Design Serialization Schema

Make versioned serialization schemas. Evolve gracefully without breaking consumers.

When Use

• Define new API contract or data interchange format
• Add fields to existing schema without breaking consumers
• Migrate between schema versions
• Pick between schema systems (JSON Schema, Protobuf, Avro)
• Document data validation rules for auto-enforcement

Inputs

• Required: Data model (entity relations, field types, constraints)
• Required: Compat needs (who consumes, how long must old formats read)
• Optional: Existing schema to evolve
• Optional: Perf needs (validation speed, schema registry integration)
• Optional: Target serialization format (JSON, binary, columnar)

Steps

Step 1: Pick Schema System

Got: Schema system picked by ecosystem, perf, evolution needs.

If fail: Unsure? Start with JSON Schema — broadest tooling, layers onto existing JSON APIs.

Step 2: Design Core Schema

JSON Schema example:

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://example.com/schemas/measurement/v1",
  "title": "Measurement",
  "description": "A sensor measurement reading",
  "type": "object",
  "required": ["sensor_id", "value", "unit", "timestamp"],
  "properties": {
    "sensor_id": {
      "type": "string",
      "pattern": "^[a-z]+-[0-9]+$",
      "description": "Unique sensor identifier (lowercase-digits format)"
    },
    "value": {
      "type": "number",
      "description": "Measured value"
    },
    "unit": {
      "type": "string",
      "enum": ["celsius", "fahrenheit", "kelvin", "percent", "ppm"],
      "description": "Unit of measurement"
    },
    "timestamp": {
      "type": "string",
      "format": "date-time",
      "description": "ISO 8601 timestamp with timezone"
    },
    "metadata": {
      "type": "object",
      "additionalProperties": true,
      "description": "Optional key-value metadata"
    }
  },
  "additionalProperties": false
}

Protocol Buffers example:

syntax = "proto3";
package sensors.v1;

import "google/protobuf/timestamp.proto";

// Measurement represents a single sensor reading.
message Measurement {
  string sensor_id = 1;         // Unique sensor identifier
  double value = 2;             // Measured value
  Unit unit = 3;                // Unit of measurement
  google.protobuf.Timestamp timestamp = 4;
  map<string, string> metadata = 5; // Optional key-value metadata
}

enum Unit {
  UNIT_UNSPECIFIED = 0;
  UNIT_CELSIUS = 1;
  UNIT_FAHRENHEIT = 2;
  UNIT_KELVIN = 3;
  UNIT_PERCENT = 4;
  UNIT_PPM = 5;
}

Apache Avro example:

{
  "type": "record",
  "name": "Measurement",
  "namespace": "com.example.sensors",
  "doc": "A sensor measurement reading",
  "fields": [
    {"name": "sensor_id", "type": "string", "doc": "Unique sensor identifier"},
    {"name": "value", "type": "double", "doc": "Measured value"},
    {"name": "unit", "type": {"type": "enum", "name": "Unit", "symbols": ["CELSIUS", "FAHRENHEIT", "KELVIN", "PERCENT", "PPM"]}},
    {"name": "timestamp", "type": {"type": "long", "logicalType": "timestamp-millis"}},
    {"name": "metadata", "type": ["null", {"type": "map", "values": "string"}], "default": null}
  ]
}

Got: Schema self-documenting. Descriptions, constraints, clear types.

If fail: Data model not stable? Mark schema draft, don't publish to registry.

Step 3: Plan Schema Evolution

Compat rules:

Safe evolution:

1. Only add optional fields with sensible defaults
2. Never remove or rename — deprecate instead
3. Version the schema in the identifier (v1, v2)
4. Use a schema registry for binary formats (Confluent Schema Registry for Avro/Protobuf)

Protobuf evolution rules:

// v1 — original
message Measurement {
  string sensor_id = 1;
  double value = 2;
  Unit unit = 3;
}

// v2 — safe evolution
message Measurement {
  string sensor_id = 1;
  double value = 2;
  Unit unit = 3;
  // NEW: added in v2 — old clients ignore this field
  google.protobuf.Timestamp timestamp = 4;
  // DEPRECATED: use sensor_id instead
  reserved 6;
  reserved "old_sensor_name";
}

JSON Schema versioning:

{
  "$id": "https://example.com/schemas/measurement/v2",
  "allOf": [
    {"$ref": "https://example.com/schemas/measurement/v1"},
    {
      "properties": {
        "location": {
          "type": "string",
          "description": "Added in v2: GPS coordinates"
        }
      }
    }
  ]
}

Got: Evolution plan documented. Safe changes vs new versions clear.

If fail: Breaking change unavoidable? Version schema (v1 → v2), keep parallel support during migration.

Step 4: Impl Schema Validation

# JSON Schema validation (Python)
from jsonschema import validate, ValidationError
import json

schema = json.load(open("measurement_v1.json"))

def validate_measurement(data: dict) -> list[str]:
    """Validate a measurement against the schema. Returns list of errors."""
    errors = []
    try:
        validate(instance=data, schema=schema)
    except ValidationError as e:
        errors.append(f"{e.json_path}: {e.message}")
    return errors

# Usage
errors = validate_measurement({"sensor_id": "s-01", "value": "not_a_number"})
# → ["$.value: 'not_a_number' is not of type 'number'"]

// TypeScript with Zod (runtime + compile-time)
import { z } from 'zod';

const MeasurementSchema = z.object({
  sensor_id: z.string().regex(/^[a-z]+-[0-9]+$/),
  value: z.number(),
  unit: z.enum(['celsius', 'fahrenheit', 'kelvin', 'percent', 'ppm']),
  timestamp: z.string().datetime(),
  metadata: z.record(z.string()).optional(),
});

type Measurement = z.infer<typeof MeasurementSchema>;

// Validation
const result = MeasurementSchema.safeParse(inputData);
if (!result.success) {
  console.error(result.error.issues);
}

Got: Validation runs on all incoming data at system boundaries (API endpoints, file ingestion).

If fail: Log validation errors with full payload (redact sensitive fields) for debugging.

Step 5: Document Schema

Make schema doc page:

# Measurement Schema (v1)

## Overview
Represents a single sensor reading with metadata.

## Fields
| Field | Type | Required | Description | Constraints |
|-------|------|----------|-------------|-------------|
| sensor_id | string | Yes | Unique sensor ID | Pattern: `^[a-z]+-[0-9]+$` |
| value | number | Yes | Measured value | Any valid IEEE 754 double |
| unit | enum | Yes | Unit of measurement | One of: celsius, fahrenheit, kelvin, percent, ppm |
| timestamp | string | Yes | Reading time | ISO 8601 with timezone |
| metadata | object | No | Key-value pairs | String keys and values |

## Changelog
| Version | Date | Changes |
|---------|------|---------|
| v1 | 2025-03-01 | Initial schema |

## Compatibility
- **Backwards**: Consumers of v1 will continue to work with future versions
- **Policy**: Only additive, optional field changes between minor versions

Got: Docs auto-generated or stay in sync with schema definition.

If fail: Docs drift from schema? Add CI check validating docs against schema source.

Checks

• Schema uses right system (JSON Schema, Protobuf, Avro)
• All fields have types, descriptions, constraints
• Required vs optional fields explicit
• Evolution strategy documented (safe changes, versioning policy)
• Validation at system boundaries
• Schema versioned with changelog
• Round-trip test: serialize → deserialize → compare, no data loss

Pitfalls

• Over-constraining too early: Strict validation on new schema blocks iteration. Start permissive (additionalProperties: true), tighten later.
• No default values: New required field without default breaks existing data. Always provide defaults for new fields.
• Ignoring null: Many schemas don't handle null/missing cleanly. Be explicit: nullable vs optional.
• Version in payload, not URL: Long-lived data (storage, events) → embed schema version in data itself, not just endpoint URL.
• Enum exhaustiveness: New enum value can crash consumers using exhaustive switch. Document: unknown values handled gracefully.

See Also

• serialize-data-formats — format selection + encoding/decoding
• implement-pharma-serialisation — pharma serialisation (regulatory schemas)
• write-validation-documentation — validation docs for regulated schemas