StatChart metricLabel migration: conceptual issues with legendFormat mapping

[abelyakin]

Description

PR #281 introduced by @AntoineThebaud changes to StatChart's Grafana migration to extract metricLabel from legendFormat based on textMode. This issue discusses the conceptual foundation of this approach and its implications.

In Grafana's time-series queries, legendFormat is a template string that controls how series names are rendered in legends and on panels.

• "{{instance}}" - Shows just the instance label
• "Server {{instance}} ({{region}})" - Combines multiple labels in a template
• "Prometheus {{job}}"- Static text with label interpolation

The legendFormat is applied at query time to format the series name based on available labels (if more than one). The series name becomes the visual representation of that time-series.

What is metricLabel in Perses?

In Perses StatChart, metricLabel is a field selector that:

• Selects a specific field/column by name (regex pattern)
• Extracts that field's value from the query result
• Displays that value instead of the calculated numeric result

This maps directly to Grafana's reduceOptions.fields configuration, which specifies which column or field to use for the stat value.

How It Works:

When a query returns data with multiple fields:

Fields: instance, status, version, uptime
Row 1: [server1, up, 1.5.0, 42d]
Row 2: [server2, down, 1.6.0, 15d]

Without metricLabel (or fields in Grafana):

Display the default calculated value (e.g., last value from time-series)

With metricLabel:

For server1 stat: Display "1.5.0" (the version field value)
For server2 stat: Display "1.6.0" (the version field value)

The metricLabel feature exists specifically because Grafana queries can return multiple fields, and sometimes you want to display a non-numeric field value (like version, status, name) as the stat's main value instead of the calculated metric value.

Series Name vs Display Value

legendFormat controls the series name (what appears in legends when showing multiple series)
metricLabel controls the display value (what numeric result is replaced with)
These two features could be used independently or together, but extracting one from the other assumes a 1:1 relationship that doesn't necessarily exist.

Examples

No legendFormat and no metricLabel (thin line on the top is actual series name from very long labels list):

Both legendFormat and metricLabel applied:

The panel.targets Dependency

I found interesting bug where stat chart migration and timeseries migration is relying on panel.targets which is being ommited in grafana.go (perses repo).

The current migration includes:

if #textMode == "name" && (*#panel.targets[0].legendFormat | null) != null {
    metricLabel: strings.Trim(#panel.targets[0].legendFormat, "{}")
}

This depends on panel.targets being present in the migration data. However:

The parent migration script in Perses main repo (grafana.go line ) omits panel.targets when converting panels
This means the condition (*#panel.targets[0].legendFormat | null) will always evaluate to null
The legendFormat extraction never actually occurs in practice. This is especially hard to catch since test input fixture includes targets making it different with actual migration runtime.

Should panel.targets be included in the parent migration?

The same pattern appears in TimeSeries migration with querySettings, creating the same dependency issue. And never migrating querySettings.

Suggestions

Was the legendFormat-to-metricLabel mapping introduced because some dashboards relied on this pattern? What was the actual use case?

I would simplify metricLabel migration relying only on fields from grafana:

#metricLabel: *#panel.options.reduceOptions.fields | null
if #metricLabel != null && #metricLabel != "" {
    metricLabel: #metricLabel
}

And I think we need to keep targets in migration script in order to fix timeseries querySettings migration.

[AntoineThebaud]

This depends on panel.targets being present in the migration data. However:

The parent migration script in Perses main repo (grafana.go line ) omits panel.targets when converting panels
This means the condition (*#panel.targets[0].legendFormat | null) will always evaluate to null
The legendFormat extraction never actually occurs in practice. This is especially hard to catch since test input fixture includes targets making it different with actual migration runtime.

Should panel.targets be included in the parent migration?

The same pattern appears in TimeSeries migration with querySettings, creating the same dependency issue. And never migrating querySettings.

This is actually something I figured out few days ago after spotting the issue with querySettings... I didnt have time to make the fix yet but yes definitely panel.targets shouldn't be omitted. And that's too bad our test framework didn't catch that indeed, 'need to think about a way to improve this

Was the legendFormat-to-metricLabel mapping introduced because some dashboards relied on this pattern? What was the actual use case?

Yes, that's what I meant here. On our side we have this pattern multiple times, looks like it works when you have only a single value returned by the query. Is the current migration logic handling both possibilities causing any issue in your case?

[AntoineThebaud]

Btw for next time, please open all issues on https://github.com/perses/perses (even when related to a specific plugin)

[abelyakin]

Thanks for explaining the use case! I think I've identified the core issue we're trying to use series names as if they're values, which is going to cause problems down the road. And thanks for the heads up about posting in the main repo will make sure to open issues there next time!

So your use case makes total sense: you have legendFormat: "{{version}}" with textMode: "name" in Grafana, and you want that formatted name ("0.1.18") to display in large text as the primary content. That's a totally valid migration requirement.

But here's where I see the issue - we're trying to solve a presentation problem (what to display as primary content) with a data transformation (metricLabel). Even if we need to process legendFormat templates, stuffing the result into metricLabel conflates "series name" with "calculated value".

I think what we actually need is Perses to support textMode (or call it displayMode to be clearer). Something like:

export interface StatChartOptions {
  displayMode?: 'auto' | 'value' | 'value-and-name' | 'name';
}

Then in the rendering code, when displayMode === 'name', we just display seriesData.name in the large font instead of the value. The data is already there and formatted - we just need to show it in the right place.
Migration becomes also super simple.

Does this make sense? The use case is valid, I just think we're solving it at the wrong layer. We need a display option, not a data transformation.

If this approach sounds reasonable, I'd be happy to work on a PR that implements the displayMode property. Would love to hear your thoughts first!

[AntoineThebaud]

I'm fine with your proposal, I'm just wondering what the Metric label field that you added with perses/perses#2574 will be used for in case of e.g displayMode: value. Would it be simply useless in this case?

[AntoineThebaud]

The parent migration script in Perses main repo (grafana.go line ) omits panel.targets when converting panels

FYI tackled by perses/perses#3726