff() implementation included some external calls (to another Worker, or external API fetch, ...)?
While await'ing for the external call to complete, will the other concurrent callers of doStuff() be able to continue execution (ie. an external API call will not cause all other concurrent callers to block)?
Thx.

Mmarko.333 I'm a little uncertain of the concurrency rules for Durable Objects RPC calls? ...

lambrospetrou•1/10/25, 9:49 AM

Yes, if you do an async IO call then other requests will run. If you want to fully block concurrency you can use https://developers.cloudflare.com/durable-objects/api/state/#blockconcurrencywhile

mihaaai•1/10/25, 1:58 PM

When using a sqlite-based Durable Objects, is Libsql for DO storage increasing the cursor memory with each read operation? Asking because we're trying to figure out final costs based also on the Duration GB-s constraint.

Rrob if im looking to do a real time chat with rooms is partyserver the bst bet to ge...

harshil1712•1/10/25, 2:28 PM

Hey, I have used PartyServer and love it because it makes it takes away a lot of boilerplate code. Also, I found this example in the repo: https://github.com/threepointone/partyserver/blob/main/fixtures/chat/README.md
Also, FYI, the repo says it is a Work in Progress

GitHub

partyserver/fixtures/chat/README.md at main · threepointone/partyse...

PartyKit, for Workers. Contribute to threepointone/partyserver development by creating an account on GitHub.

Mmihaaai When using a sqlite-based Durable Objects, is Libsql for DO storage increasing t...

lambrospetrou•1/10/25, 3:17 PM

GB/s is not checking how much memory your code uses, since you are charged for the 128MB allocated to your DO. See details in https://developers.cloudflare.com/durable-objects/platform/pricing/
So, regardless if you use 100MB or 10MB while iterating your results, the billing won't change.

5 Duration billing charges for the 128 MB of memory your Durable Object is allocated, regardless of actual usage. If your account creates many instances of a single Durable Object class, Durable Objects may run in the same isolate on the same physical machine and share the 128 MB of memory. These Durable Objects are still billed as if they are allocated a full 128 MB of memory.

Hharshil1712 Hey, I have used PartyServer and love it because it makes it takes away a lot of...

rob•1/10/25, 3:19 PM

yes thsi! i was looking if someone/s set it up before and could asisst a bit

Rraoulduke Hey I have a problem with my connections map for my web socket not maintaining s...

1984 Ford Laser•1/10/25, 8:42 PM

what happens when you do

this.state.getWebSockets()

this.state.getWebSockets()

Hharshil1712 Hey, I have used PartyServer and love it because it makes it takes away a lot of...

ray•1/10/25, 11:45 PM

I've been trying to set up partyserver for my use case and was wondering if you could share a bit of your experience; i have multiple durable objects that handle different logic and I'm not sure how best to run local dev and deploy

ray•1/10/25, 11:48 PM

Should each DO be defined with it's own wrangler.toml and have a root package.json that runs

wrangler dev

wrangler dev

in each folder? Or should i have all the DO bindings defined on a root wrangler.toml? In my case I have a sveltekit app that I'm also deploying on a worker, and was wondering if I should be defining all my DO bindings on my sveltekit wrangler.toml

peachneo•1/11/25, 2:48 AM

How do alarm() retries interact with setAlarm() ?

if alarm() throws thereby queueing a retry - will a subsequent call to setAlarm() "reset" the retry?

Rray Should each DO be defined with it's own wrangler.toml and have a root package.js...

Ephemeros•1/11/25, 11:11 AM

i dunno about partyserver but I have that setup, each DO and worker has its own wrangler and they share dbs etc like this

e.g.

durable-objects/authdo

"dev": "wrangler dev -e local --show-interactive-dev-session false --ip 0.0.0.0 --persist-to ../../.wrangler/state",

durable-objects/chatdo
"dev": "wrangler dev -e local --show-interactive-dev-session false --ip 0.0.0.0 --persist-to ../../.wrangler/state",

apps/worker-app
"dev": "wrangler dev -e local --show-interactive-dev-session false --ip 0.0.0.0 --persist-to ../../.wrangler/state",

EEphemeros i dunno about partyserver but I have that setup, each DO and worker has its own ...

ray•1/12/25, 2:28 AM

Appreciate you sharing this! Really helps me wrap my head around how I should structure things, are you also binding your authDO on your worker-app wrangler file?

Rray Appreciate you sharing this! Really helps me wrap my head around how I should st...

Ephemeros•1/12/25, 7:47 AM

Yep, I have some as service binding some as durable object bindings

Ephemeros•1/12/25, 3:19 PM

Totally, also when you separate them and can deploy only what changes it's a superpower, like zero downtime

Ephemeros•1/12/25, 3:19 PM

Especially if you can have multiple deployments for your workers and DOs and you can swap from one to another with a switch, really zero downtime

Ephemeros•1/12/25, 3:20 PM

And then we are back to.. microservices

Ephemeros•1/12/25, 3:21 PM

I made it kinda overkill with local/dev/staging/prod environments but it just was too easy to set that up lol

Rray I've been trying to set up partyserver for my use case and was wondering if you ...

BeefpOP•1/12/25, 6:25 PM

We started out with 6 DO classes in a monolith worker. In the beginning it was a lot easier to work with (single deploy, no bindings) but we hit perf issues because our monolith bundle was so large and Cf loads it on every Worker request or every individual DO request. We are currently splitting into multiple Workers which means transfer migrations as we DO storage heavily

Personally I'd love if Cloudflare improved support for smarter loading/importing based on whether the code is executing as a Worker or as a DO, and which DO class specifically. Doesn't seem necessary to load everything at once if the Worker scripts are set up the right way

BBeefp We started out with 6 DO classes in a monolith worker. In the beginning it was a...

ray•1/13/25, 4:16 AM

Thanks for the heads-up, definitely will be deploying separately based on your experiences, although local dev experience is going to suck until they release the v5 terraform sdk

Isak•1/13/25, 9:44 PM

Interesting thanks for this. How big was your worker and what performance gains did you observe? @Beefp

IIsak Interesting thanks for this. How big was your worker and what performance gains ...

BeefpOP•1/13/25, 10:13 PM

I started writing but it turned into a wall of text. I'll summarize here but I want to write a more detailed thing later on our blog. We measured two things:

Bundle startup time which you get as output from Wrangler when you deploy. We went from 200ms -> 40ms just by simplifying dependencies, lazy-loading where possible, splitting out heavier packages.
Response time hitting a simple Hello World endpoint. We had endpoints just for the Worker and one for each DO class. Response times on the Worker went from about 500ms to sub 200. Going through the Worker to the DO is a bit more.

We started off with a lot of low hanging fruit like bundles that were doing all their processing when you imported them, and some packages like limax and OpenAI which have a heavy overhead (they went into different Workers).

BeefpOP•1/13/25, 10:16 PM

if you are dealing with a monolith my recommendation is: build your worker locally and load it into Chrome with

---inspect-brk

---inspect-brk

you'll see exactly where time is being spent. I think once you've optimized on your side the only thing you can do is architect your DOs and Workers in a way that they're closer to your users which is another topic + what we're dealing with now (moveable DOs when pls)

TTyveris I’m considering using Durable Objects with WebSockets and hibernation for a game...

1984 Ford Laser•1/14/25, 2:42 AM

It makes sense to me
According to this page: https://developers.cloudflare.com/durable-objects/best-practices/websockets/#extended-methods the serialized attachment will survive hibernation, I think there is always a chance of losing the data such as in the event of CF maintenance/outages, or a reboot on the overall runtime, or you deploy a new version (I think)
Memory is reliable for 5sec, after that there's zero guarantee, if you know you won't need it after 5 seconds then you'd be fine

All of this is in theory, I'd be interested in how your solution works as I'm also interested in the websocket attachment methods

1984 Ford Laser•1/14/25, 4:05 AM

I'd love if

namespaceId

namespaceId

and

objectId

objectId

were added as dimensions to the

durableObjectsStorageGroups

durableObjectsStorageGroups

in the graphql API, so we could calculate used storage per DO instance

mihaaai•1/14/25, 4:54 PM

I noticed that in a SQLite based Durable Object, any

INSERT

INSERT

operation against the storage returns 2

rowsWritten

rowsWritten

, although I only inserted one row, is this normal behavior or am I missing something?
It's also unlikely that I exceeded any 4k limit for write units, although I'm not sure if it applies to SQLite based DOs, and I don't have any index on the table I'm writing to.

Mmihaaai I noticed that in a SQLite based Durable Object, any `INSERT` operation against ...

James•1/14/25, 4:59 PM

could be the default index for rowid, or a trigger for updated ts updates

JJames could be the default index for rowid, or a trigger for updated ts updates

mihaaai•1/14/25, 5:02 PM

Interesting, is there any way to disable that index on

rowid

rowid

?
I'm trying to optimize pricing since my use case is a write heavy one.

James•1/14/25, 5:03 PM

https://www.sqlite.org/withoutrowid.html

James•1/14/25, 5:04 PM

i wonder if the write is related to updating the meta table for bumping sequences related to rowid, and perhaps rowid doesnt cost in itsself

JJames https://www.sqlite.org/withoutrowid.html

Isaac McFadyen•1/14/25, 5:05 PM

That docs page also says that even without a Row ID you still need a primary key with a unique index.

JJames https://www.sqlite.org/withoutrowid.html

mihaaai•1/14/25, 5:06 PM

It seems to solve the problem, now I get the correct count for the written rows.

James•1/14/25, 5:07 PM

still could be sequence related, since every new row writes to the sqlite_sequence meta table, and if that counts, that kind of sucks

mihaaai•1/14/25, 5:09 PM

Indeed, I noticed that using

WITHOUT ROWID

WITHOUT ROWID

the sequence of written data is not ordered as before

James•1/14/25, 5:10 PM

temps me to switch all my PKs to random guids instead of rowid

Original message was deleted

Larry•1/14/25, 11:12 PM

UUIDv7 is the standardized version of ULID so I recommend looking for that over an ULID library

dandoen•1/15/25, 1:56 PM

hi folks,

i'm leveraging DO for a websocket w/ hubernation support, but keen to understand how i can pass on context to my "webSocketMessage". i want to simply pass on the user id of the user into the webSocketMessage handler. when it gets hibernated, it seems to get lost. i'll drop my fairly basic code in a thread.