A JVM runtime where the things that usually need a second process run in yours — at a measured, published resident cost.
Enterprise Java pays for its architecture twice. Once inside the process — abstraction layers, framework bookkeeping, classes loaded for features never used. And again around it — the saga coordinator, the event-store server, the APM sidecar: whole second processes that stay resident whether or not they are doing work.
Exeris removes both, and measures the removal.
- Exeris Kernel — an off-heap runtime kernel for the JVM (JDK 25 LTS baseline, preview-clean; a
previewline tracks the newest JDK). Request and response payloads live in PanamaMemorySegments behind deterministic buffer ownership; concurrency is virtual threads. Saga / durable execution (Flow), event sourcing (transactional outbox), and observability (JFR-first) run inside the application process — no coordinator, no event-store server, no sidecar. The runtime loads only the subsystems you declare: a plaintext deployment never carries the crypto subsystem's memory. - Exeris SDK + Tooling — the framework part of the platform, moved to build time. Annotated domain types (
@ExerisDomain,@Action,@Field,@Relationship...) are turned by the codegen pipeline into kernel-native handlers, sagas, OpenAPI, database migrations and UI components — idiomatic, editable Java, validated before the first byte of traffic. What a runtime framework does with reflection and auto-wiring on every startup happens here once, in the build — which is part of why the resident process stays small. Exeris Studio (roadmap) syncs visually with the same sources via LSP: visual speed without low-code lock-in. The same entity-first, build-time-verified pipeline is the substrate we are building toward AI-assisted generation of clean, reviewable code. - Exeris Spring Runtime — hosts existing Spring applications on the kernel. Spring keeps DI, configuration and beans; the kernel owns ingress, lifecycle and the data plane. Brownfield adoption without a rewrite.
- Exeris Benchmark Lab — the open harness behind every number we publish: matched contracts, fail-closed gates, AB/BA order control, committed raw artifacts, per-claim eligibility stamps.
Above the substrate, a composable capability ecosystem and vertical SKUs are in design — the whitepaper covers the three-tier picture.
Every comparative figure we publish carries a scope label (comparison_eligible = passed the strict fairness gate on dedicated bare metal; descriptive / exploratory = directional), a scenario contract, and raw artifacts committed next to the report. A few from the current series:
| What was measured | Result | Scope |
|---|---|---|
| Runtime-bound single-row read vs tuned pure-JDBC Quarkus and idiomatic Quarkus+Hibernate | +39% / +57% throughput at −26% / −34% CPU per request | comparison_eligible · report |
| How small a memory budget the same work fits in | Full speed in a 128 MiB budget; the comparison stack's declared heap policy does not boot below 192 MiB | descriptive · report |
| Where a Spring request's cost actually sits (one app, five hostings) | The repository layer costs headroom, not per-request latency — and the largest identified contributor is projection proxies, not the ORM's row mapping | mixed, stated per table · report |
| What declaring 155 Valhalla value classes buys on an off-heap runtime | ≈1.3 MB of permanent non-heap cost, no workload movement — published although it cuts against our own sweep | within-tier · report |
When a number loses, we retract it in public: the revision histories in those reports are load-bearing, not decorative. Review us — with or without an AI assistant; we use both — and tell us what you find. The artifacts are committed precisely so that a reviewer doesn't have to take the prose on trust.
Deep R&D, pre-1.0. The kernel is on the v0.12 line; API stability declarations have been honored since v0.9 (verified with japicmp across the full tag history). Kernel SPI, Core, the Community driver and the TCK are open under Apache-2.0; the Enterprise driver (io_uring / IOCP transport, QUIC/HTTP-3, NUMA-aware slab pools) is a commercial implementation of the same SPI — a Maven-coordinate swap, not a fork of your application.
We don't build golden cages. Business artifacts compile against the SPI, and the SPI is the contract: code detachment — taking your generated business logic and hosting it on your own infrastructure — is a documented, priced path, not a negotiation. A platform you can leave is a platform you can trust with production.
- 🌍 Website: exeris.eu
- 📊 Benchmark reports: exeris-benchmarks/results/reports
- 📩 Contact: github@exeris.eu
- 💼 LinkedIn: Exeris Systems
