Tag: karafka

Introduction

Using Kafka for Remote Procedure Calls (RPC) might raise eyebrows among seasoned developers. At its core, RPC is a programming technique that creates the illusion of running a function on a local machine when it executes on a remote server. When you make an RPC call, your application sends a request to a remote service, waits for it to execute some code, and then receives the results - all while making it feel like a regular function call in your code.

Apache Kafka, however, was designed as an event log, optimizing for throughput over latency. Yet, sometimes unconventional approaches yield surprising results. This article explores implementing RPC patterns with Kafka using the Karafka framework. While this approach might seem controversial - and rightfully so - understanding its implementation, performance characteristics, and limitations may provide valuable insights into Kafka's capabilities and distributed system design.

The idea emerged from discussing synchronous communication in event-driven architectures. What started as a theoretical question - "Could we implement RPC with Kafka?" - evolved into a working proof of concept that achieved millisecond response times in local testing.

In modern distributed systems, the default response to new requirements often involves adding another specialized tool to the technology stack. However, this approach comes with its own costs:

• Increased operational complexity,
• Additional maintenance overhead,
• And more potential points of failure.

Sometimes, stretching the capabilities of existing infrastructure - even in unconventional ways - can provide a pragmatic solution that avoids these downsides.

Disclaimer: This implementation serves as a proof-of-concept and learning resource. While functional, it lacks production-ready features like proper timeout handling, resource cleanup after timeouts, error propagation, retries, message validation, security measures, and proper metrics/monitoring. The implementation also doesn't handle edge cases like Kafka cluster failures. Use this as a starting point to build a more robust solution.

Architecture Overview

Building an RPC pattern on top of Kafka requires careful consideration of both synchronous and asynchronous aspects of communication. At its core, we're creating a synchronous-feeling operation by orchestrating asynchronous message flows underneath. From the client's perspective, making an RPC call should feel synchronous - send a request and wait for a response. However, once a command enters Kafka, all the underlying operations are asynchronous.

Core Components

Such an architecture has to rely on several key components working together:

• Two Kafka topics form the backbone - a command topic for requests and a result topic for responses.
• A client-side consumer, running without a consumer group, that actively matches correlation IDs and starts from the latest offset to ensure we only process relevant messages.
• The commands consumer in our RPC server that processes requests and publishes results
• A synchronization mechanism using mutexes and condition variables that maintain thread safety and handles concurrent requests.

Implementation Flow

A unique correlation ID is always generated when a client initiates an RPC call. The command is then published to Kafka, where it's processed asynchronously. The client blocks execution using a mutex and condition variable while waiting for the response. Meanwhile, the message flows through several stages:

• command topic persistence,
• consumer polling and processing,
• result publishing,
• result topic persistence,
• and finally, the client-side consumer matching of the correlation ID with the response and completion signaling,

Below, you can find a visual representation of the RPC flow over Kafka. The diagram shows the journey of a single request-response cycle:

Design Considerations

This architecture makes several conscious trade-offs. We use single-partition topics to ensure strict ordering, which limits throughput but simplifies correlation and provides exactly-once processing semantics - though the partition count and other things could be adjusted if higher scale becomes necessary. The custom consumer approach avoids consumer group rebalancing delays, while the synchronization mechanism bridges the gap between Kafka's asynchronous nature and our desired synchronous behavior. While this design prioritizes correctness over maximum throughput, it aligns well with typical RPC use cases where reliability and simplicity are key requirements.

Implementation Components

Getting from concept to working code requires several key components to work together. Let's examine the implementation of our RPC pattern with Kafka.

Topic Configuration

First, we need to define our topics. We use a single-partition configuration to maintain message ordering:

topic :commands do
  config(partitions: 1)
  consumer CommandsConsumer
end

topic :commands_results do
  config(partitions: 1)
  active false
end

This configuration defines two essential topics:

• Command topic that receives and processes RPC requests
• Results topic marked as inactive since we'll use a custom iterator instead of a standard consumer group consumer

Command Consumer

The consumer handles incoming commands and publishes results back to the results topic:

class CommandsConsumer < ApplicationConsumer
  def consume
    messages.each do |message|
      Karafka.producer.produce_async(
        topic: 'commands_results',
        # We evaluate whatever Ruby code comes in the payload
        # We return stringified result of evaluation
        payload: eval(message.raw_payload).to_s,
        key: message.key
      )

      mark_as_consumed(message)
    end
  end
end

We're using a simple eval to process commands for demonstration purposes. You'd want to implement proper command validation, deserialization, and secure processing logic in production.

Synchronization Mechanism

To bridge Kafka's asynchronous nature with synchronous RPC behavior, we implement a synchronization mechanism using Ruby's mutex and condition variables:

class Accu
  include Singleton

  def initialize
    @running = {}
    @results = {}
  end

  def register(id)
    @running[id] = [Mutex.new, ConditionVariable.new]
  end

  def unlock(id, result)
    return false unless @running.key?(id)

    @results[id] = result
    mutex, cond = @running.delete(id)
    mutex.synchronize { cond.signal }
  end

  def result(id)
    @results.delete(id)
  end
end

This mechanism maintains a registry of pending requests and coordinates the blocking and unblocking of client threads based on correlation IDs. When a response arrives, it signals the corresponding condition variable to unblock the waiting thread.

The Client

Our client implementation brings everything together with two main components:

1. A response listener that continuously checks for matching results
2. A blocking command dispatcher that waits for responses

class Client
  class << self
    def run
      iterator = Karafka::Pro::Iterator.new(
        { 'commands_results' => true },
        settings: {
          'bootstrap.servers': '127.0.0.1:9092',
          'enable.partition.eof': false,
          'auto.offset.reset': 'latest'
        },
        yield_nil: true,
        max_wait_time: 100
      )

      iterator.each do |message|
        next unless message

        Accu.instance.unlock(message.key, message.raw_payload)
      rescue StandardError => e
        puts e
        sleep(rand)
        next
      end
   end

   def perform(ruby_remote_code)
      cmd_id = SecureRandom.uuid

      Karafka.producer.produce_sync(
        topic: 'commands',
        payload: ruby_remote_code,
        key: cmd_id
      )

      mutex, cond = Accu.instance.register(cmd_id)
      mutex.synchronize { cond.wait(mutex) }

      Accu.instance.result(cmd_id)
    end
  end
end

The client uses Karafka's Iterator to consume responses without joining a consumer group, which avoids rebalancing delays and ensures we only process new messages. The perform method handles the synchronous aspects:

• Generates a unique correlation ID
• Registers the request with our synchronization mechanism
• Sends the command
• Blocks until the response arrives

Using the Implementation

To use this RPC implementation, first start the response listener in a background thread:

# Do this only once per process
Thread.new { Client.run }

Then, you can make synchronous RPC calls from your application:

Client.perform('1 + 1')
#=> Remote result: 2

Each call blocks until the response arrives, making it feel like a regular synchronous method call despite the underlying asynchronous message flow.

Despite its simplicity, this implementation achieves impressive performance in local testing - roundtrip times as low as 3ms. However, remember this assumes ideal conditions and minimal command processing time. Real-world usage would need additional error handling, timeouts, and more robust command processing logic.

Performance Considerations

The performance characteristics of this RPC implementation are surprisingly good, but they come with important caveats and considerations that need to be understood for proper usage.

Local Testing Results

In our local testing environment, the implementation showed impressive numbers.

A single roundtrip can be completed in as little as 3ms. Even when executing 100 sequential commands:

require 'benchmark'

Benchmark.measure do
  100.times { Client.perform('1 + 1') }
end
#=> 0.035734   0.011570   0.047304 (  0.316631)

However, it's crucial to understand that these numbers represent ideal conditions:

• Local Kafka cluster
• Minimal command processing time
• No network latency
• No concurrent load

Summary

While Kafka wasn't designed for RPC patterns, this implementation demonstrates that with careful consideration and proper use of Karafka's features, we can build reliable request-response patterns on top of it. The approach shines particularly in environments where Kafka is already a central infrastructure, allowing messaging architecture to be extended without introducing additional technologies.

However, this isn't a silver bullet solution. Success with this pattern requires careful attention to timeouts, error handling, and monitoring. It works best when Kafka is already part of your stack, and your use case can tolerate slightly higher latencies than traditional RPC solutions.

This fascinating pattern challenges our preconceptions about messaging systems and RPC. It demonstrates that understanding your tools deeply often reveals capabilities beyond their primary use cases. While unsuitable for every situation, it provides a pragmatic alternative when adding new infrastructure components isn't desirable.