Skip to content

Quickstart

Jump to bottom
Matthieu Monsch edited this page Jan 22, 2017 · 44 revisions

Types

One of the main features provided by Avro is a way to encode (serialize) data. Once data is encoded, it can be stored in a file or a database, sent across the internet to be decoded on another computer, etc.

Many other encodings exist. JSON for example is very commonly used from JavaScript: it's built-in (via JSON.parse and JSON.stringify), reasonably fast, and produces human-readable encodings.

> pet = {kind: 'DOG', name: 'Beethoven', age: 4};
> str = JSON.stringify(pet);
'{"kind":"DOG","name":"Beethoven","age":4}' // Encoded data (still readable).
> str.length
41 // Number of bytes in the encoding.

JSON isn't always the most adequate encoding though. It produces relatively large encodings since keys are repeated in the output (kind, name, and age above). It also doesn't enforce any properties on the data, so any validation has to be done separately (other encodings with the same schema-less property include xml, MessagePack).

Avro types provide an alternate serialization mechanism, with a different set of properties:

  • Schema-aware: each type is tied to a particular data structure and will validate that any encoded data matches this structure.
  • Compact: Avro's binary encoding isn't meant to be human-readable, it is instead optimized for size and speed. Depending on the data, avsc can be an order of magnitude faster and smaller than JSON.

What is a Type?

A type is an JavaScript object which knows how to decode and encode a "family" of values. Examples of supported families include:

  • All strings.
  • All arrays of numbers.
  • All Buffers of length 4.
  • All objects with an integer id property and string name property.

By default, Avro uses [schemas][avro-schemas] to define which family a type supports. These schemas are written in JSON-format (the same encoding described earlier).

> avro = require('avsc');
> inferredType = avro.Type.forValue(pet);
> inferredType.schema();
{ type: 'record', // "Record" is Avro parlance for "structured object".
  fields:
   [ { name: 'kind', type: 'string' }, // Each field corresponds to a property.
     { name: 'name', type: 'string' },
     { name: 'age', type: 'int' } ] }

TODO...

> buf = inferredType.toBuffer(pet);
> buf.length
15 // 60% smaller than JSON!
> inferredType.isValid({kind: 'CAT', name: 'Garfield', age: 5.2});
false // The age isn't an integer.
> inferredType.isValid({name: 'Mozart', age: 3});
false // The kind field is missing.
> inferredType.isValid({kind: 'PIG', name: 'Babe', age: 2});
true // All fields match.
> exactType = avro.Type.forSchema({
...   type: 'record',
...   fields: [
...     {name: 'kind', type: {type: 'enum', symbols: ['CAT', 'DOG']}},
...     {name: 'name', type: 'string'},
...     {name: 'age', type: 'int'}
...   ]
... });
> buf = exactType.toBuffer(pet);
> buf.length
12 // 70% smaller than JSON!
> exactType.isValid({kind: 'PIG', name: 'Babe', age: 2});
false // The pig kind wasn't defined in our enum.
> exactType.isValid({kind: 'DOG', name: 'Lassie', age: 5});
true // But dog was.

Services

As well as defining an encoding, Avro defines a standard way of executing remote procedure calls (RPCs), exposed via services. By leveraging these services, we can implement portable and "type-safe" APIs:

  • Clients and servers can be implemented once and reused over many different transports (in-memory, TCP, HTTP, etc.).
  • All data flowing through the API is automatically validated: call arguments and return values are guaranteed to match the types specified in the API.

In this section, we'll walk through an example of building a simple link management service similar to bitly.

Defining a Service

The first step to creating a service is to define its protocol, describing the available API calls and their signature. There are a couple ways of doing so; we can write the JSON declaration directly, or we can use Avro's IDL syntax (which can then be compiled to JSON). The latter is typically more convenient so we will use this here.

/** A simple service to shorten URLs. */
protocol LinkService {

  /** Map a URL to an alias, throwing an error if it already exists. */
  null createAlias(string alias, string url);

  /** Expand an alias, returning null if the alias doesn't exist. */
  union { null, string } expandAlias(string alias);
}

With the above spec saved to a file, say LinkService.avdl, we can instantiate the corresponding service as follows:

// We first compile the IDL specification into a JSON protocol.
avro.assembleProtocol('./LinkService.avdl', function (err, protocol) {
  // From which we can create our service.
  const service = avro.Service.forProtocol(protocol);
});

The service object can then be used generate clients and servers, as described in the following sections.

Server implementation

So far, we haven't said anything about how our API's responses will be computed. This is where servers come in, servers provide the logic powering our API.

For each call declared in the protocol (createAlias and expandAlias above), servers expose a similarly named handler (onCreateAlias and onExpandAlias) with the same signature:

const urlCache = new Map(); // We'll use an in-memory map to store links.

// We instantiate a server corresponding to our API and implement both calls.
const server = service.createServer()
  .onCreateAlias(function (alias, url, cb) {
    if (urlCache.has(alias)) {
      cb(new Error('alias already exists'));
    } else {
      urlCache.set(alias, url); // Add the mapping to the cache.
      cb();
    }
  })
  .onExpandAlias(function (alias, cb) {
    cb(null, urlCache.get(alias));
  });

Notice that no part of the above implementation is coupled to a particular communication scheme (e.g. HTTP, TCP, AMQP): the code we wrote is transport-agnostic.

Calling our service

The simplest way to call our service is use an in-memory client, passing in our server above as option to service.createClient:

const client = service.createClient({server});

// We first send a request to create an alias.
client.createAlias('hn', 'https://news.ycombinator.com/', function (err) {
  // Which we can now expand.
  client.expandAlias('hn', function (err, url) {
    console.log(`hn is currently aliased to ${url}`);
  });
});

The above is handy for local testing or quick debugging. More interesting perhaps is the ability to communicate with our server over any binary streams, for example TCP sockets:

const net = require('net');

// Set up the server to listen to incoming connections on port 24950.
net.createServer()
  .on('connection', function (con) { server.createChannel(con); })
  .listen(24950);

// And create a matching client:
const client = service.createClient({transport: net.connect(24950)});

Note that RPC calls messages are always sent asynchronously and in parallel: requests do not block each other. Furthermore, responses are available as soon as they are received from the server; the client keeps track of which calls are pending and triggers the right callbacks as responses come back.

Both above transports (in-memory and TCP) have the additional property of being stateful: each connection can be used to exchange multiple messages, making them particularly efficient (avoiding the overhead of handshakes). These aren't the only kind though, it is possible to exchange messages over stateless connections, for example HTTP:

const http = require('http');

// Each HTTP request/response will correspond to a single API call.
http.createServer()
  .on('request', function (req, res) {
    server.createChannel(function (cb) { cb(null, res); return req; });
  })
  .listen(8080);

// Similarly, an HTTP client:
const client = service.createClient({transport: function (cb) {
  return http.request({method: 'POST', port: 8080})
    .on('response', function (res) { cb(null, err); })
    .on('error', cb);
}});

Next steps

The API documentation provides a comprehensive list of available functions and their options. The Advanced usage section goes through a few more examples of advanced functionality.

Clone this wiki locally