---

name: crystal-concurrency description: Use when implementing concurrent programming in Crystal using fibers, channels, and parallel execution patterns for high-performance, non-blocking applications. allowed-tools: [Bash, Read]

Crystal Concurrency

You are Claude Code, an expert in Crystal's concurrency model. You specialize in building high-performance, concurrent applications using fibers, channels, and Crystal's lightweight concurrency primitives.

Your core responsibilities:

• Implement fiber-based concurrent operations for non-blocking execution
• Design channel-based communication patterns for inter-fiber coordination
• Build parallel processing pipelines with proper synchronization
• Implement worker pools and task distribution systems
• Handle concurrent resource access with mutexes and atomic operations
• Design fault-tolerant concurrent systems with proper error handling
• Optimize fiber scheduling and resource utilization
• Implement backpressure and flow control mechanisms
• Build real-time data processing systems
• Design concurrent I/O operations for network and file systems

Fibers: Lightweight Concurrency

Crystal uses fibers (also known as green threads or coroutines) for concurrency. Fibers are cooperatively scheduled by the Crystal runtime and are much lighter weight than OS threads.

Basic Fiber Spawning

# Simple fiber spawning
spawn do
  puts "Running in a fiber"
  sleep 1
  puts "Fiber completed"
end

# Fiber with arguments
def process_data(id : Int32, data : String)
  puts "Processing #{data} with id #{id}"
  sleep 0.5
  puts "Completed #{id}"
end

spawn process_data(1, "task A")
spawn process_data(2, "task B")

# Wait for fibers to complete
sleep 1

Fiber with Return Values via Channels

# Fibers don't return values directly, use channels instead
result_channel = Channel(Int32).new

spawn do
  result = expensive_computation(42)
  result_channel.send(result)
end

# Do other work...
puts "Doing other work"

# Wait for result
result = result_channel.receive
puts "Got result: #{result}"

def expensive_computation(n : Int32) : Int32
  sleep 1
  n * 2
end

Named Fibers for Debugging

# Give fibers descriptive names for debugging
spawn(name: "data-processor") do
  process_large_dataset
end

spawn(name: "cache-updater") do
  update_cache_periodically
end

# Fiber names appear in exception backtraces
spawn(name: "failing-worker") do
  raise "Something went wrong"
end

Channels: Inter-Fiber Communication

Channels are the primary mechanism for communication between fibers. They provide thread-safe message passing with optional buffering.

Unbuffered Channels

# Unbuffered channel - blocks until both sender and receiver are ready
channel = Channel(String).new

spawn do
  puts "Sending message"
  channel.send("Hello")
  puts "Message sent"
end

spawn do
  sleep 0.1  # Small delay
  puts "Receiving message"
  msg = channel.receive
  puts "Received: #{msg}"
end

sleep 1

Buffered Channels

# Buffered channel - allows sending without blocking up to buffer size
channel = Channel(Int32).new(capacity: 3)

# These sends won't block
channel.send(1)
channel.send(2)
channel.send(3)

# This would block until someone receives
# channel.send(4)

# Receive values
puts channel.receive  # 1
puts channel.receive  # 2
puts channel.receive  # 3

Channel Closing and Iteration

# Producer-consumer with channel closing
channel = Channel(Int32).new

# Producer
spawn do
  5.times do |i|
    channel.send(i)
    sleep 0.1
  end
  channel.close  # Signal no more values
end

# Consumer - iterate until channel is closed
spawn do
  channel.each do |value|
    puts "Received: #{value}"
  end
  puts "Channel closed, consumer exiting"
end

sleep 1

Checking if Channel is Closed

channel = Channel(String).new

spawn do
  channel.send("message 1")
  channel.send("message 2")
  channel.close
end

sleep 0.1

# Check before receiving
unless channel.closed?
  puts channel.receive
end

# Or handle the exception
begin
  puts channel.receive
  puts channel.receive
  puts channel.receive  # Will raise Channel::ClosedError
rescue Channel::ClosedError
  puts "Channel is closed"
end

Select: Multiplexing Channels

The select statement allows waiting on multiple channel operations simultaneously, similar to Go's select statement.

Basic Select with Multiple Channels

ch1 = Channel(String).new
ch2 = Channel(Int32).new

spawn do
  sleep 0.2
  ch1.send("from channel 1")
end

spawn do
  sleep 0.1
  ch2.send(42)
end

# Wait for whichever channel is ready first
select
when msg = ch1.receive
  puts "Got string: #{msg}"
when num = ch2.receive
  puts "Got number: #{num}"
end

sleep 1

Select with Timeout

channel = Channel(String).new

spawn do
  sleep 2  # Takes too long
  channel.send("delayed message")
end

# Wait with timeout
select
when msg = channel.receive
  puts "Received: #{msg}"
when timeout(1.second)
  puts "Timed out waiting for message"
end

Select with Default Case (Non-blocking)

channel = Channel(Int32).new

# Non-blocking receive
select
when value = channel.receive
  puts "Got value: #{value}"
else
  puts "No value available, continuing immediately"
end

Select in a Loop

results = Channel(String).new
done = Channel(Nil).new
output = [] of String

# Multiple workers sending results
3.times do |i|
  spawn do
    sleep rand(0.5..1.5)
    results.send("Worker #{i} done")
  end
end

# Collector fiber
spawn do
  3.times do
    output << results.receive
  end
  done.send(nil)
end

# Wait for completion with timeout
select
when done.receive
  puts "All workers completed"
  output.each { |msg| puts msg }
when timeout(5.seconds)
  puts "Timeout - not all workers completed"
end

Worker Pools

Worker pools distribute tasks across a fixed number of concurrent workers.

Basic Worker Pool

class WorkerPool(T, R)
  def initialize(@size : Int32)
    @tasks = Channel(T).new
    @results = Channel(R).new
    @workers = [] of Fiber

    @size.times do |i|
      @workers << spawn(name: "worker-#{i}") do
        worker_loop
      end
    end
  end

  private def worker_loop
    @tasks.each do |task|
      result = process(task)
      @results.send(result)
    end
  end

  def process(task : T) : R
    # Override in subclass or pass block
    raise "Not implemented"
  end

  def submit(task : T)
    @tasks.send(task)
  end

  def get_result : R
    @results.receive
  end

  def shutdown
    @tasks.close
  end
end

# Usage example
class IntSquarePool < WorkerPool(Int32, Int32)
  def process(task : Int32) : Int32
    sleep 0.1  # Simulate work
    task * task
  end
end

pool = IntSquarePool.new(size: 3)

# Submit tasks
10.times { |i| pool.submit(i) }

# Collect results
results = [] of Int32
10.times { results << pool.get_result }

pool.shutdown
puts results.sort

Worker Pool with Error Handling

struct Task
  property id : Int32
  property data : String

  def initialize(@id, @data)
  end
end

struct Result
  property task_id : Int32
  property success : Bool
  property value : String?
  property error : String?

  def initialize(@task_id, @success, @value = nil, @error = nil)
  end
end

class RobustWorkerPool
  def initialize(@worker_count : Int32)
    @tasks = Channel(Task).new(capacity: 100)
    @results = Channel(Result).new(capacity: 100)

    @worker_count.times do |i|
      spawn(name: "worker-#{i}") do
        process_tasks
      end
    end
  end

  private def process_tasks
    @tasks.each do |task|
      begin
        result_value = process_task(task)
        @results.send(Result.new(
          task_id: task.id,
          success: true,
          value: result_value
        ))
      rescue ex
        @results.send(Result.new(
          task_id: task.id,
          success: false,
          error: ex.message
        ))
      end
    end
  end

  private def process_task(task : Task) : String
    # Simulate processing that might fail
    raise "Invalid data" if task.data.empty?
    sleep 0.1
    "Processed: #{task.data}"
  end

  def submit(task : Task)
    @tasks.send(task)
  end

  def get_result : Result
    @results.receive
  end

  def shutdown
    @tasks.close
  end
end

Parallel Map and Reduce

Implement parallel processing of collections.

Parallel Map

def parallel_map(collection : Array(T), workers : Int32 = 4, &block : T -> R) : Array(R) forall T, R
  tasks = Channel(Tuple(Int32, T)).new
  results = Channel(Tuple(Int32, R)).new

  # Spawn workers
  workers.times do
    spawn do
      tasks.each do |index, item|
        result = yield item
        results.send({index, result})
      end
    end
  end

  # Send tasks
  spawn do
    collection.each_with_index do |item, index|
      tasks.send({index, item})
    end
    tasks.close
  end

  # Collect results in order
  result_map = {} of Int32 => R
  collection.size.times do
    index, result = results.receive
    result_map[index] = result
  end

  collection.indices.map { |i| result_map[i] }
end

# Usage
numbers = (1..100).to_a
squares = parallel_map(numbers, workers: 8) do |n|
  sleep 0.01  # Simulate work
  n * n
end

puts squares.first(10)

Parallel Reduce with Pipeline

def parallel_reduce(collection : Array(T), workers : Int32 = 4, initial : R, &block : R, T -> R) : R forall T, R
  chunk_size = (collection.size / workers.to_f).ceil.to_i
  chunks = collection.each_slice(chunk_size).to_a

  results = Channel(R).new

  chunks.each do |chunk|
    spawn do
      chunk_result = chunk.reduce(initial) { |acc, item| yield acc, item }
      results.send(chunk_result)
    end
  end

  # Reduce the partial results
  final_result = initial
  chunks.size.times do
    final_result = yield final_result, results.receive
  end

  final_result
end

# Usage - sum of squares
numbers = (1..1000).to_a
sum = parallel_reduce(numbers, initial: 0) do |acc, n|
  acc + n * n
end

puts "Sum of squares: #{sum}"

Mutex: Protecting Shared State

When fibers need to share mutable state, use mutexes to prevent race conditions.

Basic Mutex Usage

require "mutex"

class Counter
  def initialize
    @count = 0
    @mutex = Mutex.new
  end

  def increment
    @mutex.synchronize do
      current = @count
      sleep 0.001  # Simulate some work
      @count = current + 1
    end
  end

  def value : Int32
    @mutex.synchronize { @count }
  end
end

counter = Counter.new

# Spawn 100 fibers that each increment 10 times
100.times do
  spawn do
    10.times { counter.increment }
  end
end

sleep 2
puts "Final count: #{counter.value}"  # Should be 1000

Read-Write Lock Pattern

require "mutex"

class CachedData
  def initialize
    @data = {} of String => String
    @mutex = Mutex.new
    @version = 0
  end

  def read(key : String) : String?
    @mutex.synchronize do
      @data[key]?
    end
  end

  def write(key : String, value : String)
    @mutex.synchronize do
      @data[key] = value
      @version += 1
    end
  end

  def batch_update(updates : Hash(String, String))
    @mutex.synchronize do
      updates.each do |key, value|
        @data[key] = value
      end
      @version += 1
    end
  end

  def snapshot : Hash(String, String)
    @mutex.synchronize do
      @data.dup
    end
  end
end

Atomic Operations

For simple counters and flags, atomic operations are more efficient than mutexes.

Atomic Counter

require "atomic"

class AtomicCounter
  def initialize(initial : Int32 = 0)
    @count = Atomic(Int32).new(initial)
  end

  def increment : Int32
    @count.add(1)
  end

  def decrement : Int32
    @count.sub(1)
  end

  def value : Int32
    @count.get
  end

  def compare_and_set(expected : Int32, new_value : Int32) : Bool
    @count.compare_and_set(expected, new_value)
  end
end

counter = AtomicCounter.new

# Safe concurrent increments without mutex
1000.times do
  spawn { counter.increment }
end

sleep 1
puts "Count: #{counter.value}"

Atomic Flag for Coordination

require "atomic"

class ShutdownCoordinator
  def initialize
    @shutdown_flag = Atomic(Int32).new(0)
  end

  def shutdown!
    @shutdown_flag.set(1)
  end

  def shutdown? : Bool
    @shutdown_flag.get == 1
  end

  def run_until_shutdown(&block)
    until shutdown?
      yield
      sleep 0.1
    end
  end
end

coordinator = ShutdownCoordinator.new

# Worker that checks shutdown flag
spawn(name: "worker") do
  coordinator.run_until_shutdown do
    puts "Working..."
  end
  puts "Worker shutdown gracefully"
end

sleep 1
coordinator.shutdown!
sleep 0.5

When to Use This Skill

Use the crystal-concurrency skill when you need to:

• Process multiple I/O operations concurrently (network requests, file operations)
• Implement real-time data processing pipelines
• Build worker pools for parallel task processing
• Handle multiple client connections simultaneously (web servers, chat systems)
• Perform background processing without blocking main execution
• Aggregate results from multiple concurrent operations
• Implement producer-consumer patterns
• Build rate limiters and backpressure mechanisms
• Process large datasets in parallel
• Coordinate multiple asynchronous operations
• Implement timeout and cancellation patterns
• Build concurrent caches with synchronized access
• Stream data processing with multiple stages
• Implement fan-out/fan-in patterns

Best Practices

1. Always Close Channels: Close channels when done sending to signal completion to receivers
2. Use Buffered Channels for Performance: Buffer channels when producers/consumers run at different speeds
3. Limit Fiber Count: Don't spawn unlimited fibers; use worker pools for bounded concurrency
4. Handle Channel Closure: Always handle Channel::ClosedError or check closed? before operations
5. Use Select for Timeouts: Implement timeouts with select and timeout() to prevent infinite blocking
6. Prefer Channels Over Shared State: Use message passing (channels) instead of shared memory when possible
7. Synchronize Shared State: Always use Mutex or atomics when sharing mutable state between fibers
8. Clean Up Resources: Use ensure blocks to guarantee resource cleanup even on errors
9. Name Your Fibers: Give fibers descriptive names for easier debugging and profiling
10. Avoid Blocking Operations in Fibers: Use non-blocking I/O; blocking operations prevent other fibers from running
11. Use Atomic Operations for Counters: Atomics are more efficient than mutexes for simple counters and flags
12. Implement Graceful Shutdown: Design systems to shut down cleanly, draining channels and waiting for fibers
13. Handle Fiber Panics: Wrap fiber code in exception handlers to prevent silent failures
14. Size Channel Buffers Appropriately: Too small causes blocking; too large wastes memory
15. Use Select Default for Polling: Non-blocking checks with select ... else for polling patterns

Common Pitfalls

1. Forgetting to Close Channels: Receivers will wait forever if channels aren't closed after sending completes
2. Deadlocks from Unbuffered Channels: Sending to unbuffered channel blocks until receiver is ready
3. Race Conditions on Shared State: Not using mutexes/atomics when multiple fibers access same data
4. Channel Buffer Overflow: Sending more items than buffer capacity without receivers causes blocking
5. Not Handling Closed Channels: Receiving from closed channel raises exception; always handle it
6. Spawning Too Many Fibers: Unlimited fiber spawning exhausts memory; use worker pools instead
7. Blocking the Scheduler: CPU-intensive work in fibers prevents other fibers from running
8. Resource Leaks: Not closing channels, files, or connections in all code paths including errors
9. Order Assumptions: Fibers execute in non-deterministic order; don't assume execution sequence
10. Timeout Too Short: Aggressive timeouts cause false failures; balance responsiveness with reliability
11. Mutex Held Too Long: Long critical sections reduce concurrency; minimize mutex hold time
12. Send/Receive Mismatch: Imbalanced producers/consumers leads to memory buildup or starvation
13. Ignoring Fiber Exceptions: Exceptions in fibers don't propagate to spawner; handle explicitly
14. Nested Mutex Locks: Can cause deadlocks; avoid acquiring multiple mutexes or use consistent order
15. Not Using synchronize: Forgetting to wrap mutex usage in synchronize block causes race conditions

Resources

• Crystal Concurrency Guide
• Crystal API - Fiber
• Crystal API - Channel
• Crystal API - Mutex
• Crystal API - Atomic
• Crystal Book - Concurrency
• Effective Crystal - Concurrency Patterns