This is a library for working with Avro schemas and the Confluent® Schema Registry, primarily focused on working with Kafka streams. It relies on erlavro for encoding and decoding data and confluent_schema_registry to look up schemas using the Schema Registry API.
Its primary value is that it caches schemas for performance and to allow programs to work independently of the Schema Registry being available. It also has a consistent set of functions to manage schema tags, look up schemas from the Schema Registry or files, and encode/decode data.
Much thanks to Klarna for Avlizer, which provides similar functionality to this library in Erlang, erlavro for Avro, and brod for dealing with Kafka.
Add the package to your list of dependencies in mix.exs:
def deps do
[{:avro_schema, "~> 0.1.0"}]
endThen run mix deps.get to fetch the new dependency.
Documentation is on HexDocs.
To generate a local copy, run mix docs.
Add the cache GenServer to your application's supervision tree:
def start(_type, _args) do
cache_dir = Application.get_env(:yourapp, :cache_dir, "/tmp")
children = [
{AvroSchema, [cache_dir: cache_dir]},
]
opts = [strategy: :one_for_one, name: LogElasticsearch.Supervisor]
Supervisor.start_link(children, opts)
endWhen using Kafka, producers and consumers are separated, and schemas may evolve over time. It is common for producers to tag data indicating the schema that was used to encode it. Consumers can then look up the corresponding schema version and use it decode the data.
This library supports two tagging formats, Confluent wire format, and Avro single object encoding.
With the Confluent wire format, Avro binary encoded objects are prefixed with a five-byte tag.
The first byte indicates the Confluent serialization format version number, currently always 0. The following four bytes encode the integer schema ID as returned from the Schema Registry in network byte order.
When used without a schema registry, it's common to prefix binary data with a hash of the schema that created it. In the past, that might be something like MD5.
The Avro "Single-object encoding" formalizes this, prefixing Avro binary data with a two-byte marker, C3 01, to show that the message is Avro and uses this single-record format (version 1). That is followed by the 8-byte little-endian CRC-64-AVRO fingerprint of the object's schema.
The CRC64 algorithm is uncommon, but used because it is relatively short,
while still being good enough to detect collisions. The fingerprint function is
implemented in fingerprint_schema/1.
In a relatively static system, it's not too hard to exchange schema files between producers and consumers. When things are changing more frequently, it can be difficult to keep files up to date. It's also easy for insignificant differences such as whitespace to result in a different schema hashes.
The Schema Registry solves this by providing a centralized service which producers and consumers can call to get a unique identifier for a schema version. Producers register a schema with the service and get an id. Consumers look up the id to get the schema. The Schema Registry also does validation on new schemas to ensure that they meet the backwards compatibility policy for the organization.
The disadvantage of the Schema Registry is that it can be a single point of failure. Different schema registries will in general assign a different numeric id to the same schema.
This library provides functions to register schemas with the Schema Registry and look them up by id. It caches the results in RAM (ETS) for performance, and optionally also on disk (DETS). This gives good performance and allows programs to work without needing to communicate with the Schema Registry. Once read, the numeric IDs never change, so it's safe to cache them indefinitely.
The library also has support for managing schemas from files. It can add files to the cache by fingerprint, registering the same schema under multiple fingerprints, i.e. the raw JSON, a version in Parsing Canonical Form and with whitespace stripped out. You can also manually register aliases for the name and fingerprint to handle legacy data.
A Kafka producer program needs to be able to encode the data with an Avro schema and tag it with the schema ID or fingerprint. It normally stores the schema in the code or reads it from a file.
It can then call the Schema Registry to get the ID matching the Avro schema and subject:
iex> schema_json = "{\"name\":\"test\",\"type\":\"record\",\"fields\":[{\"name\":\"field1\",\"type\":\"string\"},{\"name\":\"field2\",\"type\":\"int\"}]}"
iex> subject = "test"
iex> {:ok, ref} = AvroSchema.register_schema(subject, schema_json)
{:ok, 21}The subject is a name which identifies the type of data. Multiple Kafka topics
may carry the same data, so they have the same subject. It is normally the Avro
"full name", e.g.
com.example.X. In an Avro schema, it is the "name" field.
If the schema has already been registered, then the Schema Registry will return the current id. If you are registering a new version of the schema, then the Schema Registry will first check if it is compatible with the old one. Depending on the compatibility rules, it may reject the schema.
The producer next needs to get an encoder for the schema.
The encoder is a function that takes Avro key/value data and encodes it to a binary.
iex> {:ok, encoder} = AvroSchema.make_encoder(schema_json)
{:ok, #Function<2.110795165/1 in :avro.make_simple_encoder/2>}
Next we encode some data:
iex> data = %{field1: "hello", field2: 100}
iex> encoded = AvroSchema.encode(data, encoder)
[['\n', "hello"], [200, 1]]Finally, we tag the data:
iex> tagged_confluent = AvroSchema.tag(encoded, 21)
[<<0, 0, 0, 0, 21>>, [['\n', "hello"], [200, 1]]]If you are using files, the process is similar. First create a fingerprint for the schema:
iex> fp = AvroSchema.fingerprint_schema(schema_json)
<<172, 194, 58, 14, 16, 237, 158, 12>>Next tag the data:
iex> tagged_avro = AvroSchema.tag(encoded, fp)
[
<<195, 1>>,
<<172, 194, 58, 14, 16, 237, 158, 12>>,
[['\n', "hello"], [200, 1]]
]Now you can send the data to Kafka.
The process for a consumer is similar.
Receive the data and get the registration id in Confluent format:
iex> tagged_confluent = IO.iodata_to_binary(AvroSchema.tag(encoded, 21))
<<0, 0, 0, 0, 21, 10, 104, 101, 108, 108, 111, 200, 1>>
iex> {:ok, {{:confluent, regid}, bin}} = AvroSchema.untag(tagged_confluent)
{:ok, {{:confluent, 21}, <<10, 104, 101, 108, 108, 111, 200, 1>>}}Get the schema from the Schema Registry:
iex> {:ok, schema} = AvroSchema.get_schema(regid)
{:ok,
{:avro_record_type, "test", "", "", [],
[
{:avro_record_field, "field1", "", {:avro_primitive_type, "string", []},
:undefined, :ascending, []},
{:avro_record_field, "field2", "", {:avro_primitive_type, "int", []},
:undefined, :ascending, []}
], "test", []}}Create a decoder and decode the data:
iex> {:ok, decoder} = AvroSchema.make_decoder(schema)
{:ok, #Function<4.110795165/1 in :avro.make_simple_decoder/2>}
iex> decoded = AvroSchema.decode(bin, decoder)
%{"field1" => "hello", "field2" => 100}
The process is similar with a fingerprint. In this case, we get the schema from files and register it in the cache using the schema name and fingerprints. There is more than one fingerprint because we register it with the raw schema from the file and the normalized JSON for better interop.
iex> {:ok, files} = AvroSchema.get_schema_files("test/schemas")
{:ok, ["test/schemas/test.avsc"]}
iex> for file <- files, do: AvroSchema.cache_schema_file(file)
[
ok: [
{"test", <<172, 194, 58, 14, 16, 237, 158, 12>>},
{"test", <<194, 132, 80, 199, 36, 146, 103, 147>>}
]
]To decode, separate the fingerprint from the data:
iex> tagged_avro = IO.iodata_to_binary(AvroSchema.tag(encoded, fp))
<<195, 1, 172, 194, 58, 14, 16, 237, 158, 12, 10, 104, 101, 108, 108, 111, 200,
1>>
iex> {:ok, {{:avro, fp}, bin}} = AvroSchema.untag(tagged_avro)
{:ok, {{:avro, <<172, 194, 58, 14, 16, 237, 158, 12>>},
<<10, 104, 101, 108, 108, 111, 200, 1>>}}Get the decoder and decode the data:
iex> {:ok, schema} = AvroSchema.get_schema({"test", fp})
{:ok,
{:avro_record_type, "test", "", "", [],
[
{:avro_record_field, "field1", "", {:avro_primitive_type, "string", []},
:undefined, :ascending, []},
{:avro_record_field, "field2", "", {:avro_primitive_type, "int", []},
:undefined, :ascending, []}
], "test", []}}Decoding works the same as with the Schema Registry:
iex> {:ok, decoder} = AvroSchema.make_decoder(schema)
{:ok, #Function<4.110795165/1 in :avro.make_simple_decoder/2>}
iex> decoded = AvroSchema.decode(bin, decoder)
%{"field1" => "hello", "field2" => 100}
For best performance, save the encoder or decoder in your process state to avoid the overhead of looking it up for each message.
An in-memory ETS cache maps the integer registry ID or name + fingerprint to the corresponding schema and decoder. It also allows lookups using name and fingerprint as a key.
The fingerprint is CRC64 by default. You can also register a name with your own fingerprint.
This library also allows consumers to look up the schema on demand from the Schema Registry using the name + fingerprint as the registry subject name.
This library can optionally persist the cache data on disk using DETS, allowing programs to work without continuous access to the Schema Registry.
Programs which use Kafka may process high message volumes, so efficiency is important. They generally use multiple processes, typically one per topic partition or more. On startup, each process may simultaneously attempt to look up schemas.
The cache lookup runs in the caller's process, so it can run in parallel. If there is a cache miss, then it calls the GenServer to update the cache. This has the effect of serializing requests, ensuring that only one runs at a time. See https://www.cogini.com/blog/avoiding-genserver-bottlenecks/ for discussion.
In order to improve interoperability, the schema should be put into standard form.
It might also call the schema registry to get the schema for a given subject:
iex> ConfluentSchemaRegistry.get_schema(client, "test")
{:ok,
%{
"id" => 21,
"schema" => "{\"type\":\"record\",\"name\":\"test\",\"fields\":[{\"name\":\"field1\",\"type\":\"string\"},{\"name\":\"field2\",\"type\":\"int\"}]}",
"subject" => "test",
"version" => 13
}}Avro timestamps are in Unix format with microsecond precision:
iex> datetime = DateTime.utc_now()
~U[2019-11-08 09:09:01.055742Z]
iex> timestamp = AvroSchema.to_timestamp(datetime)
1573204141055742
iex> datetime = AvroSchema.to_datetime(timestamp)
~U[2019-11-08 09:09:01.055742Z]