Hi I'm looking to use Cloudflare DO x Websockets, can anyone share their experience on reliability a
Hi I'm looking to use Cloudflare DO x Websockets, can anyone share their experience on reliability and scalability of it?
LThe key is to assume the connection will break often and design for reconnection and resumability. If the server never needs to initiate communication, I use a lazy [re-]connection strategy where the next time the client needs to send something, it first checks if the connection is up and if not re-connects. For situations like chat where you need the server to reach out to the client, you will have to use the hooks for sensing when the connection is broken and reconect. There is a package that Cloudflare Agents use for this purpose. I have forked that for my own purposes but you could just use that.
SIs there a way to register an unhandledrejection event handler in Durable Objects?
LWhat do you mean by "rejection"?
SCloudflare Docs
Standardized APIs for use by Workers running on Cloudflare's global network.
LWhen I return a DO stub over Workers RPC, in vitest environment, the connection breaks unless I use setTimeout. Is this a known limitation of vitest, also what happens in prod, or am I just doing something wrong? I can share code if you like?
1Are you returning an
RpcTargetAlso for testing, is the DO class being returned in the same Worker? If not, you might need to run both Workers concurrently but I don't have much experience there
Jlooks like JSONB support is now available in sqlite DOs, thanks so much to whoever made this change!
1There needs to be a better way to track logs on a DO, especially when working with WebSockets. I have a DO that manages inbound (hibernated) sockets as well as multiple outgoing socket connections, as well as alarms and HTTP requests. Every WS message I receive I log, but those logs can be split between the different events that the DO manages. I can see the full list of logs for the current IOContext() of the outgoing sockets within the event invocation of the alarms, which feels very unintuitive. Should I be pushing the logs to observability or elsewhere instead of relying on the built-in logic?
LI had a chance over the last week to do a deep dive on Cap'n Web. Here is my multi-page writeup as a comparison to my own Lumenize RPC. Please let me know what you think. I'm particularly interested in hearing if I've misconfigured anything on the Cap'n Web side which would explain some of the troubles I had: https://lumenize.com/docs/rpc/capn-web-comparison
Comparing Lumenize RPC to Cloudflare Cap'n Web
LIs there a way for a DO instance to know what binding name was used to connect to it without explicitly passing that in? I know it's tricky with service bindings to it's not consistent so I'm guessing the answer is "no" but thought I'd ask.
HNot yet. The tricky part is that names aren’t bounded by size, so theoretically you could have a 50 MB name, which is not great to be shipping around the world. Though iirc the team is looking into supporting it somehow
AI'm seeing some strange behavior after destroying a durable object.
I have this destory method():
When it executes, I see normal logs:
Then strangely, a few seconds later I see this:
After it's destroyed, something else wakes up the DO only to warn me that an alarm was replaced. Why is this happening?
I have this destory method():
When it executes, I see normal logs:
Then strangely, a few seconds later I see this:
After it's destroyed, something else wakes up the DO only to warn me that an alarm was replaced. Why is this happening?
AYes, the destroy() is called from within an alarm.. it's a sort of TTL where I tell the DO to destroy itself sometime in teh future. I do catch the exception thrown by abort() tho...
Aand the destory is working, i can see everything wiped in the sqlite file
Athat I don't know, just following code i saw in some examples
LThanks, while we're on the subject, is
ctx.idgetByName()
HIt should be, since
getByName.get(DO.idFromName()
LYes, the ID is always populated and correct. And you can route to a DO by the ID even if other callers route by name.
LI don't see a channel for Cloudflare Queues but please excuse me if this is the wrong channgel to ask on, but what's the expected latency for Cloudflare Queues between the time you push a message in and the time your Queue consumer is called?
C#queues
Cit can def take a few seconds though How to make queues pick up messages
LI am never sure I get the unit conversions correct. I've seen folks say it costs about $4/month to run a DO 24/7, but I've also seen $0.34/day which works out to about $10/month. Can someone show me the math?
L1 DO running 24/7 needs:
You get 400,000 GB-s for free.
The above is duration only charges.
edit: using the correct
0.128 GB-s * 3600 seconds_per_hour * 744 hours_in_month = 342_835 GB-s/month$12.50/million GB-s342_835/1_000_000*12.50You get 400,000 GB-s for free.
The above is duration only charges.
edit: using the correct
0.128 GB-s0.125GB-sThanks! I only need duration charge help. I can count requests, just not unit convert. 
PIs there any way to get a hook to run code before a DO is evicted from memory? Would love to use it to flush in memory stuff to sql storage rather than needing to write a value that will be modified often in short bursts.
LThere is no such hook and even if there were, it wouldn't always work because it can evicted for crazy reasons. The saving grace is that reading from and writing to storage is very fast due to caching and locality. I have switched to a coding style where I don't have instance variables unless the in-memory representation is an expensive transformation from the in-storage representation. It's counter-intuitive but think of storage like you used to think of instance variables.
LFor the cases where the in-memory transformation is expensive, I have also adopted a coding style where I save and retrieve that largish structure in one whole shot using the fact that the KV API let's you store just about everything (objects with cycles or aliases, Date, Map, Set, Infinity, NaN, etc.). See the complete list in the second column here
PThat's pretty fair, wishful thinking I guess. Wanted to avoid the $1/million write charges because it's an extremely bursty workflow (counting a rapid stream of events), but honestly not that bad
Pthe storage backend KV-like API for sqlite is OP though
LReads are so inexpensive I ignore the cost. Write costs can add up.
PYup, that's why I was wondering if I could just flush it with a handle
PGuess this might be a better task for a container
LI don't know the costs for containers but I find it hard to image it would be less under the vast majority of circumstances.
POn average it will undergo ~250,000 events in a 8 minute span, and then go silent for up to a month, so about $0.25 in writes vs probably a couple cents for the container
Pvery niche/weird usecase lol
1Could the events be grouped in some way? Store groups of events per write request
LFor $0.25, I wouldn't bother optimizing. Not even for $25. Maybe at $250 if it was simple to implement and manage but containers are generally harder to manage in my experience.
PUnfortunately I don’t control that side. Been asking for that, but they don’t want to do it
PI totally agree, but I’m basically donating the cost to manage this out of pocket currently. I was originally going to just spin up a VPS but figured scale to zero could save a bit (and scale horizontally since I don’t know how many events per month will be happening initially)
PIt’s for managing some stuff for a highschool competition being spun up. Infra is a bit of a mess but until it goes public and gets sponsored basically just doing it for free.
1If it is actually only once a month then you'll stay within the included limits with workers paid anyway
1Doing lots of writes sucks but I wouldnt worry unless your use case would easily push you over included limits
PIt’s once a month per team per event
1ahhh
PAnd I have no clue what the number of teams or events will be
1THat might be an issue lol
PYeah
1What are you doing with the data afterwards?
PIt gets processed into a leaderboard, basically