High-Throughput Event Streaming: Processing 18,000 Events/sec in Real-Time

Sub-Second Contextual Decisioning Under Peak Load

In mobile telecom networks, customer context shifts continuously. Network events—such as balance exhaustion, top-up transactions, roaming handovers, and USSD queries—generate high-volume transaction bursts requiring instantaneous evaluation.

Event Pipeline Architecture Working across tier-1 operators with SAS Middle East and Telenor, we engineered real-time decisioning pipelines capable of sustaining 18,000 events per second:

1. 1. **Ingestion & Protocol Normalization**: Asynchronous listeners built with Java and Apache Camel ingest network probes and CDR streams over TCP/IP sockets.
2. **Stream Filtering**: SAS Event Stream Processing (ESP) aggregates rolling time-windowed metrics and discards redundant events before forwarding to decision engines.
3. **Decision Arbitration**: SAS Real-Time Decision Manager (RTDM) evaluates rule matrices, customer eligibility, and propensity scores to select the single most relevant offer.
4. **Channel Dispatch**: Sub-second dispatch to USSD gateways, SMSCs (via SMPP protocol), and mobile app push listeners in under 300 milliseconds.

Core Design Decisions - **Non-Blocking I/O**: Utilizing asynchronous socket processing (Apache MINA) to prevent thread exhaustion during network surges. - **In-Memory Caching**: Redis clusters caching subscriber profile snapshots for fast sub-millisecond lookups. - **Deterministic Fallbacks**: Circuit breakers ensuring default offers are returned immediately if downstream predictive scoring models exceed latency SLAs.