SPSC RingBuffer

SpscRingBuffer[A] is optimized for the simplest and fastest case: exactly one producer thread and one consumer thread. It uses the FastFlow algorithm to eliminate all cross-core cache traffic, achieving nanosecond-scale latencies with no volatile reads on the fast path.

Why FastFlow?

Understanding the FastFlow algorithm helps explain why SpscRingBuffer achieves such high performance.

Imagine two threads sharing data with a traditional lock-based queue like ArrayBlockingQueue. They use a mutex (lock) to coordinate:

1. Producer thread: acquires lock, adds element, releases lock
2. Consumer thread: acquires lock, removes element, releases lock

This works, but locks have a cost: when threads contend for the same lock, one thread blocks (goes to sleep) while waiting. Waking a thread is expensive (thousands of CPU cycles). Even lock-free queues using synchronized or volatile reads can cause cache-coherency traffic: when one CPU core writes to a variable another core reads, the entire cache line must be invalidated and transferred — a costly operation that slows down both cores.

The fundamental issue: if the producer has to read what the consumer wrote (or vice versa), their CPU caches constantly fight. This is called false sharing and can reduce throughput by 10x or more on heavily loaded systems.

The problem is that any read by the producer of the consumer's state (or vice versa) causes cache line bouncing. So FastFlow's radical idea: neither side should ever read the other's state.

Instead, the producer simply writes data into slots and marks them non-null. The consumer independently reads slots and takes any non-null values. The array slot's null/non-null status itself is the only coordination needed — a happens-before relationship written once by the producer, read once by the consumer. No locks, no atomic operations on the fast path, no reading the other side's counters.

The note-passing analogy:

• You (producer) have a row of empty desks between you and your friend (consumer)
• Empty desk = null (available)
• Note on desk = non-null value (message ready)
• You never look at your friend's side — you just place notes on empty desks
• Your friend never looks at your side — they just pick up notes they see
• No shouting "are you ready?", no waiting, no lock contention

This is how FastFlow solves the cache-coherency problem: since producer and consumer never read each other's counters, there's no cache line bouncing between CPU cores. All coordination happens through the slots themselves, which are written once and read once.

The look-ahead cache is a further optimization: the producer maintains a local cached limit (producerLimit) so they don't have to check every slot individually. It's like glancing ahead at the next N desks to see if they're empty, without actually bending over to look. This keeps the fast path extremely fast — just a local counter check and an array write.

Why FastFlow is fast:

• No locks — no thread ever blocks another
• Minimal cache coordination — producer and consumer touch separate memory locations; the array slot is written once, read once
• Write-once, read-once semantics — the slot's null/non-null status is the handshake
• Cache-line padding — producer and consumer indices padded to separate cache lines, eliminating false sharing

The result: lock-free communication that scales linearly with CPU count and achieves nanosecond-scale latencies. This is why FastFlow is the algorithm of choice for SPSC ring buffers in high-performance systems like trading platforms, game engines, and network stacks. (The slow path's retry on full buffers means the algorithm is not wait-free, though contention is rare.)

Algorithm

SpscRingBuffer uses the FastFlow pattern, originally developed for the C++ FastFlow framework and popularized in Java by JCTools and the LMAX Disruptor. The core insight: the array element's null/non-null state is the synchronization signal. The producer never reads consumerIndex; the consumer never reads producerIndex. This eliminates all cross-core cache traffic between the two sides.

How it works:

• The producer walks forward through the array, placing elements into slots that are currently null. It marks a slot non-null after writing. The producer never reads what the consumer has consumed — it only checks its own cached producerLimit to know how many slots are available.
• The consumer walks forward through the array, reading any slot that contains a non-null value. After reading, it writes null to clear the slot. The consumer never reads the producer's index.
• The slot's nullness itself is the coordination: written once by the producer, read once by the consumer. No locks, no atomic operations on the fast path, no reading the other side's counters.

Look-ahead cache: The producer maintains a local producerLimit (derived from consumerIndex during slow path). This allows the fast path to check a simple local counter instead of reading the volatile consumerIndex on every offer. The look-ahead step is max(1, min(capacity/4, 4096)), balancing between reducing consumer reads and memory usage.

Diagram

To see the single-producer single-consumer FastFlow algorithm in action, use this interactive stepper. Type any label, click Offer to enqueue or Take to dequeue, and watch how the producer and consumer coordinate using only release/acquire semantics — no CAS on either side.

SPSC Ring Buffer · Capacity 4

Step summary

step summary will appear here after each operation

Algorithm trace

algorithm trace will appear here after each operation

Operation history

#1initial state

Type any label · Offer to enqueue · Take to dequeue · no CAS on either side — only release/acquire stores and loads · stale slots show the no-null-clear design · use Back / Forward to replay any step

Here is a complete walkthrough of every variable in the trace, in the order the algorithm computes them.

pIdx — the monotonic producer counter (plain load/store)

The producer reads pIdx with a plain load — no atomic operations at all. Since only one thread produces, no synchronization is needed. The producer maintains its own count, never reads consumerIndex, and advances pIdx after writing with a release store.

cIdx — the monotonic consumer counter

SPSC avoids reading consumerIndex entirely. Instead of checking the occupancy formula (pIdx - cIdx), the producer reads the array slot at producerIndex + lookAheadStep to determine if a slot is free (on the slow path when the look-ahead cache is exhausted). This slot-based approach eliminates all counter reads between producer and consumer.

size = pIdx − cIdx — the occupancy check

Both sides use the same occupancy formula, but only during slow paths (rare). On the fast path, neither side reads the other's counter; they rely on slot nullness (FastFlow) for synchronization.

slot = pIdx & mask or slot = cIdx & mask — the circular array index

Identical bitmask arithmetic: & mask replaces modulo, giving a slot in [0, capacity).

element = buf[slot] — no CAS, plain slot reads

Unlike SPMC and MPMC, the consumer reads the slot directly with an acquire load — because there's only one consumer, no read-before-CAS is needed. The consumer is the sole owner of the take path. After reading, the consumer nulls the slot with a release store, which serves two purposes: (1) signals to the producer that the slot is free, and (2) releases any bookkeeping updates to other cores via the release semantics.

Why SPSC doesn't read producer state

The consumer never reads pIdx. Instead, it reads the slot directly: if the slot is non-null, data is ready. If null, the buffer is empty. The producer wrote a non-null value via a release store, and the consumer reads via an acquire load, forming a happens-before pair. This is the core of FastFlow: eliminate cross-core reads entirely.

The diagram marks slots as "live" (still owned by the producer-consumer pair) or "stale" (consumed but not yet overwritten). Unlike SPMC where consumers CAS and have ordering concerns, SPSC clears immediately because only one consumer exists — there's no race.

Creating Instances

Ring buffers are instantiated via the companion object's apply method. SpscRingBuffer[A] uses the FastFlow pattern with a look-ahead cache. On the fast path, the producer checks a cached producerLimit to avoid reading consumerIndex. When the cached limit is exhausted, the slow path reads the array slot at producerIndex + lookAheadStep — never consumerIndex directly. This keeps the producer and consumer cache lines fully independent. The consumer uses null/non-null slot reads (FastFlow semantics). Together, these avoid repeated volatile reads and minimize cross-core cache traffic.

object SpscRingBuffer {
  def apply[A <: AnyRef](capacity: Int): SpscRingBuffer[A]
}

Creates a new SPSC ring buffer with the given capacity. The capacity must be a positive power of two.

We create an SPSC buffer as follows:

import zio.blocks.ringbuffer.SpscRingBuffer

val rb = SpscRingBuffer[String](16)  // capacity must be power of 2

Operations

SpscRingBuffer provides operations for sending and receiving elements, as well as querying buffer state:

Inserting Elements — SpscRingBuffer#offer

Use SpscRingBuffer#offer to insert an element with this signature:

final class SpscRingBuffer[A <: AnyRef](val capacity: Int) {
  def offer(a: A): Boolean
}

Inserts the element without blocking. Returns true on successful insertion, false when the buffer is full. Raises NullPointerException if the element is null. Call only from the producer thread.

To handle backpressure by checking if insertion succeeds:

import zio.blocks.ringbuffer.SpscRingBuffer

val rb = SpscRingBuffer[String](4)

When the buffer becomes full, offer returns false:

val result1 = rb.offer("a")
val result2 = rb.offer("b")
val result3 = rb.offer("c")
val result4 = rb.offer("d")
val result5 = rb.offer("e")

Removing Elements — SpscRingBuffer#take

Use SpscRingBuffer#take to remove an element with this signature:

final class SpscRingBuffer[A <: AnyRef](val capacity: Int) {
  def take(): A
}

Retrieves and removes an element from the front of the buffer. Returns immediately without blocking, providing the element or null if the buffer is empty. Call only from the consumer thread.

To retrieve elements in FIFO order:

rb.take()
rb.take()
rb.take()
rb.take()
rb.take()

Checking State — size, isEmpty, isFull

Ring buffers provide three query methods to check their state. Note that under concurrent access, these results are approximate — by the time the method returns, other threads may modify the buffer.

Use SpscRingBuffer#size to get the approximate element count with this signature:

final class SpscRingBuffer[A <: AnyRef](val capacity: Int) {
  def size: Int
}

Returns the approximate number of elements currently in the buffer. Under concurrent access, the result may be stale. O(1).

Use SpscRingBuffer#isEmpty to check if the buffer is empty with this signature:

final class SpscRingBuffer[A <: AnyRef](val capacity: Int) {
  def isEmpty: Boolean
}

Returns true if the buffer contains no elements (approximate). O(1).

Use SpscRingBuffer#isFull to check if the buffer is at capacity with this signature:

final class SpscRingBuffer[A <: AnyRef](val capacity: Int) {
  def isFull: Boolean
}

Returns true if the buffer is at capacity (approximate). O(1).

All four implementations provide the same three methods. To check state after operations:

import zio.blocks.ringbuffer.SpscRingBuffer

val rb2 = SpscRingBuffer[String](8)

State queries are cheap but approximate under concurrency:

rb2.offer("x")
rb2.offer("y")

rb2.size
rb2.isEmpty
rb2.isFull

Batch Operations — SpscRingBuffer#drain and SpscRingBuffer#fill

Use SpscRingBuffer#drain to consume up to N elements with this signature:

final class SpscRingBuffer[A <: AnyRef](val capacity: Int) {
  def drain(consumer: A => Unit, limit: Int): Int
}

Removes up to limit elements from the buffer, passing each to the consumer callback. Returns the number of elements actually drained. Raises IllegalArgumentException if limit is negative. Call only from the consumer thread. O(n) where n is the number of elements drained. If the callback raises an exception, all elements passed to it up to that point remain consumed and the buffer stays in a consistent state.

Use SpscRingBuffer#fill to produce up to N elements with this signature:

final class SpscRingBuffer[A <: AnyRef](val capacity: Int) {
  def fill(supplier: () => A, limit: Int): Int
}

Inserts up to limit elements by calling the supplier for each new element. Returns the number of elements actually inserted. Raises IllegalArgumentException if limit is negative. Raises NullPointerException if the supplier returns null. Call only from the producer thread. O(n) where n is the number of elements inserted.

Batch operations amortize synchronization costs. To drain multiple elements at once:

import zio.blocks.ringbuffer.SpscRingBuffer

val rb3 = SpscRingBuffer[java.lang.Integer](16)
(1 to 5).foreach(i => rb3.offer(Integer.valueOf(i)))

val collected = scala.collection.mutable.Buffer[java.lang.Integer]()
val drained = rb3.drain(collected += _, 10)  // drained = 5

println(s"Drained items: ${collected.mkString(", ")}")

To avoid repeated offer calls when producing elements, use fill:

import zio.blocks.ringbuffer.SpscRingBuffer
import java.util.concurrent.atomic.AtomicInteger

val rb4 = SpscRingBuffer[String](8)
val counter = new AtomicInteger(0)
val filled = rb4.fill(() => { counter.incrementAndGet(); s"item-${counter.get()}" }, 5)

println(s"Filled $filled items")

Primitive variants

The generic SpscRingBuffer[A] stores AnyRef slots, which boxes Java primitives on the producer side. For SPSC workloads handing off Int, Long, Float, or Double values between two threads, four zero-boxing companions are provided:

Type	Class	Slot encoding
Int	IntSpscRingBuffer	one Long per slot: high 32 bits hold an occupancy tag (0 = empty, 1 = data, 2 = DONE sentinel), low 32 bits hold the Int payload. The whole 64-bit word is 0L when empty, so any tagged value is non-zero regardless of the payload bits.
Long	LongSpscRingBuffer	one Long per slot. Long.MinValue is reserved as the empty marker and Long.MinValue + 1L as the in-band DONE sentinel; offering either value throws IllegalArgumentException. All other Long values pass through unchanged.
Float	FloatSpscRingBuffer	one Long per slot: high 32 bits tag (0 = empty, 1 = data, 2 = DONE), low 32 bits hold floatToRawIntBits(value). The whole word is 0L when empty. Every Float (including 0.0f, NaN, ±Inf) is representable.
Double	DoubleSpscRingBuffer	one Long per slot. doubleToRawLongBits(value) is stored directly; two reserved non-canonical NaN bit patterns mark empty (0xFFF8_0000_0000_0001L) and DONE (0xFFF8_0000_0000_0002L). Any NaN payload is canonicalized to Double.NaN on offer so it can never collide with the reserved markers.

All four use the same FastFlow algorithm as the generic version — single backing Array[Long], padded indices, look-ahead cache — and expose the same core operations (offer, take, peek, isEmpty, isFull, size, capacity). The only API difference is that take returns a primitive instead of AnyRef, so callers should check peek() first to know whether the next read will be a real value or just the empty marker. (The generic drain/fill batch helpers are not provided on the primitive variants; loop peek/take directly.)

In addition, each primitive variant exposes two methods for in-band end-of-stream signalling, used by the concurrent stream operators to propagate "no more data" through a primitive queue without needing a separate side channel:

• offerDone(): Boolean — write the DONE sentinel into the next free slot, returning true on success or false if the queue is full. Once offerDone() succeeds the producer should not offer any further values.
• pollPacked(): Long — atomically poll the next slot and return its packed encoding: EMPTY_PACKED/EMPTY / EMPTY_BITS when empty (consumer index NOT advanced), DONE_PACKED/DONE / DONE_BITS for the DONE sentinel (index advanced), or a non-sentinel encoded value otherwise (index advanced). The exact constant names live on each companion object (IntSpscRingBuffer.DONE_PACKED, LongSpscRingBuffer.DONE, FloatSpscRingBuffer.DONE_PACKED, DoubleSpscRingBuffer.DONE_BITS).

pollPacked is intended for advanced consumers that want a single atomic operation distinguishing empty / data / DONE. Most application code that simply needs primitive throughput is fine using the higher-level concurrent stream operators below.

import zio.blocks.ringbuffer.{IntSpscRingBuffer, LongSpscRingBuffer, DoubleSpscRingBuffer}

val ints = new IntSpscRingBuffer(16)
ints.offer(42)
if (ints.peek()) {
  val v: Int = ints.take()     // 42
}

val longs = new LongSpscRingBuffer(16)
longs.offer(Long.MaxValue)
if (longs.peek()) {
  val w: Long = longs.take()   // Long.MaxValue
}

val doubles = new DoubleSpscRingBuffer(16)
doubles.offer(Double.NaN)      // canonicalized to Double.NaN's bit pattern on offer

These variants underpin the primitive-specialized concurrent stream operators (mapPar, mergeAll, flatMapPar) so primitive streams do not box during cross-thread handoff.

Examples

SPSC: Producer-Consumer Ping-Pong

In a single-producer, single-consumer setup, use SpscRingBuffer for maximum throughput. This example demonstrates how two threads communicate efficiently using FastFlow signaling.

import docs.SourceFile

SourceFile.print("zio-blocks-examples/src/main/scala/ringbuffer/SpscExample.scala")

(source)

sbt "zio-blocks-examples/runMain ringbuffer.SpscExample"

Batch Fill and Drain (SPSC)

Use fill and drain for efficient batch operations. This example shows how to amortize synchronization costs by processing multiple elements at once.

import docs.SourceFile

SourceFile.print("zio-blocks-examples/src/main/scala/ringbuffer/BatchExample.scala")

(source)

sbt "zio-blocks-examples/runMain ringbuffer.BatchExample"