Shapes: Type-Safe Schemas

Data structures in punchcard are like ordinary collections such as an Array<T> or Map<K, V>, except their type is explicitly defined with a “virtual type-system”, called Shapes.

Shapes are an in-code abstraction for (and agnostic to) data and schema formats such as JSON Schema, Glue Tables, DynamoDB, and (soon) Avro, Protobuf, Parquet, Orc.

class NotificationRecord extends Type({
  key: string,
  count: integer
    .apply(Minimum(0)), // apply a Trait to constrain the value of this integer
  tags: optional(array(string()))
}) {}

const topic = new SNS.Topic(stack, 'Topic', {
  // "shape" of data in the SNS Topic
  shape: NotificationRecord

This Topic’s Shape is now encoded in its type definition:


So, a type-safe interface can also be derived at runtime:

async function publish(notification: NotificationRecord): Promise<AWS.SNS.PublishResponse>;

As opposed to the un-safe opaque types when using the low-level AWS SDK:

async function publish(request: {
  TopicArn: string;
  Message: string | Buffer;
  // etc.

That type-machinery is achieved by mapping a Shape to its Value (the representation at runtime). It should look and feel like an in-memory array:


Validation and Serialization

The framework makes use of Shapes to type-check your code against its schema and safely serialize and deserialize values at runtime. The application code is only concerned with a deserialized and validated value, and so the system is protected from bad data at both compile time and runtime.

For reference, the above Topic’s Shape:

class NotificationRecord extends Type({
  key: string,
  count: integer
  tags: optional(array(string()))
}) {}

Is the same as this JSON Schema:

  "type": "object",
  "properties": {
    "key": {
      "type": "string"
    "count": {
      "type": "integer",
      "maximum": 10
    "timestamp": {
      "type": "string",
      "format": "date-time"
    "tags": {
      "type": "array",
      "items": {
        "type": "string"
  "required": [

Data Types

Shapes are a format-agnostic Data Definition Language (DDL) for JSON, Glue (Hive SQL), (and soon) Avro, Orc and Parquet. For example, the Topic’s Shape maps to this Glue Table Schema:

create table myTable(
  key string,
  count int,
  timestamp timestamp,
  tags array<string>

Below is a table of supported Data Types with their corresponding mappings to different domains:

Shape Runtime JSON Schema Dynamo Glue Usage
BooleanShape boolean boolean BOOL boolean boolean
TimestampShape string string (format: date-time) S timestamp timestamp
BinaryShape Buffer string
(contentEncoding: base64)
B binary binary
StringShape string string S string string
IntegerShape number integer N int integer
BigIntShape number integer N bigint bigint
SmallIntShape number integer N smallint smallint
TinyIntShape number integer N tinyint tinyint
FloatShape number number N float float
DoubleShape number number N double double
ArrayShape<T> Array<T> array L array array(string
SetShape<T> Set<T> array
(uniqueItems: false)
array set(string
MapShape<T> {[K: string]: T} object
(additionalProperties: true)
M map<string, V> map(string
UnknownShape unknown {} (AWS Document Client) Error unknown
AnyShape any {} (AWS Document Client) Error any


Shapes form the foundation on which DSLs are built for interacting with services. This is best demonstrated with AWS DynamoDB.

Next: Dynamic (and safe) DynamoDB DSL