Support for user defined functions

[anthony-c-martin]

See https://github.com/Azure/bicep/blob/ant/exp/func/src/Bicep.Core.Samples/Files/Functions_LF/main.bicep for a syntax example.

Notes/limitations:

• UDFs in JSON require declaring the return type, so I've modified the spec to require the user to declare the return type.
• User-defined types in UDF params & outputs are unsupported.
• UDFs are currently quite limited - they can't call UDFs, or access outer scoped variables.
• This is currently behind an experimental feature flag ("userDefinedFunctions")

Closes #447
Closes #9239

Microsoft Reviewers: Open in CodeFlow

[dciborow]

Here is an example of something I would replace with a UDF. I use this to generation the connection string for Cassandra DB, which includes each region.

param location string
param secondaryLocations array = []

var secondaryRegions = [for (region, i) in secondaryLocations: {
  locationName: contains(region, 'locationName') ? region.locationName : region
  failoverPriority: contains(region, 'failoverPriority') ? region.failoverPriority : i + 1
  isZoneRedundant: contains(region, 'isZoneRedundant') ? region.isZoneRedundant : isZoneRedundant
}]

var locations = union([
    {
      locationName: location
      failoverPriority: 0
      isZoneRedundant: isZoneRedundant
    }
  ], secondaryRegions)


var unwind = [for location in locations: '${toLower(cassandraName)}-${location.locationName}.cassandra.cosmos.azure.com']
var locationString = replace(substring(string(unwind), 1, length(string(unwind))-2), '"', '')

[jeskew]

• We don't support user-defined types in UDF params & outputs. Should we block this, or attempt to support it?

It would be a small-ish change in the ARM runtime to support user-defined types, and if it were up to me, I would rather put in the effort to close that feature gap than document and field questions about the limitation.

But it did get me thinking: because types can contain symbolic references to other types, I can't think of any reason we would allow user-defined types and still disallow references to variables and parameters. Should we look at removing that restriction on UDFs, too?

[anthony-c-martin]

But it did get me thinking: because types can contain symbolic references to other types, I can't think of any reason we would allow user-defined types and still disallow references to variables and parameters. Should we look at removing that restriction on UDFs, too?

I think so - I just wanted to keep the initial implementation as simple as possible. One complexity is that UDFs overloads parameters(...) to refer to parameters of the UDF and not parameters of the template. We'd most likely need to introduce scoping rules to deal with this.

[jeskew]

Were these new diagnostics intentional?

[majastrz]

BlockUserDefinedFunctionsWithoutExperimentalFeaure

This is certainly a convenient place to put these checks, but I wonder if we need to rename this class. (Blocking UDFs if the experimental feature is disabled isn't really due to JSON limitations.)

[anthony-c-martin]

Good point, it's a convenient place to put any sort of ad hoc check. Will raise an issue to rename it.

[SimonWahlin]

🤩🎉🚀

[ayoussef-insight]

Support to outer scoped variables and include/import feature would be very useful