.PB

Introduction

Protocol Buffers is Google’s format for structured data, used inside Google since the early 2000s and open source since 2008. The fields of each message are declared in a .proto schema, and the protoc compiler turns the schema into code that reads and writes the compact binary form.

A file holding one encoded message is commonly given .pb; the Protocol Buffers documentation recommends .binpb. Many larger formats are Protocol Buffers messages underneath, including ONNX models.

Software

• protoc and the official libraries for C++, Java, Python, Go, C# and more
• Protoscope, which prints any message as readable text, with or without the schema
• Buf, a command-line tool for schemas and encoded messages
• Netron, for machine learning models stored as Protocol Buffers

Structure

An encoded message has no header and no magic number. It is a run of fields, and each field starts with a tag: the field number from the schema shifted left three bits, plus a three-bit wire type saying how the value is stored.

Tags, lengths and integers are varints: seven bits per byte, least significant group first, with the top bit set on every byte except the last. Fixed-size values are little-endian. The field names are not stored, so the schema is needed to give the numbers meaning.