Skip to content

Aeron® Open Source vs Premium: When the Commercial Stack Pays

Transport, Archive and Cluster — everything you need to run a replicated matching engine — are open source in the Aeron® repository under Apache-2.0. Aeron Premium, Adaptive’s subscription offering, adds seven components around that core; three of them change what your system can do: Kernel Bypass (DPDK) when you hit the kernel drivers’ packet-rate ceiling, Cluster Standby when you need cross-AZ or cross-region recovery without putting Raft on the WAN, and Aeron Transport Security (ATS) when traffic crosses a network you don’t trust. If none of those three describes your system, the open-source stack is enough, and the money is better spent on the application layer.

NeedOpen-source pathPremium componentWhat changes (measured on this site)When open source is enough
Sustain a high datagram rate with a flat tailJava or C media driver, batching, host tuningKernel Bypass (DPDK)AWS measured ~1.4× on cluster P99 at 100k msg/s; past the kernel knee, the gap is saturation, not speed — DPDK guideRate comfortably under the packet ceiling, or traffic that batches
Survive an AZ or region lossRaft across AZs, or ClusterBackup (cold copy + replay)Cluster StandbyWarm, non-voting standbys; Raft stays in one AZ — Standby & Multi-AZ HAAn RTO of minutes to hours (log-size dependent) is acceptable
Snapshots that don’t pause the leaderSnapshot on every node (cluster-wide pause)Cluster Standby (standby snapshots)Snapshot taken off the voting members — Snapshot failure & recoverySnapshot pause fits inside your max-latency budget
Replace a node or move regions liveStatic membership; rebuild from ClusterBackupCluster Standby (transition-as-member)Follower swaps: 0 missing messages (single host) — Cross-region migrationYou can take a maintenance window
Encrypt and authenticate trafficNetwork-layer encryption (IPsec, WireGuard) or an isolated networkAeron Transport SecurityAES-256-GCM per frame, RSA-authenticated drivers; no latency figure published or measured hereTrusted, isolated network
Dashboards and alertsOwn exporter over the CnC countersAeron Insights (beta)A curated subset of the counters, packaged; no election-count or leadership-term metric — Metrics & alertingYou already run Prometheus and can own the exporter
Bound archive disk useManual purge disciplineData Retention RegulatorPolicy-driven retention — Archive retention & purgePurge is a scheduled, reviewed runbook
Many low-rate streams without burning coresSleeping / yielding idle strategyAeron Selectorepoll-based wake-up instead of polling; no figure publishedA few streams, or cores to spare

Kernel Bypass (DPDK): the packet-rate ceiling

Section titled “Kernel Bypass (DPDK): the packet-rate ceiling”

Adaptive ships Kernel Bypass as pre-compiled bundles of the C media driver with DPDK, launched as aeronmd_dpdk. Its documented limits matter before its speed does: IPv4 only, unicast only, one DPDK port per driver, and the conductor, sender and receiver pinned to separate cores in DEDICATED threading mode. A design built on multicast cannot use it. The setup guide expects the NIC bound to vfio-pci and hugepages reserved, and lists tested poll-mode drivers net_ena (AWS), net_gve (GCP) and net_ixgbe (Intel).

What it buys, from AWS’s 2025 benchmark on c6in.16xlarge (cluster placement group, MTU 1408), as worked through in Reading the Benchmark Honestly:

Aeron Cluster, P99OSS JavaOSS CPremium (DPDK)
100k msg/s136 µs139 µs98 µs
1M msg/s8,479 µs8,577 µs143 µs

At 100k msg/s the gap is ~1.4×. At 1M msg/s the kernel drivers have saturated and DPDK has not; the headline “59×” measures that, not a faster datapath. For the Transport (non-cluster) path, OSS Java vs DPDK P99 is 32 µs vs 29 µs at 100k and 57 µs vs 39 µs at 1M.

The price, covered in the DPDK decision guide: dedicated cores at 100% busy-spin, reserved hugepages, a dedicated ENI whose bind a reboot loses, and a failure mode where a mis-set bind degrades latency silently, with no error.

Open source is enough when your datagram rate sits well under the kernel ceiling, your traffic coalesces into datagrams, or your latency is spent in matching and risk logic rather than in the datapath.

Cluster Standby: HA without Raft on the WAN

Section titled “Cluster Standby: HA without Raft on the WAN”

A standby is a consensus-module variant that receives the cluster’s log, runs your clustered service on it, and does not vote: Adaptive’s docs say it “won’t vote in a leader election” and applies no back-pressure to the cluster’s log. That one property drives everything else:

  • Quorum stays local. Put the three voting members in one AZ and the standbys in another; commits never wait on a cross-AZ ack. The open-source alternative is spreading Raft across AZs, which puts that round-trip on every commit — see Standby & Multi-AZ HA.
  • Snapshots move off the voters. PremiumClusterTool standby-snapshot (triggered via the leader) takes the snapshot on the standby; a restarting member fetches newer standby snapshots before loading its own. Both standbySnapshotsEnabled and acceptStandbySnapshots default to false — turn them on or you have paid for the feature without using it.
  • Node replacement is a command, not a rebuild. transition-as-member promotes a standby into a failed member’s ID; members must be addressed by hostname so the name can be repointed (DNS, /etc/hosts or a custom resolver). This site’s migration runbook measured it on a single host (3-node cluster, no WAN): follower swaps lost 0 messages. Promoting a standby into the leader’s slot did not re-converge on that harness, so treat that step as unproven until you have demonstrated a clean crossover on your own rig.
  • Live membership change is Premium-only now. The open-source changelog deprecated dynamic join in 1.41.0 (“to be replaced with a more robust and user friendly premium offering”) and removed the remaining APIs in 1.47.0. In open source, the members list is static.

Open source is enough when an RTO of minutes to hours is acceptable: ClusterBackup replay time depends on log size. ClusterBackup copies the log and snapshots from a running cluster; rebuilding a member means copying that archive and recording.log to a node with the same clusterMembers entry and replaying. The Backups & DR comparison lays out warm-vs-cold side by side. Snapshot safety is identical in both editions — an interrupted snapshot recovers the same way, and the integrity tooling (invalidate-latest-snapshot, validate-recording-log) is open source.

Aeron Transport Security: frames on an untrusted network

Section titled “Aeron Transport Security: frames on an untrusted network”

The open-source SECURITY.md is blunt: “Aeron does not authenticate or encrypt transport traffic by default”, and on untrusted networks, “including most cloud deployments”, it points to ATS or your own IPsec/WireGuard. ATS encrypts each datagram with AES-256-GCM (header sent in the clear as authenticated data), authenticates drivers by RSA key pair, and uses ECDH for unicast keys. It covers unicast, multicast and MDC.

Constraints that decide feasibility: C media driver only (the Java driver errors on the ats URI parameter), OpenSSL 3.0.0, a unique RSA private key per driver, and loading via LD_PRELOAD or AERON_DRIVER_DYNAMIC_LIBRARIES. ATS and non-ATS endpoints cannot see each other on the same channel, so roll it out per channel (ats=false opts out). Adaptive publishes no latency figure (its overview says only “minimal latency”) and this site has not measured one — measure p99 with and without it before committing.

Open source is enough when the network is trusted and isolated, or a network-layer tunnel’s latency cost fits your budget. Archive and Cluster authentication and authorisation are open source, but not off. By default authentication accepts every connection (DefaultAuthenticatorSupplier) and the Archive’s authorisation allows every command. The Cluster’s default authorisation (AllowBackupAndStandbyAuthorisationService) allows only the requests Cluster Backup and Standby need, so client admin requests to the consensus module are refused. Your application commands arrive as ingress messages and never reach it; the clustered service must authorise them itself. Configure an authenticator and an authorisation service either way.

Aeron Insights, Data Retention Regulator, Selector

Section titled “Aeron Insights, Data Retention Regulator, Selector”

These three change effort, not latency.

  • Aeron Insights (beta) ships Prometheus exporters, a sample Grafana dashboard and a CLI. It reads the same CnC file an open-source exporter reads, but exports a curated subset of it: notably no metric for cluster election count, leadership term or append position, the most useful stability alerts, as covered in Metrics & Alerting. Run your own exporter for those counters alongside it.
  • Data Retention Regulator applies retention policies to Archive and Cluster recordings — for example RetainEnoughDataForRecoveryPolicy keeps the last N snapshots and the log after the earliest of them. Its default backup strategy does nothing. The open-source equivalent is the manual workflow in Archive Retention & Purge.
  • Aeron Selector is a Linux daemon (aeronsd, epoll-based, media driver 1.43.0+) that lets an application thread sleep on many subscriptions instead of busy-spinning. Adaptive states the wake-up latency is the trade. The open-source lever is the idle strategy — see Smart Batching & Idle Strategy.

Adaptive also lists SBE Domain Mapper (Java 17+, generates SBE ↔ domain-object adapters). It is a developer-productivity tool with no operational effect, so it doesn’t belong in this decision.

  1. Is your first wall packets per second, with smooth arrivals you can’t batch and unicast IPv4 only? → evaluate Kernel Bypass, and budget dedicated isolated cores per driver.
  2. Must the cluster survive an AZ or region loss faster than a ClusterBackup replay, without a cross-AZ ack on every commit? → evaluate Cluster Standby, and design for async-replication data loss.
  3. Does the snapshot pause show up in your max latency? → standby snapshots; verify both standbySnapshotsEnabled and acceptStandbySnapshots are set.
  4. Does traffic cross a network you don’t control? → ATS (C driver) or a network-layer tunnel; measure p99 either way.
  5. None of the above? → stay on open source. Write the exporter, script the purge, tune the idle strategy.

Adaptive sells Premium on subscription with what it calls “SLA-backed support”, and lists it on the AWS, Google Cloud and Azure marketplaces. It publishes no prices or SLA terms, so this page has none.

This site is not affiliated with, endorsed by, or sponsored by Adaptive Financial Consulting Limited or the Aeron project. Aeron is a registered trademark of Adaptive Financial Consulting Limited.

Aeron is a trademark of Adaptive Financial Consulting Limited in the United Kingdom and other countries.