Skip to content
Last updated

Real Time 2.0 System Performance Limitations and Expected Behavior

Overview

Real-Time 2.0 is engineered for high-throughput event ingestion, low-latency personalization, and large-scale triggered activations. To ensure system stability and predictable performance, specific limits and Service Level Agreements (SLAs) are in place. This document outlines the expected behavior and performance characteristics when operating Real-Time workloads.Performance and Limits

CategoryLimit / BehaviorNotes
Event TablesUp to 100 event tables per Parent SegmentIncludes both batch and real-time event tables.
ID Stitching KeysUp to 100 IDs can be configured for stitchingExample: email, user_id, td_client_id.
Event DefinitionsNo hard limitCombined filter length must be ≤1000 characters.
Real-Time AttributesMax 200 attributes per profile (≤500 bytes each)Includes single, list, and counter attributes.
Ingestion Payload1–500 events per batch; 1MB per event, 5MB per batchApplies to Streaming Ingest API and SDKs.
Personalization APIResponse size: ≤10KB per requestLarger payloads may increase latency.
Ingestion ThroughputDefault: 2,000 events/sec per customerScales horizontally to 100k+ events/sec globally.
Decisioning Throughput8,000 events/sec per customerScales horizontally, expected upper bound 100x per customer.
Triggered Activations8,000 activations/sec per customerScales horizontally, expected upper bound 100x per customer.
Personalization API SLA≤100ms (p95) response time for up to 20 attributesEnterprise Support required for SLA guarantee.
Triggered Activation SLAUp to 3 minutes for activations to be deliveredIncludes evaluation, ID stitching, and activation dispatch.
Data SyncBatch → Real-Time: Supported. Real-Time → Batch: Not supportedRT attributes do not sync back into batch storage.
ScalabilityAll core components scale horizontallyVery large deployments (100+ Parent Segments) should validate in PoC.

Expected Behavior

  • Horizontal Scale-Out: Real-Time 2.0 automatically scales across infrastructure to manage traffic spikes. Extremely large deployments should undergo validation during a Proof of Concept (PoC) phase.

  • Latency Expectations: Personalization responses are nearly instantaneous, typically within 100ms (p95). Triggered activations may take up to 3 minutes end-to-end. Ingestion is designed to support sustained high throughput with linear scalability.

  • Data Sync Characteristics: Real-time and batch attributes are complementary, with data flowing exclusively from batch into real-time. Real-time attributes are not written back to batch storage.

  • Error Handling: All APIs are idempotent and retry-safe, allowing for safe reattempting of requests. If system limits are exceeded, errors will be returned with standard HTTP codes and clear messages.

Notes and Recommendations

  • Customers operating at extreme scales (e.g., 100+ Parent Segments or 10M+ events/sec) should plan staged rollouts to validate performance thoroughly.

  • If you anticipate exceeding any of the stated limits, please contact Treasure Data Support to discuss potential configuration adjustments or scaling options.

  • Certain advanced features, such as A/B testing within Real-Time Journeys and the deletion of attributes/profiles, are currently under development and may be subject to future changes.