Cloudflare Developers•13mo ago

Can we create workflows with workers for platforms?

Mmakrdev Can we create workflows with workers for platforms?

Matt Silverlock•12/5/24, 11:06 AM

You can’t accept user-created scripts as Workflows that your WfP dispatcher calls, if that’s what you’re asking.

MMatt Silverlock You can’t accept user-created scripts as Workflows that your WfP dispatcher call...

makrdevOP•12/5/24, 11:09 AM

What if I create the initial workflow definition and run dynamic steps? Im trying to give each tenant its own instance with separate limits.

Mmakrdev What if I create the initial workflow definition and run dynamic steps? Im tryin...

Matt Silverlock•12/5/24, 11:11 AM

You can run whatever code you like: you specifically asked about WfP. If users aren’t writing any code then I wouldn’t call that WfP.

If they (untrusted users) are supplying code, then you’ll want WfP.

Sskip Is there a way to trigger another workflow then pause until it completes and rec...

skip•12/5/24, 11:12 AM

So, sort of like making a DO step that forks to another seperate workflow

MMatt Silverlock You can run whatever code you like: you specifically asked about WfP. If users a...

makrdevOP•12/5/24, 11:15 AM

Maybe I explained in a wrong way. Basically what I want to achieve is deploying separate worker for each tenant with a pre-defined workflow ( I assume they have separate concurrency limits etc.) and running dynamic workflows with a config(created by a no-code app you shared as an example)

Sskip Is there a way to trigger another workflow then pause until it completes and rec...

DeChill•12/5/24, 11:30 AM

I'm looking for something similar. I cannot find a way to get the "Workflow" output. It is shown in the Portal after the Worfklow finished, but

await instance.status()

await instance.status()

only gets me an array of the output from the individual steps. Not the final Worfklow return value.

Focux•12/5/24, 3:59 PM

@skip i asked that some days ago, and currently, you can't. they said they will take the feedback into account

Focux•12/5/24, 3:59 PM

workflows

Focux•12/5/24, 4:01 PM

i wish there were a public roadmap or some way to add/upvote feature requests

DeChill•12/5/24, 6:22 PM

How would I programmatically get the Workflow

Output

Output

? Is it only available via the GraphQL API?

DDeChill How would I programmatically get the Workflow `Output`? Is it only available via...

Seekerdasbatatas•12/5/24, 9:52 PM

You can get it from api.cloudflare.com (sorry I'm on the phone)

Seekerdasbatatas•12/5/24, 9:52 PM

Or via the binding

SSeekerdasbatatas Or via the binding

DeChill•12/5/24, 9:59 PM

Thanks for the reply, but how can I get it via the binding?

await instance.status()

await instance.status()

only gives me the output of the steps inside the workflow, not the workflow return value.

DDeChill Thanks for the reply, but how can I get it via the binding? `await instance.stat...

Focux•12/5/24, 10:42 PM

await instance.status().output

await instance.status().output

FFocux `await instance.status().output`

Isaac McFadyen•12/5/24, 10:48 PM

^ but you'll need brackets around it because the promise is the

status()

status()

not the

output

output

, so

(await instance.status()).output

(await instance.status()).output

FFocux `await instance.status().output`

DeChill•12/5/24, 11:47 PM

I'm reading the output value, but as I wrote in my message, the

output

output

property of the

InstanceStatus

InstanceStatus

does not contain the Workflow return value, even when the workflow is completed. I'm only seeing an array of the outputs from the individual steps called inside the workflow.

oof2win2•12/5/24, 11:51 PM

would this be valid workflows code? everything is deterministic as all results are computed from steps, and i just await them all at the end

const generateTitleStep = step.do(
  "Generate the project title",
  async () => {
    console.log("Generate the project title");
    const anthropic = createAnthropic({
      apiKey: this.env.ANTHROPIC_API_KEY,
    });

    return {};
  },
);
const generateNotesStep = step.do(
  "Generate the project notes",
  async () => {
    console.log("Generate the project notes");
    return {};
  },
);
// option A
await step.do("Await all project generation steps", async () => {
  await generateTitleStep;
  await generateNotesStep;
});
// option B
await Promise.all([generateTitleStep, generateNotesStep]);

const generateTitleStep = step.do(
  "Generate the project title",
  async () => {
    console.log("Generate the project title");
    const anthropic = createAnthropic({
      apiKey: this.env.ANTHROPIC_API_KEY,
    });

    return {};
  },
);
const generateNotesStep = step.do(
  "Generate the project notes",
  async () => {
    console.log("Generate the project notes");
    return {};
  },
);
// option A
await step.do("Await all project generation steps", async () => {
  await generateTitleStep;
  await generateNotesStep;
});
// option B
await Promise.all([generateTitleStep, generateNotesStep]);

oof2win2•12/6/24, 12:25 AM

and can i await the output of one step in another step? i first need to generate data and then i need to save data - don't want to regen data if only the saving failed

const generateNotesStep = step.do(
  "Generate the project notes",
  async () => {
    console.log("Generate the project notes");
    return "These are some notes returned after awaiting";
  },
);

const saveNotesStep = step.do(
  "Save the project notes",
  async () => {
    const notes = await generateNotesStep
    // save the data to the DB here. notes have been generated and won't regenerate if this retries
    return {};
  },
);

const generateNotesStep = step.do(
  "Generate the project notes",
  async () => {
    console.log("Generate the project notes");
    return "These are some notes returned after awaiting";
  },
);

const saveNotesStep = step.do(
  "Save the project notes",
  async () => {
    const notes = await generateNotesStep
    // save the data to the DB here. notes have been generated and won't regenerate if this retries
    return {};
  },
);

oof2win2•12/6/24, 12:27 AM

basically i want to start multiple AI processes asynchronously so that they can finish not in order but all are essentially guaranteed to finish

kchro3•12/6/24, 12:51 AM

we're seeing an issue where some workflows are getting stuck in the queue. I only have two queued workflows & nothing concurrently executing & i can't terminate them from the dashboard.

any suggestions here or known issues?

FFocux @skip i asked that some days ago, and currently, you can't. they said they will ...

skip•12/6/24, 6:43 AM

Ah thanks for the reply. This is a feature that would make workflows hugely useful, imo!

Ooof2win2 would this be valid workflows code? everything is deterministic as all results a...

DeChill•12/6/24, 8:11 AM

According to https://developers.cloudflare.com/workflows/build/rules-of-workflows/, awaiting multiple steps with

Promise.all

Promise.all

is fine.

But nested steps seem like they might cause issues with the retry logic? I would suggest just placing the

saveNotesStep

saveNotesStep

after the

Promise.all

Promise.all

and reading the output of the step. In that case if the "save" step fails, the

generateNotesStep

generateNotesStep

will have cached the response. It might only be an issue if the serialized data exceeds the limits.

DDeChill According to https://developers.cloudflare.com/workflows/build/rules-of-workflow...

oof2win2•12/6/24, 8:14 AM

the issue is that generatesNotesStep might complete earlier than generateTextStep, so i want to save it to the db as soon as it generates

oof2win2•12/6/24, 8:15 AM

or should i just make it two nested steps? or an async function that awaits the first step and then saves it (two await step.do calls)

oof2win2•12/6/24, 8:15 AM

the last one kinda makes sense

DeChill•12/6/24, 8:21 AM

If you nest the save step inside the generate step, then I don't think it will cache the generate response if the inner save fails. So I think you need to keep them separate. 2 await steps are probably most idiomatic, but the second will only run when the first is finished.

SSeekerdasbatatas Or via the binding

DeChill•12/6/24, 3:19 PM

I really hope getting the workflow output by binding will be implemented in the near future! Is there any roadmap or feedback site for workflows?

D Trombett•12/6/24, 7:33 PM

From the workflows get started guide:

Because Workflows can be long running, it is possible to have running instances that represent different versions of your Workflow code over time.

Does it mean that if, for example, I update my code while the workflow is waiting a

sleep()

sleep()

step, when the sleep is over it will execute the old code and not the updated steps (considering that the step names are the same if it matters)?

DD Trombett From the workflows get started guide: > Because Workflows can be long running, i...

Hello, I’m Allie!•12/7/24, 12:45 AM

Yeah, that is correct, since Workflows itself doesn’t know whether the names are the same or not, or whether it is safe to mix a run like that

Warren•12/7/24, 2:44 AM

I am using Workers in Workflows and I am running up against the 1000 subrequests limit. my question is is there a way to clearly see what are the subrequests made in a Worker?

so for context, my worker is trying to do long running job in batches. the issue is I am using another Durable Object as a rate limiter. the problem with rate limiting is I can’t estimate how frequently the rate limiter will be called in a single worker run.

now that I’ve gotten rid of the rate limiter, it seems to not run into the subrequest issue anymore.

my feedback would be:

if a Worker in a Workflow has hit the subrequest limit, it seems pointless to retry since any retries would also by definition hit the same limit
I personally feel that this subrequest limit really dampens the promise of Cloudflare platform being serverless and scales infinitely, esp if it includes calling internal Cloudflare resources like Durable

WWarren I am using Workers in Workflows and I am running up against the 1000 subrequests...

Hello, I’m Allie!•12/7/24, 3:30 AM

On your first point, not necessarily, since if the first run makes the 1,000 requests, stores its state, and is restarted, it will skip those first 1k requests

HHello, I’m Allie!Yeah, that is correct, since Workflows itself doesn’t know whether the names are...

D Trombett•12/7/24, 12:46 PM

I see, thanks! So, if I want to "switch" to the new code, should I implement it myself like cancelling the sleeping workflow and running a new one?

D Trombett•12/7/24, 12:49 PM

Oh, wait, is it possible to call a worker from a workflow with rpc, like we do from workers? This way I could implement the logic that may change in the worker and call it from the workflow

D Trombett•12/7/24, 3:09 PM

this seems much better, thanks a lot!

Warren•12/7/24, 3:15 PM

I have observed that the try-catch behavior described in the docs do not reliably happen. i.e. sometimes when the workflow fails, the catch part of the code does not run. this is quite bad when I build systems that rely on the catch to maintain state...

Cloudflare Docs

Sleeping and retrying · Cloudflare Workflows docs

This guide details how to sleep a Workflow and/or configure retries for a Workflow step.

Dmitry•12/7/24, 10:56 PM

Any ideas about when performance issues will be fixed? (In relation to starting a bunch in a short time span and dashboard just loading)? Assuming with holidays right around the corner.. prob some time Jan?

Dmitry•12/7/24, 10:56 PM

Would love for the dashboard for Workflows to be more real time

Dmitry•12/7/24, 10:57 PM

I know it refreshes every x seconds.. but would be nice for things to stream in/auto update in realtime

Dmitry•12/7/24, 10:57 PM

Also adding a refresh button so I don't have to reload the entire page would be awesome too

inaridiy•12/8/24, 4:55 AM

Is there an upper limit to the size of the Workflow Payload? If I pass a payload that exceeds 1 MiB in size, the instance will not be triggered.

oof2win2•12/8/24, 5:44 AM

do queued instances get killed after a certain time? they aren't getting triggered and are constantly being pushed to the front of the dashboard so it's just annoying

oof2win2•12/8/24, 5:49 AM

also is there a limit to how many workflows can be created per second? i have a database webhook that can execute say 40 calls per second and some instances just don't seem to get a payload and execute

oof2win2•12/8/24, 5:50 AM

like the input is say 500 bytes so don't think i should have any issues there

oof2win2•12/8/24, 6:13 AM

yeah it just happens that the tail end (last ~12 of my ~50 workflow invocations) get empty params and don't execute at all because of that

oof2win2•12/8/24, 6:16 AM

code works in all cases except this tail end which is just frustrating. an empty object as a body doesn't make sense here - i check for props on

body.record

body.record

and they aren't null

const workflowBody: WorkflowParams = {
  coursework: {
    id: body.record.id,
...
  },
  prompt_id: body.record.prompt_id,
};
await env.grading_workflow.create({
  id: submissionId,
  params: workflowBody,
});

const workflowBody: WorkflowParams = {
  coursework: {
    id: body.record.id,
...
  },
  prompt_id: body.record.prompt_id,
};
await env.grading_workflow.create({
  id: submissionId,
  params: workflowBody,
});

Iinaridiy Is there an upper limit to the size of the Workflow Payload? If I pass a payload...

oof2win2•12/8/24, 6:59 AM

the failure rate is fairly low but it is there and it scales a lot when i scale the requests - from 12k executions i got around 600 that didn't execute at all, then from like 400 today i got 50. is there a way to debug these issues? i feel like i have the same issue as above but without the 1MB requirement - all payloads are ~400 characters/bytes

Original message was deleted

oof2win2•12/8/24, 7:06 PM

thank you! is there any eta on this?

oof2win2•12/8/24, 7:28 PM

thank you very much!

Warren•12/9/24, 12:45 AM

ah for anyone else facing the

Serialized RPC arguments or return values are limited to 1MiB

Serialized RPC arguments or return values are limited to 1MiB

error

I don’t think it’s possible to catch and handle this error directly since it happens at the infra rather than the app level
something I’ve found that works is to check the return size before returning, in which case you could either throw an error or internalize the error handling somehow

const generateNotesStep = step.do( "Generate the project notes", async () => { console.log("Generate the project notes"); return "These are some notes returned after awaiting"; }, ); const saveNotesStep = step.do( "Save the project notes", async () => { const notes = await generateNotesStep // save the data to the DB here. notes have been generated and won't regenerate if this retries return {}; }, );

const workflowBody: WorkflowParams = { coursework: { id: body.record.id, ... }, prompt_id: body.record.prompt_id, }; await env.grading_workflow.create({ id: submissionId, params: workflowBody, });

Can we create workflows with workers for platforms?

Similar Threads

Similar Threads

Similar Threads