Aeron: Inside the Messaging System That Powers Half of the HFT Industry

Aeron: Media Driver, shared memory, triple-buffered logs — the architecture that set the standard for low-latency messaging.

When it comes to messaging systems for high-frequency trading, one name comes up in every conversation — Aeron. Developed by Martin Thompson and the Real Logic team (later Adaptive Financial Consulting), Aeron became the de facto standard for data transmission in a world where microseconds decide everything.

In this article, we break Aeron down into its parts: Transport, Archive, Cluster, and Sequencer. We explain how it works inside, where its strengths lie — and where the problems begin.

TL;DR

• Aeron — open-source (Apache 2.0) messaging system for low-latency applications
• IPC latency: ~250 ns round-trip via shared memory
• Throughput: 20+ million messages per second
• Four products: Transport (core), Archive (record/replay), Cluster (Raft), Sequencer (total order)
• Language: Java (primary) + C client (less complete)
• Used by: dozens of HFT firms, market makers, and exchanges

Part 1: Aeron Transport — The Core

Architecture: Media Driver

The central component is the Media Driver — a separate process (or embedded library) that manages all data transmission. Applications communicate with it via shared memory (mmap files in /dev/shm).

Key data structures:

• ManyToOneRingBuffer (MPSC) — commands from clients to Media Driver
• BroadcastTransmitter/Receiver — responses from Media Driver to clients
• Log Buffers — triple-buffered append-only log for data
• Position Counters — atomic counters for position coordination

Performance

Part 2: Aeron Archive — Record and Replay

Archive records message streams to disk for replay from any position. Use cases: regulatory audit, crash recovery, backtesting, debugging.

Part 3: Aeron Cluster — Raft Consensus

Fault-tolerant replicated state machine for systems where losing a message is unacceptable (matching engines, order management). Leader election, log replication, strongly consistent reads.

Part 4: Aeron Sequencer — Total Ordering

New product (2025) optimized for capital markets. Provides a single global ordering of all events — critical for matching engines and multi-venue market making. Built on Cluster. Commercial product (closed-source).

Weaknesses

1. JVM dependency — safepoints, GC pauses, JIT warm-up
2. Media Driver overhead — extra hop through shared memory
3. No native kernel bypass — UDP only (io_uring, DPDK not supported in open-source)
4. SBE is separate — XML schemas, Java code generator, separate build step
5. No zero-copy networking — data copied from socket to log buffer

Alternatives

We at Marketmaker.cc developed ZigBolt — an open-source Aeron alternative in Zig. No JVM, no GC, comptime codecs, 20 ns SPSC latency. Read more in our ZigBolt article.

Links:

• Aeron GitHub: github.com/real-logic/aeron
• Martin Thompson: mechanical-sympathy.blogspot.com
• ZigBolt (our alternative): article | site
• Marketmaker.cc: marketmaker.cc

Citation

@article{soloviov2026aeron,
  author = {Soloviov, Eugen},
  title = {Aeron: Inside the Messaging System That Powers Half of the HFT Industry},
  year = {2026},
  url = {https://marketmaker.cc/en/blog/post/aeron-messaging-overview},
  description = {Architecture deep-dive: Aeron Transport, Archive, Cluster, Sequencer.}
}