A Detailed Introduction to the Avro Data Format

By SQream

3.21.2022 twitter linkedin facebook


In this article, we’re going to take a look at Apache Avro Data Format, which is a data serialization system that converts an object in memory to a stream of bytes.

Visit SQream documentation to learn how to insert data from Avro into SQream


Let’s start with an example. Say we have a software application dealing with orders. Each order has data like the ID, date, customer ID, type, etc. This data is loaded in memory as an object/struct as per the programming language. Below are some examples of when we might want to serialize this object:

  • Storing the object in a file: Suppose we want to share a list of orders with our suppliers. Without giving them access to our system, we may share a file containing orders with them. To write the orders from memory to a file would require us to serialize each order into a stream of characters/bytes. Another scenario for serializing to a file is when the orders are too numerous to all fit into memory, so the application stores them temporarily to a file and only loads those for the current page.
  • Sending over a network connection: Suppose we want to share the orders with another system in our company—say the delivery system, which runs its own web service. To send an order object to that service over the network, we would have to send it as a stream of bytes (i.e., serialize it).

And what will we do with the serialized data? At some later point, we’ll need to deserialize it or read it back from the stream of bytes to an object. This implies that the serialization needs to have a particular format so that we are able to deserialize it. There are already some formats in use for this (each with their own drawback):

  • CSV: No support for complex data types like structs, arrays, etc.
  • XML: Verbose
  • JSON: Still verbose, as field names and string representation of all types take up space

Why Avro Data Format?

Why use Avro when there are already other formats present? There are quite a few reasons. Avro:

  • Has a compact and fast binary data format
  • Is a documented format that makes use of schemas for correctness
  • Has rich data types (e.g., arrays, maps, enumerations, objects, etc.)
  • Provides a container file format that is splitable into chunks for distributed processing and contains the schema along with the data
  • Ships with integration with popular languages like Python, Java, C++, etc. and can work with map/dictionary-like objects to represent records; does not need to have strongly typed classes generated from schemas
  • Supports schema evolution (i.e., schema changes); for example, if a new field is added later to the schema, reading serialized data without the field won’t cause a problem

Avro Data Format

The Avro format consists of two parts: the schema and the serialized data. Both need to be present when serializing/writing or deserializing/reading data.

Avro Schema

The schema specifies the structure of the serialized format. It defines the fields that the serialized format contains, their data types (whether optional or not), default values, etc. The schema itself is in JSON format. A schema can be any of the below:

  • A JSON string containing a type name—e.g., “string” or “int
  • Or a JSON object that defines a new record type in the format { “type” : “typeName”, attributes }
  • Or a JSON array specifying a union of multiple types—e.g., [ “null“, “string“, “int“] is a union of three types and means that the field using this type can be either null OR a string OR an int

The built-in data types are as follows:

  • boolean: a binary value
  • int: 32-bit signed integer
  • long: 64-bit signed integer
  • float: single precision (32-bit) IEEE 754 floating-point number
  • double: double precision (64-bit) IEEE 754 floating-point number
  • bytes: sequence of 8-bit unsigned bytes
  • string: Unicode character sequence
  • record: a type that contains named fields, each having a type, required, default, etc.
  • enum: an enumeration of given values; other values aren’t allowed
  • array: an array of items
  • map: a map/dictionary of key-value pairs, where the keys must be strings
  • fixed: a fixed number of bytes
  • union: an array containing type names; the resulting type can be any one of these types

The complex types each have different attributes (e.g., record has fields, array has items). There are some attributes common to them though.

  • The doc attribute is used to document what the type is for.
  • The namespace attribute helps to avoid name collisions among types with the same name. For example, there could be multiple types named “document.” Each would have a different namespace (e.g., “com.company1” and “org.org2”). And we would then refer to a type by its full name to get the correct one (e.g., “com.company1.document,” instead of just “document”).
  • The default attribute specifies the default value for a type.

The above is a simplified explanation of the schema. The full Avro format specification can be found here.


As an example, the schema for the order data that we saw above would be as follows:


  "namespace" : "org.example",

  "type": "record", 

  "name": "order", 

   "fields" : [

    {"name": "orderId", "type": "long"}, 

    {"name": "orderDate", "type": "int", "logicalType" : "date"}, 

    {"name": "customerId", "type": "string"}, 

    {"name": "orderType", "type":{ "type" : "enum", "name" : "orderType", 

         "symbols" : ["NORMAL", "INTERNAL", "CUSTOM"] }},

     {"name": "notes", "type": [ "null", "string"], "default" : null }



Here are some things to note about the above example:

  • The date type is not supported out of the box and is represented as int that is the number of days from the start of the epoch. An extra attribute, logicalType, is supported so that applications handling this data may process/convert it further.
  • When we use a complex type in a field, it needs to use the object syntax: { type:XXX name:YYY }.
  • The “notes” string field can have a null value. We say this by using the union of null and string for its type. It also has a default value of null. The default value will be used when we’re reading the serialized data if this field is missing from the data.

Avro Data

The serialized data can be created in three formats/encodings. See this reference for details.

  1. JSON format: This is the verbose, human-readable format, useful for debugging. But it’s not performant for parsing and takes up more space.
  2. Binary format: This is the compact, performant default format. Strings are encoded as UTF-8, the rest as binary. As a simplified example, consider how 2^16 = 65536 will take up five characters to encode as a string, but only two bytes as binary. Thus, the nonstring types are serialized to an optimized binary format. One important point is that this format doesn’t store the field names/IDs, so the schema used to write an object is always required when reading back the serialized data in order to identify the fields.
  3. Single object encoding: In addition to the binary format, this also contains a fingerprint or hash of the schema with which it was created. This is useful for scenarios where the schema has changed a few times. The serialized data that we’re reading may belong to any of those schemas, and we want to know which data/message belongs to which schema. In this case, the fingerprint/hash can be used to find the matching schema. This could also be achieved by introducing a custom schema field in each message.

Object Container Files

Avro also provides a format for storing Avro records as a file. The records are stored in the binary format. The schema used to write the records is also included in the file. The records are stored in blocks and can be compressed. Also, the blocks make the file easier to split, which is useful for distributed processing like Map-Reduce. This format is supported by many tools/frameworks like Hadoop, Spark, Pig, and Hive.

Writing to Avro

Many popular languages have APIs for working with Avro. Here, we’re going to see an example of writing to Avro in Python. Note how the objects we’re writing are just generic dictionaries and not classes with strongly typed methods like getOrderId(). We need to convert the date into an int containing the number of days since epoch, as Avro uses this for dates. At the end, we’ll have printed the number for bytes used for encoding the objects.

Here we’re using the DatumWriter to write individual records; we’re not using the Avro object container file format that also contains the schema. This will be the case when we’re writing individual items to a message queue or sending them as params to a service. The DataFileWriter should be used if we want to write an Avro object container file.

from avro.io import DatumWriter, DatumReader, BinaryEncoder, BinaryDecoder

import avro.schema

from io import BytesIO

from datetime import datetime

schemaStr = ”’

{ "type": "record", 

"name": "order", 

"fields" : [ 

    {"name": "orderId", "type": "long"}, 

    {"name": "orderDate", "type": "int", "logicalType" : "date"}, 

    {"name": "customerId", "type": "string"}, 

    {"name": "orderType", "type": { "type" : "enum", "name" : "orderType", "symbols" : ["NORMAL", "INTERNAL", "CUSTOM"] } },

    {"name": "notes", "type": [ "null", "string"], "default" : null }



def daysSinceEpoch( dateStr) :

    inpDate = datetime.strptime( dateStr, '%Y-%m-%d')

    epochDate = datetime.utcfromtimestamp(0)

    return( inpDate - epochDate ).days

schema = avro.schema.parse( schemaStr)

wbuff = BytesIO()

avroEncoder = BinaryEncoder(wbuff)

avroWriter = DatumWriter(schema)

avroWriter.write( { "orderId":11, "orderDate":daysSinceEpoch( "2022-01-27"), "customerId":"CST223", 

    "orderType":"INTERNAL" }, avroEncoder)

avroWriter.write( { "orderId":12, "orderDate":daysSinceEpoch( "2022-02-27"), "customerId":"MST001", 

    "orderType":"NORMAL", "notes" : "Urgent" }, avroEncoder)

print( wbuff.tell())

Reading From Avro Data Format

Writing to Avro is straightforward. We have the schema and data to be written. Reading Avro, however, may involve two schemas: the writer schema, which was used to write the message, and the reader schema that is going to read the message. This is necessary if there are different schema versions (maybe fields have been added or removed), and so the writer’s schema is different from the reader’s schema. It could also happen that the reader needs to read only a subset of the fields written. Below is the code in Python, appended to the one we already have for the writer. In this case, we’ve used a different schema for the reader, which contains only the orderId and a new field, extraInfo. Since extraInfo wasn’t in the written data, its default value will be used by the reader.

schemaStrRdr = ''' { "type": "record",  "name": "order",  "fields" : [      {"name": "orderId", "type": "long"},      {"name": "extraInfo", "type": "string", "default" : "NA" }     ] } ''' schemaRdr = avro.schema.parse( schemaStrRdr) rbuff = BytesIO(wbuff.getvalue()) avroDecoder = BinaryDecoder(rbuff) avroReader = DatumReader(schema, schemaRdr) # Use both schemas msg = None while True :     try :         msg = avroReader.read(avroDecoder)     except Exception as ex:          msg = None         print ( ex) # no proper EOF handling for DatumReader     if msg == None : break     print(msg, rbuff.tell())


Schema Evolution

By using both the writer and reader schemas when reading data, Avro allows the schema to evolve in a limited way—for example, when adding/removing fields or when fields are changed in a way that is allowed (e.g. int to long, float, or double; string to and from bytes). We’ve already seen an example of this in the reader code above. Note that if a field is missing from the data being read and does not have a default value in the reader’s schema, an error will be thrown. More here.

Avro Schema Registry

Given the many potential factors (e.g., that Avro messages could be consumed by many different applications, each is going to need a schema to read/write the messages, the schemas could change, and there could be multiple schema versions in use), it makes sense to keep the schemas versioned and stored in a central registry, the Schema Registry. Typically, this would be a distributed, REST API service that manages and returns schema definitions.


We’ve seen what the Avro data format is, how to define schemas, and how to read and write objects to and from Avro. If you want a fast, compact, and well-supported format that’s easy to use and validate, you should consider Avro. And if you’re looking for some useful tools to help you with Avro, check out this Avro schema validator.

This post was written by Manoj Mokashi. Manoj has more than 25 years of experience as a developer, mostly on java, web technologies, and databases. He’s also worked on PHP, Python, Spark, AI/ML, and Solr and he really enjoys learning new things.


SQream Avro Support


Avro is a popular binary row-based serialized textual format. It can be seen as a binary alternative to JSON – drawing inspiration from its flexibility and nesting, while offering a much more efficient storage method.  Avro data consists of a JSON-formatted schema and a set of data payloads, which are either serialized in the Avro binary format or in JSON – the latter mostly used for debugging, and not supported by itself in all Avro-consuming products.  SQream supports Avro with simple data ingestion capabilities for both batch and streaming (for example, Kafka).  Once ingested, (see image below) Avro object(s) are mapped into rows. All existing data types can be represented in JSON fields, including JSONPath support and encoding. It is a simple plug and play solution for Avro objects/files to reduce the complexity of and accelerate TTTI.