Semantic model roll back problems

[MorganLi0109]

Description

Our team manages semantic models using TE2 and TE3. One of our semantic models contains a large number of measures. We have observed that the model occasionally experiences rollback issues: some measures disappear unexpectedly, forcing us to rewrite the calculations. This issue occurs roughly once every few months at irregular intervals.

According to some references, this behavior might be related to version conflicts during collaborative saving. However, the problem still occurs even when we confirm that only a single user is modifying the model. As a result, we have to back up the model frequently, yet we are unable to identify the exact cause or scope of the rollbacks.

We have not been able to find any official documentation or community posts regarding this specific issue. We would appreciate your support in confirming whether similar cases exist and if there are any recommended solutions or troubleshooting steps.

Tabular Editor 3 Version

3.20.0

Screenshots

No response

Steps to Reproduce

No response

Expected behavior

No response

Crash Report

No response

Windows Version

No response

[otykier]

Hi @MorganLi0109
That sounds super weird, and not something we have experienced before. A couple of questions to hopefully shed more light on the situation:

• When are measures disappearing? I assume they're not disappearing out of the blue, so it much be when some action is performed against the semantic model, such as a save/deploy operation (from the TE2/TE3 UI or perhaps the TE2 CLI)?
• Are you performing any automated operations on the model, such as running TMSL scripts in a scheduled job, etc.?
• Are you seeing this behavior only on one specific model, or have you experienced it on other models?
• How are you persisting model metadata on disk / source control? Are you using TMDL, .bim or the Tabular Editor "legacy" save-to-folder format?

Any additional information you can provide would be helpful. Thanks!

[MorganLi0109]

Hi @MorganLi0109 That sounds super weird, and not something we have experienced before. A couple of questions to hopefully shed more light on the situation:

• When are measures disappearing? I assume they're not disappearing out of the blue, so it much be when some action is performed against the semantic model, such as a save/deploy operation (from the TE2/TE3 UI or perhaps the TE2 CLI)?
• Are you performing any automated operations on the model, such as running TMSL scripts in a scheduled job, etc.?
• Are you seeing this behavior only on one specific model, or have you experienced it on other models?
• How are you persisting model metadata on disk / source control? Are you using TMDL, .bim or the Tabular Editor "legacy" save-to-folder format?

Any additional information you can provide would be helpful. Thanks!

Thank you for your reply. I will answer your questions one by one in detail below.

• Members of our team alternate between TE2 and TE3 to make modifications to the semantic model, and occasionally use both at the same time. All semantic models have been deployed to workspaces on Power BI Server, and we only perform modify, save, and refresh operations (either for individual tables or the entire semantic model). After creating and saving each new measure, we always verify immediately to ensure the model works correctly. However, rollbacks typically occur several days after a successful save, at which point the newly created measures can no longer be found.

-Refreshes of the semantic model are usually executed via scheduled automatic refresh through the gateway. Sometimes we also refresh individual tables via the XMLA Endpoint, using either SSMS or the built-in refresh functionality in TE3. I believe this process does invoke JSON commands, but I am not certain whether this is equivalent to the TMSL scripts you mentioned.

-This issue almost exclusively occurs on one specific semantic model named Model M. A key characteristic of this model is that all tables are set to Direct Query mode, connecting to other semantic models that contain the actual data. Across our entire data project, all measures are centralized and maintained within Model M for unified management.In short, Model M does not store any data directly; it consists of a set of Direct Query items pointing to other semantic models, along with a large number of measures.

-We frequently back up the semantic model locally in the form of .bim files. Model M itself remains deployed in the production workspace at all times, and we update it using the “Save” function in TE2/TE3. No additional storage or source control methods are used.

Thanks!

[otykier]

This issue almost exclusively occurs on one specific semantic model named Model M. A key characteristic of this model is that all tables are set to Direct Query mode, connecting to other semantic models that contain the actual data. Across our entire data project, all measures are centralized and maintained within Model M for unified management.In short, Model M does not store any data directly; it consists of a set of Direct Query items pointing to other semantic models, along with a large number of measures.

This is a key piece of information! You're referring to Composite models on Power BI Semantic Models and Analysis Services. These models are notoriously tricky, because they manage the lifecycle of certain objects on their own. Specifically, if [Measure A] exists in "Source Model X", and your "Model M" connects to "Source Model X" in DirectQuery mode, you would typically also see [Measure A] get imported to "Model M", with a DAX expression that uses the EXTERNALMEASURE function. If [Measure A] gets renamed or removed in the source model, it would also (eventually) be removed from "Model M" - although unfortunately, Microsofts documentation doesn't make it clear when this "housekeeping" happens.

Now consider the following scenario; you add a measure directly to "Model M" with an expression that may or may not reference measures from one or more source models. I have a hypothesis that if you add this measure to a table that originates from a source model, then Power BI may remove that measure eventually, because it doesn't exist in the source. We have the option to specify child objects that should not be included from the source, but as far as I can tell, we don't have the opposite ability: I.e. ensuring that objects that are added locally, are persisted even when the underlying source is changed.

One thing you could try, is to create a similar composite model entirely through Power BI Desktop, without using Tabular Editor. Then, try to see if you can replicate the measures that you have seen disappear. Perhaps Power BI Desktop adds another type of annotation on them, or something like that, to ensure that they are not removed by the "housekeeping". It's not uncommon for Microsoft to have undocumented features like that. It may also be the case that the result is exactly the same as with Tabular Editor, and that the measure you created through Power BI Desktop eventually disappears. In any case, it may make sense to reach out to Microsoft for further support on this.

[MorganLi0109]

Thank you for your prompt and detailed response!

• As a matter of fact, we have already tried reaching out to Microsoft official technical support and requested their assistance in checking rollback logs and related records. Unfortunately, we have not received any effective explanation or solution from them. They emphasized that Tabular Editor is not covered under their technical support scope.

• I understand the approach to measure creation you described, but our setup is slightly different. In our Model M, we have created a group (essentially an empty table with no fields) to organize all measures. We do not create measures on any of the tables imported from source models. Instead, all measures are created directly within this group in Model M. However, their DAX expressions referred table names from the source semantic models to perform calculations, if this information is useful.

The testing method you suggested sounds very reasonable. However, as I mentioned, the rollback issue occurs only infrequently and unpredictably, so it would be difficult to actually reproduce it reliably.

My team members and I are curious: is there any way to retrieve the version number or snapshot timestamp of a semantic model within Tabular Editor? This would help us track whether a rollback has occurred and to what extent.

Thank you!

[otykier]

Thanks for the added details, @MorganLi0109. So you're saying that you've seen measures disappear from the empty "group" table, which only exists in Model M (and not in any of the source models)? That indeed can't be explained by what I wrote above - unless perhaps something more advanced is going on - i.e. a measure that is defined on the group table references a measure that was imported from a source model. I don't know if they do this kind of impact analysis, i.e. if the referenced measure was removed from the source - would they then also remove measures in the composite model that point to it? But it could at least be an explanation...

As to your question, the Database object does have a Version property which should get incremented on every metadata change to the model. That might be a good thing to look out for:

[MorganLi0109]

Thanks for the added details, @MorganLi0109. So you're saying that you've seen measures disappear from the empty "group" table, which only exists in Model M (and not in any of the source models)? That indeed can't be explained by what I wrote above - unless perhaps something more advanced is going on - i.e. a measure that is defined on the group table references a measure that was imported from a source model. I don't know if they do this kind of impact analysis, i.e. if the referenced measure was removed from the source - would they then also remove measures in the composite model that point to it? But it could at least be an explanation...

As to your question, the Database object does have a Version property which should get incremented on every metadata change to the model. That might be a good thing to look out for:

Super useful !

Thank you for all the information and responses. We will check version frequently to track status in the future and try to figure out the roll back root cause.

Best wishes ~