CONCEPT Cited by 1 source
Single-sink telemetry architecture¶
Single-sink telemetry architecture is the structural posture of routing observability traffic (traces / logs / metrics) directly from instrumented clients to a single durable analytical store through one managed ingestion endpoint — collapsing away the conventional intermediate-broker hop (Kafka / Pulsar / Kinesis) that traditional multi-hop telemetry pipelines use.
What it replaces¶
Conventional telemetry pipeline:
Single-sink shape:
The defining property: one hop, one schema boundary, one operational system to own.
Canonical instance: Databricks Zerobus + UC OTel Trace Tables (2026-05-22)¶
The 2026-05-22 OTel-tracing launch frames the architecture explicitly:
"With a 'single-sink' architecture, Zerobus Ingest simplifies observability by streaming data directly to the lakehouse. Existing OLTP-compatible collectors can point directly to this endpoint via gRPC, entirely bypassing intermediate message buses like Kafka." — Source: sources/2026-05-22-databricks-observability-any-agent-anywhere-otel-unity-catalog
Zerobus Ingest is the managed receiver; UC OTel Trace Tables are the storage destination; the single-sink claim is the absence of a broker between them.
Structural arguments for single-sink¶
| Argument | Detail |
|---|---|
| Fewer hops | Each hop is a failure-mode / latency / throughput-bottleneck candidate. Removing one collapses the surface area. |
| One schema boundary | Multi-hop pipelines often re-serialize at each hop (OTel proto → Avro → Parquet). Single-sink lets the receiver write the storage format directly. |
| Operational simplification | No broker to size, scale, secure, patch, monitor, upgrade. Managed ingestion absorbs that. |
| Lower latency to query | Trace ingest → queryable storage in one path; broker buffering removed. |
| Drop-in compatibility | OTLP-compatible collectors that already exist can re-point at the new endpoint without re-architecture. |
Structural arguments against (when single-sink is the wrong shape)¶
- Multi-destination fan-out. If telemetry must land in two places (lakehouse for analytics + APM for real-time alerting + cold archive for compliance), a single sink can't satisfy all destinations. Brokers fan out cleanly; single-sink doesn't.
- Buffering for back-pressure. Brokers absorb bursts that overwhelm downstream consumers; single-sink pushes back-pressure to the client (OTel SDK retries, client-side buffer, or drops).
- Cross-team aggregation. Brokers are natural rendezvous points where multiple producers and multiple consumers meet; collapsing the broker collapses that property.
- Replay. Brokers retain events for replay; a write-direct-to-storage path requires the storage layer itself to support replay (Delta CDF / Iceberg time travel can — but the property has to be designed in, not assumed).
When the single-sink trade is favourable¶
The 2026-05-22 architecture argues this for agent traces specifically:
- One destination is the answer — UC Delta is the lakehouse the customer already governs; alternative destinations are not part of the design intent.
- Burst tolerance comes from managed receiver — Zerobus is "fully managed, serverless" with the implication that the platform absorbs spikes, though the post does not benchmark burst behaviour.
- Replay is provided by storage — Delta + CDF gives the same replay property a Kafka topic would, on the storage side.
- Cross-team aggregation lives in UC — the lakehouse is itself the aggregation point; teams query the same tables.
The trade favours single-sink when the lakehouse is already the canonical analytical destination and when the workload is not real-time alerting (which still wants APM / Prometheus / sub-second tooling alongside).
Relationship to other concepts¶
- Specialisation of concepts/lakehouse-native-observability: lakehouse-native observability says "observability lives in the lakehouse"; single-sink architecture is the ingestion-side discipline that gets it there.
- Composes with concepts/instrumentation-storage-decoupling: OTel-as-protocol-boundary lets clients emit anywhere; single-sink architecture decides where they land.
- Sibling to patterns/streaming-broker-as-lakehouse-bronze-sink (the opposite shape) — that pattern uses a broker as the bronze-tier sink. Single-sink architecture is the no-broker alternative when broker-as-bronze isn't needed.
- Operationalised by patterns/managed-otel-ingestion-direct-to-lakehouse — the canonical pattern instance.
Caveats¶
- The marketing argument is "fewer hops". The structural argument is "fewer systems to operate". These are correlated but not identical — a single highly-complex managed receiver can be operationally heavier than two simple components, depending on where the operational burden actually lands.
- "Bypassing Kafka" is not always a win. Kafka exists in many telemetry pipelines for legitimate reasons (multi-consumer fan-out, durable replay, cross-team rendezvous). Single-sink architecture works when those reasons don't apply for the specific workload.
- The claim is rarely benchmarked. Vendor framings (Zerobus, similar managed-ingest products) emphasise architectural elegance over throughput / latency comparisons.
Seen in¶
- sources/2026-05-22-databricks-observability-any-agent-anywhere-otel-unity-catalog — first wiki disclosure; the "single-sink" term is named verbatim, the Kafka-bypass argument is explicit, and the canonical instance (systems/zerobus-ingest + systems/uc-otel-trace-tables) is the substrate for the entire 2026-05-22 product launch.
Related¶
- concepts/instrumentation-storage-decoupling — sibling concept (client-side decoupling vs ingest-side simplification).
- concepts/lakehouse-native-observability — the architectural posture single-sink enables.
- concepts/observability — parent concept.
- systems/zerobus-ingest — canonical instance.
- systems/uc-otel-trace-tables — destination of the canonical instance.
- systems/opentelemetry — wire protocol.
- systems/unity-catalog — governance substrate.
- systems/delta-lake — physical storage format.
- patterns/managed-otel-ingestion-direct-to-lakehouse — canonical pattern.
- patterns/telemetry-to-lakehouse — generalised pattern.
- companies/databricks