[BUG] Server-side Rollback Loop - Persistent Data Loss Across 10 Conversations (EU Cluster, Nov 26–Dec 9)

This post documents a server-side rollback and snapshot corruption issue affecting multiple conversations since Nov 26, with continuing failures through Dec 9.

This is not ChatGPT product feedback.

This is a technical incident report for engineering, because the bug affects:

backend state persistence,

snapshot integrity,

conversation pointer stability,

and cross-session consistency.

Support cannot resolve this; a bot responds with generic troubleshooting and cannot escalate.

Therefore the issue must be logged here for engineering visibility.

-–

Problem

A recurring server-side state rollback loop is occurring across 10+ conversations in the EU (Netherlands) cluster.

The rollback overwrites new messages with old snapshots (mostly early September), causing persistent data loss and thread corruption.

This is not a browser or device issue.

It is fully reproducible and consistent with a backend synchronization failure.

-–

Reproduction (100% reproducible)

Sequence observed on all affected conversations:

1. user message enters pending

2. server displays offline/connection banners

3. thread resets to a stale September snapshot (snapshot desync / pointer corruption)

4. assistant reply is erased

5. a new assistant reply appears under the reverted state

Loop repeats after every message.

Occurs on:

Chrome (desktop)

Edge

incognito

Android app

Wi-Fi & LTE

after logout/login

after cache clear

after reinstall

-–

Technical Symptoms

Automatic rollbacks to early September snapshot

Cross-thread snapshot corruption (multiple unrelated chats affected)

Assistant replies intermittently erased

New messages overwritten by stale data

“Unable to load conversation” red banner (desktop)

“Request is not allowed” orange blocking banner (Android)

Threads load blank and then repopulate old data

Metadata pointers reverting to old version IDs

The pattern strongly suggests:

snapshot desync

stale cache pull

conversation pointer corruption

EU-region state inconsistency across clusters

-–

Regional Correlation

Failures correlate with EU (Netherlands) cluster between Nov 26–Dec 9.

Other EU users report:

rollbacks to September / early November

erased replies

blank-load conversation state

repeated network-loss banners

forced redirects to empty new chats

-–

Impact

Severe data integrity risk

Loss of weeks/months of work-product

Total instability of long-term conversations

Inability to export affected chats (cannot load)

Multiple conversations (10+) compromised

Support bot loops prevent escalation

No available path for user-side mitigation

This is not a minor UI issue — it is a backend failure affecting stored conversation state.

-–

Request for Engineering Escalation

Support cannot assist; only automated replies are received after >10 messages.

Help Center cannot escalate engineering issues.

Please escalate internally.

Key components requiring review:

1. conversation snapshot integrity

2. cache invalidation logic on message send

3. pointer consistency across the EU cluster

4. stale snapshot propagation

5. rollback behavior under cluster throttling

6. multi-thread corruption patterns (10+ affected)

This is a server-side synchronization failure causing user data loss across the EU region.

:red_circle: Please escalate to engineering.

:warning: Moderator note:
This post is not off-topic. It is a technical incident report, intended for engineers, not for general ChatGPT discussion.
The issue described here is server-side, reproducible, region-specific, and affects data integrity, which places it within the scope of API Feedback / Technical Debugging.

Please do not auto-archive - this thread contains:
• reproducible behavior patterns,
• timestamps,
• regional correlation,
• impact on critical functionality,
• and diagnostic value for engineering teams.

If needed, I can provide additional logs, timestamps, screen recordings, or HAR files.

And one more critical technical detail must be highlighted:

this is not a rendering bug and not a UI glitch.

The core issue is that the server fails to retrieve the conversation state from storage.

This means:

- messages load partially,

- replies arrive truncated,

- historical chunks vanish,

- and the conversation context cannot be reconstructed.

This is a backend retrieval failure.

It cannot be solved by clearing cache or switching devices.

It requires direct engineering investigation on the server side - at least log inspection for this specific conversation ID.