The World of Splectrum

Mycelium Message

The message is the tree in motion. When the identifier structure becomes active — when it travels — it goes into a Kafka record. The Kafka record does not introduce new structure. It introduces directionality. The tree is symmetric. The Kafka record says: this part is the question, this part is the answer. Same structure, now polarised.

Headers and Value

Headers — the piece of tree that describes what you want done. Context, intent, operator arguments. Tree structure serving as input.

Value — the piece of tree that holds the result. Starts empty, fills during execution. Tree structure serving as output.

Every message is an operator invocation. Pure data transfer is not a special case. It is noop — an operator with fully qualified identity spl.splectrum.operator.noop, a specific contract (value passes through unchanged), and args: null as its legitimate interface. The RPC server has exactly one code path.

The Onion — Nested Records

Every message is a nested Kafka record. Each layer has the same two sides: headers and value. Peel a layer — same structure underneath.

The outer layer is the execution envelope. The inner layer is the protocol operator. Both are Kafka records. Same shape at every nesting level.

Headers as Property Bag

Headers is a property bag — a context with an AVRO schema assigning namespace to short names within it. Properties inside headers are just properties. Fully qualified or not, based on the schema.

The headers schema contains:

record — a property whose schema identifies the operator and its arguments. record is itself a property bag. Its schema gives logicalType and args their namespace.

Additional properties alongside record carry execution context — tracing, routing, processing metadata. These get their namespace from the headers schema.

headers: {
  record: {
    logicalType: 'spl.splectrum.operator.noop',
    args: null
  },
  spl.trace.id: ...,
  spl.mycelium.context: ...
}

The record property is the what — the operation. Everything else in headers is the how — execution context. Two concerns, cleanly separated within the same property bag.

Dispatch

Dispatch reads one path: headers.record.logicalType. That is the routing key. The single dispatch mechanism for every message.

Schema as Namespace Authority

The namespace resolution is uniform at every level of the message:

Kafka record has an AVRO schema. That schema gives headers its namespace. headers becomes spl.mycelium.message.headers.
Headers is a property bag with an AVRO schema. That schema gives record its namespace. record becomes spl.mycelium.operator.record.
Record is a property bag with an AVRO schema. That schema gives logicalType and args their namespace.

Each schema is the namespace authority for the names it contains. No level is special. No level resolves differently. One mechanism applied recursively.

Concrete Message

An RPC server execution envelope wrapping a get request:

spl.mycelium.process.execute.exec {
  headers: {
    record: {
      logicalType: 'spl.mycelium.process.execute',
      args: { mode: 'sync' }
    }
  },
  key: "/blog/submissions",
  value: xpath.data.uri.get {
    headers: {
      record: {
        logicalType: 'spl.splectrum.operator.get',
        args: { filter: ... }
      }
    },
    key: "/blog/submissions",
    value: <output>
  }
}

The outer layer is the execution envelope — its headers identify the execution operator and mode. Its value carries the inner operator as a complete computation record.

The inner layer is the protocol operation — its headers identify the data access operator and args. Its value starts empty and fills with the retrieved data.

Processing Pipeline

arrive   -> headers: args          value: empty
validate -> headers: args          value: validated input + output schema
execute  -> headers: args          value: output
error    -> headers: args + error  value: partial/empty

Same pattern at both levels. The pipeline is uniform because the entry point is always the same shape. Validation is: does the logical type resolve to a known operator schema? Do the args conform?

Request and Response

The response returns the request enriched with the result. No separate response message. The same message at every stage, just more resolved.

get — request: args in headers, value empty. Response: value filled with retrieved data.

put — request: args in headers, value carries data to write. Response: value confirmed.

delete — request: args in headers, value empty. Response: value carries what was removed.

noop — request: args null, value carries data payload. Response: value passed through unchanged.

Design Properties

Self-contained — pick the message up at any stage, read what was asked (headers) and how far it got (value).

Enrichment not replacement — the message accumulates. Request context preserved.

Errors don’t break the shape — error conditions add metadata to headers. No separate error envelope.

Same shape sync and async — the echo-back pattern works whether response returns immediately or via a queue.

Uniform dispatch — every message is an operator invocation. No branching on message category. The logical type in headers.record is the single dispatch mechanism.