Workflow retry mechanism

[ezyyeah]

Hi!

I'm curious about the retry mechanisms.

For example, let’s say I have 3 actions in a workflow:

• The first action creates a threadId for an agent
• The second action uses that threadId to call Gemini via the "agent" component
• The third action fails 50/50 (for testing)

The response from Gemini is quite large, and I’d like to avoid repeating it if a later step fails. I know there’s a built-in retry mechanism, which works fine, but what happens if maxAttempts is reached and the workflow ends in a "failed" state?

Ideally, I don’t want to repeat the earlier steps. I would like to resume the workflow from the point of failure. Is there a way to manually trigger this kind of retry, for example using the workflowId? Or would it be better to store the workflow progress in db and check before each action whether that step has already been completed?

[ianmacartney]

Hey sorry I missed this.
The maxAttempts is per-action. It sounds a bit like you'd want to just increase the number of maxAttempts?
Or catch the error in the workflow and decide whether to run again?
Or perhaps if it fails you'd want to pause the workflow and be able to resume it later?

Since the workflow is already tracking each step, it'd be a shame to have to duplicate all that effort of short-circuiting.
What are the situations where you want to restart it, but you wouldn't want to just set a higher maxRetries?

Maybe something like the action cache would make it easier for Gemini. Or if this happens super infrequently, maybe just re-generating those results on those occasions would be okay

[srich36]

I'd also want this feature (to be able to re-run failed workflows from the dashboard / even if code if needed).

My use case is to re-run failed workflows after making code changes - though there are more use cases for this feature than just that.

So the flow would be:

• workflow (and all retries for a given step) fails
• fix bug in code
• re-run the workflow from the dashboard given a workflow ID

I've used hatchet.run and AWS step functions in the past and re-running workflows from the UI has always been incredibly useful.

Eventually it would be nice to re-run from failed steps only, but I'd be fine to start by having a way to easily re-run entire workflows with the same arguments.

[ianmacartney]

Resuming from a failed step sounds great - and definitely possible. Until then, the challenge is getting the original arguments?

…

On Fri, Jul 11, 2025 at 3:34 PM Sean Rich ***@***.***> wrote: *srich36* left a comment (get-convex/workflow#38) <#38 (comment)> I'd also want this feature (to be able to re-run failed workflows from the dashboard / even if code if needed). My use case is to re-run failed workflows after making code changes - though there are more use cases for this feature than just that. So the flow would be: - workflow (and all retries for a given step) fails - fix bug in code - re-run the workflow from the dashboard given a workflow ID I've used hatchet.run and AWS step functions in the past and re-running workflows from the UI has always been incredibly useful. Eventually it would be nice to re-run from failed steps only, but I'd be fine to start by having a way to easily re-run entire workflows with the same arguments. — Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACZQW6RNTXXOFKOPW2XZPD3IA3YFAVCNFSM6AAAAAB6CHQ3V6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTANRUGA4DANRQGU> . You are receiving this because you commented.Message ID: ***@***.***>

[srich36]

@ianmacartney

Resuming from a failed step sounds great - and definitely possible.
Until then, the challenge is getting the original arguments?

Would be amazing to resume from a failed step! This would be a huge unlock.

Until then, yes, the challenges are:

1: Having an easy way to run a workflow from the dashboard with arguments we can pass in (is there an easy way to do this?)
2: Re-running a workflow given a workflow id (so under the hood this would need to get the original arguments I'm assuming). This is a nice-to-have beyond #1.

Happy to clarify anything else there!

[ianmacartney]

For (1) you can write an internalMutation that enqueues the work from arguments, then in the dashboard copy the args from the component's table? It's a workaround for sure, but something you could do today vs. when I have bandwidth to add features

…

On Sat, Jul 12, 2025 at 11:15 AM Sean Rich ***@***.***> wrote: *srich36* left a comment (get-convex/workflow#38) <#38 (comment)> @ianmacartney <https://github.com/ianmacartney> Resuming from a failed step sounds great - and definitely possible. Until then, the challenge is getting the original arguments? Would be amazing to resume from a failed step! This would be a huge unlock. Until then, yes, the challenges are: 1: Having an easy way to run a workflow from the dashboard with arguments we can pass in (is there an easy way to do this?) 2: Re-running a workflow given a workflow id (so under the hood this would need to get the original arguments I'm assuming). This is a nice-to-have beyond #1 <#1>. Happy to clarify anything else there! — Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACZQWZ25VZL5TA6N65YVG33IFGEFAVCNFSM6AAAAAB6CHQ3V6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTANRVHEZTAMJYHE> . You are receiving this because you were mentioned.Message ID: ***@***.***>