Deterministic step budget · 2 MB scored in ~3.4 ms · 947-byte edge bundle, no database · Blue/green rollback
See how it compares to ODMPlatform/Decision Server
Decision Server is the hot path — where rule versions authored in Decision Manager are deployed, exercised in the API workbench or probe, executed against live traffic, and measured.
Get started
Runtime & observability
The hot path: deploy, execute, and measure decisions in production.
Capabilities
Real operator surfaces: API workbench, probe, deploy matrix, warehouse — deploy, execute, measure without owning the policy text.
A deployment matrix across environments, with sub-second hot-deploy, blue/green slots and sticky A/B — no application redeploy to change a rule.
Compiled multi-key indexes on a Rust service over PostgreSQL. Operators use the API workbench (collections, environments, assertions, multi-language snippets — decision-native) and the probe console; production traffic uses the same execute path.
Execution KPIs, the decision warehouse browser, and writer health — so a decision can be explained after the fact. Published latency: ~3.4 ms end-to-end for a 2 MB request against 500 candidates (including parse).
Decision engine · XU honesty
Live — Shipped and available in product todayDecision Server runs a compiled single-pass engine with multi-key equality indexes, residual_rules, hit policies, and a hard step budget. We name residuals so leads never confuse Live product with classic ODM XU marketing.
These classic ODM/XU surfaces are residual. DecisionManager does not ship them as product:
Performance & high availability
Live — Shipped and available in product todayFigures come from the product binaries under documented conditions — not customer averages or freestanding sub-2 ms floors.
~3.4 ms
End-to-end execute
2 MB request · 500 candidates · includes parse
Sub-second
Hot-deploy path
Blue/green slots & sticky A/B · no app redeploy
~20 ms
Cold start (edge)
Debug build · Postgres stopped · method stated
Every decision stays inside a deterministic step budget; residual_rules keep non-indexed paths honest and measurable.
Process admission and L1 deploy cache protect the hot path under load; cache invalidation follows version publish.
Hot prewarm/swap and canary/blue-green without restarting the application that calls execute.
Idempotent batch execute and SERVICE_BUSY shed when capacity is exhausted — Live signals, not invented autoscalers.
Residual (not Live): multi-algo Rete/GPU/JIT product, multi-region active-active mesh, closed-loop auto-rollback. Named so enterprise buyers can plan honestly.
Observability
Live — Shipped and available in product todaySee what fired, how long it took (engine.steps on every response), and writer health — so a production decision is explainable after the fact. We publish ~3.4 ms (2 MB / 500 candidates), not a freestanding sub-2 ms average.
Warehouse matched[] traces, topology health, SLO burn, and rule telemetry — explain a production decision after the fact without inventing multi-XU JMX maps.
API workbench
Live — Shipped and available in product todayOperators exercise collections, environments, assertions and multi-language snippets against the real execute API — plus probe and idempotent batch — without a separate mock server.
Workbench is included with Decision Server on every plan — same runtime production traffic uses.
Go deeper
Dedicated feature pages for the capabilities that usually decide a BRMS shortlist — with Live / Residual honesty carried through.
Single-pass indexes, hit policies, residual_rules — and residual Rete/JCA named plainly.
Open feature pageMeasured ~3.4 ms path, admission, deploy cache, hot/canary — residual multi-region stated.
Open feature pageWarehouse, telemetry, topology, SLO burn — post-hoc explanation without invented badges.
Open feature pageContainers, K8s, hash-verified edge bundles — account-scoped edge, not a global multi-region mesh we don't operate.
Open feature pageAudience
Platform and SRE teams: the people who own uptime, latency and rollback. Rule authors do not need it to change a rule — that happens in Decision Manager and arrives here as a published version.
Migration
It replaces the ODM execution tier: RuleApp deployment cycles onto a Java EE rule execution server running on WebSphere Liberty or Tomcat, where a bad version is reverted by redeploying the previous RuleApp and restarts carry risk. Decision Server is a stateless container over one PostgreSQL database, with no provider-specific code.
The latency and architecture comparison used above is the same one on the home page battlecard; the execution API is documented without an account.
Authoring is one of three consoles over the same repository.
Get started
Self-serve trial includes Decision Server and the API workbench — no card, no auto-convert. Prefer a seller who can map ODM XU residual to your estate? Book a meeting.
No credit card · Trial does not auto-convert · Residual Rete / JCA stated on this page