Production incident: Exponential CPU Growth Over Time Until Container Restart

[matteotarantino-algor]

Description

We are experiencing a critical anomaly where CPU usage grows exponentially over time despite no increase in client load or traffic.
The issue affects an Hocuspocus server running on AWS ECS Fargate, where CPU and memory remain stable for hours, then suddenly accelerate upward until the container becomes saturated.

The anomaly persists indefinitely and is only resolved by manually redeploying / restarting the container, which immediately resets CPU and memory back to their normal baseline.
This strongly suggests an internal infinite loop, runaway task, or repeated command cycle inside the server (potentially related to Yjs updates, PubSub, or markdown parsing), which is not triggered by real user actions.

No spikes in connected clients, document updates, or network traffic are observed that would explain the CPU increase.

Screenshot, video, or GIF

Environment?

• operating system: AWS Fargate (node image)
• Hocuspocus version: 3.4.1
• TipTap version: 3.11.0

Additional Context
The server uses TipTap v3 with Markdown Manager v3 to perform conversions between
Y.XmlFragment <-> markdown string with these extensions:

export const DefaultTipTapExtensions: Extensions = [
  StarterKit,
  Mathematics,
  Table,
  TableCell,
  TableHeader,
  TableRow
];

This parsing step might be involved, but the CPU growth happens even during periods of extremely low usage and when no active conversions are expected.

The problem appears unrelated to load and looks instead like an internal loop or uncontrolled background task that keeps consuming CPU until the container is restarted.

[janthurau]

Hi @matteotarantino-algor,

sorry to hear you're having issues in production. Please note that I'm unable to provide mission-critical support here ; if you are using Hocuspocus in production scale, you way want to consider the cloud or on-prem offering of Tiptap Collaboration (contact humans@tiptap.dev or signup on the website).

I have seen similar issues really rarely in our clusters. They were usually related to the Yjs processing taking a lot of time and thus blocking the main thread, but debugging this is usually tricky. Unfortunately there are no ways to limit the Yjs updates during integration. You can either try adding manual logging, or using commands like strace to debug what exactly is happing during these high-cpu times (but strace probably won't work if you use Fargate ECS).

If you can reproduce the issue and provide a code sample to do that, I'm happy to see if we can implement a fix in Hocuspocus.

[matteotarantino-algor]

Hi @janthurau , thanks a lot for the quick reply!

At this point, the next logical step on our side is to add proper monitoring. Since your experience with Hocuspocus goes far beyond mine, if the issue is really caused by Yjs updates blocking the main thread, I’d love to understand where you would place the instrumentation.

In short: what is the smartest place to add logs so that we can capture the right signals as soon as the issue happens again?
Any guidance on which hooks or events are most likely to reveal what’s going on would be extremely helpful.

About hosting: I don’t think this is related to our infrastructure. Even with over-provisioned resources on AWS ECS Fargate and autoscaling, the behaviour looks much more like a logic-level issue rather than something caused by where the server is running. We also use the latest versions of both Tiptap and Hocuspocus, so I would expect the same behaviour anywhere.

That said, I’d really like to get closer to you and the team behind the technology. We are definitely interested in some form of premium or priority support. Is there a more direct path for getting in touch with you, aside from the humans@tiptap.dev email?

I also believe this could benefit both sides. Our platform (Algor Education) has around 10k daily active users creating long, complex documents in a Notion-style editor that relies heavily on BlockNote. We are about to release a new editor that combines all BlockNote text features with our custom tools (mind maps, quizzes, flashcards, etc.), all built on top of Tiptap. This will significantly increase the amount of real-time editing happening every day, and if the issue is really tied to Yjs updates, the chances of triggering it will also increase.

For you, this is a chance to see your technology pushed to its limits in large-scale real-world usage. We’re also happy to contribute improvements back to the open source codebase, as we did recently with the 3.4.1 version.

For us, having a closer connection with the maintainers helps us keep production stable and our users happy.

Let me know the best way to move forward. We’d really appreciate the opportunity to collaborate more closely.

[Palanikannan1437]

I think this was related to 1f0b5a4, is it still happening for you after this fix @matteotarantino-algor, I believe this was stopping from hocuspocus from unloading docs from memory due to how process.nextTick works

[matteotarantino-algor]

Hi @Palanikannan1437,

Thanks for your input!
I agree that change could be a root cause. In fact, I upgraded Hocuspocus to the latest version (3.4.3) about ~20 days ago.

The incident I observed happened ~15 days before that upgrade, and it hasn’t happened again in the period between those two dates (it was a rare event anyway), so we can’t empirically conclude that this fix solved it yet: not enough time has passed.

For this reason I’ll keep the issue open for now. If it doesn’t happen again for another 4–6 weeks, I’ll close it and report back that it’s very likely this fix resolved the severe issue we were seeing at scale (roughly ~20k people/day using the platform).

Thanks again for contributing, hopefully I’ll be back with good news!