Temporal logic is native
Windowed counts, filtered aggregates, distinct sets, and sequences live inside the serving engine instead of being scattered across stream jobs, cache keys, and application code.
TemporalStore
Temporal decision-state serving for risk, recommendations, and AI context.
TemporalStore is the MatrixArk engine for high-cardinality entity state that changes over time: rolling aggregates, filtered windows, distinct sets, behavior sequences, frequency caps, and structured context that must be answered inside the online request path.
Product Advantages
Why TemporalStore over a generic cache plus precomputed streams?
Lower pipeline complexity
Teams can serve many risk and feature policies directly from online state rather than building a new stream job and cache layout for every window.
High-cardinality friendly
The architecture is shaped around entity keys, model-aware objects, in-memory indexes, logs, and persisted pages for large online state.
Fresh online decisions
Risk, fraud, ads, recommendation, and agent systems can query recent state at request time instead of waiting for offline materialization.
System Architecture
Feature API / SDK
Router
Partition worker
Model engine
WAL / oplog
Page + index store
TemporalStore keeps model-aware temporal state close to the serving path. Entity updates are routed to the right partition, applied through the configured data model, persisted for replay, and queried through APIs built for time windows, filters, distinct state, sequences, counters, and context assembly.
Serving Workflow
- Ingest an event, counter update, behavior append, or context record.
- Resolve namespace, table, data model, partition, and entity key.
- Apply updates inside the model instead of in a separate application layer.
- Answer window, filter, aggregate, distinct, context, and sequence queries online.
Scenario Architecture
One serving path for event ingestion and temporal feature reads.
Ingest
Application events, stream consumers, and repair jobs write the same entity timeline or aggregate object.
Store
Partition workers update hot objects, append logs, and maintain page/index state for replay and persistence.
Serve
Online systems request windows, filters, distinct counts, behavior sequences, context, and recent entity state through one serving API.
Cloud-Native Operations
Separate the public website, customer console, and private metrics plane.
A test deployment can host the marketing site and observability pages on one node and one port. A production TemporalStore deployment should split those concerns: the public website stays public, the console is authenticated, and metrics endpoints stay private inside the customer or MatrixArk VPC.
Public website
Marketing, product pages, blogs, and documentation are served separately from runtime operations.
Authenticated console
Cluster health, node inventory, partition state, cache usage, scale tests, and deployment status live behind customer access control.
Private metrics
Prometheus-compatible metrics, logs, data-node variables, and storage signals stay on private networking or controlled observability integrations.
ASG node registration
New data nodes launched by autoscaling register capacity with the metaserver; the metaserver assigns partitions and replicas.
Graceful scale-down
Before an instance terminates, it should enter draining state so primaries move away and replicas catch up.
Versioned runtime packages
Release packages should include binaries, dynamic libraries, service files, health checks, registration scripts, and rollback metadata.
Data Models
Built for temporal online decisions.
TemporalCounter
Counts and sums over keyed time buckets for velocity, caps, and online policies.
TemporalAggregate
Filtered sum, min, max, count, and model-specific rollups over recent event state and bucketed dimensions.
TemporalDistinct
Unique merchant, device, campaign, IP, or session counts inside time windows.
Sequence
Recent behavior history, sorted timelines, long user action sequences, and inference features.
Hash/Profile
Latest user, item, merchant, campaign, or device state beside temporal features.
Context
Structured LLM and agent context, tool events, session memory, retrieval metadata, and safety/rate counters.
Comparison
| Alternative | Good at | TemporalStore difference |
|---|---|---|
| Redis / Redis Enterprise | General cache, strings, hashes, modules | Model-aware temporal windows, filters, sequences, and persistence-first online features. |
| Feast online store | Latest feature lookup and registry-driven materialization | Native online temporal APIs and storage-side aggregation logic. |
| Flink + cache | Streaming pre-aggregation | Less pipeline sprawl when business asks for ad hoc windows and entity-specific filters. |
| OpenMLDB / feature SQL | Feature computation through SQL and pre-aggregation | Serving-engine design for hot high-cardinality windows and sequence reads. |
Where TemporalStore fits
TemporalStore is the product engine for temporal online state: event ingestion, rolling windows, filtered aggregates, distinct counts, behavior sequences, frequency caps, persistence, metrics, and customer-facing SDK access.
Talk to MatrixArk