SE Radio 717 — Eric Tschetter on Decoupling Observability

3 min read

I hosted Apache Druid co-founder Eric Tschetter on Software Engineering Radio to talk about breaking the observability stack into four layers — and what it takes to escape a coupled vendor stack.

I recently sat down with Eric Tschetter — co-founder of Apache Druid and Chief Architect at Imply — for episode 717 of Software Engineering Radio to dissect a shift I keep seeing in production systems: the move toward decoupling observability.

Also on Apple Podcasts.

Why decouple at all?

We started from first principles: the three pillars — logs, metrics, and traces — and why the rise of microservices has turned traditional, tightly coupled observability stacks into a source of real pain. When ingest, storage, query, and dashboards all live inside one vendor's box, you inherit vendor lock-in, scaling costs that grow prohibitive, and operational complexity that compounds with every new service you deploy.

The four-layer architecture

Drawing a parallel to how the business intelligence world separated its stack, Eric lays out an architecture with four distinct layers:

  1. Ingest / Route — get telemetry in and send it where it needs to go
  2. Data Storage — own your data, independent of any one tool
  3. Query / Compute — the engine that answers questions about it
  4. Visualization — whatever front-end suits the team

The framework's promise is flexibility: each layer can evolve — or be swapped — without dragging the rest of the stack along.

Data portability and OpenTelemetry

A big portion of the conversation is about what makes decoupling practical: data portability. Standards like OpenTelemetry standardize schemas so telemetry can flow freely between multiple back-ends instead of being held hostage by one ingest format.

Druid, and querying hot + cold data

Since Eric co-authored Apache Druid, we went deep on the Query/Compute layer: how Druid addresses the demands of real-time analytics on observability data, the indexing strategies involved, and how it unifies query results across hot and cold storage tiers.

Operational survival

The part I found most practical for anyone running this in anger:

  • Smart sampling that preserves high-value signals instead of blindly dropping data
  • Buffering and backpressure practices for when the pipeline gets swamped
  • Governance models that let multiple teams safely share one data lake

The honest trade-offs

We closed with a candid look at the complexity you take on when you decouple — this is not a free lunch — and a roadmap for organizations weighing a migration away from a coupled vendor stack.

About Eric

Eric is Chief Architect at Imply and one of the original authors of the open-source Apache Druid project, with experience across engineering and leadership roles at Splunk and Yahoo. He previously served as VP of Engineering at Metamarkets — where Druid originated — and was a founding team member of the diabetes data nonprofit Tidepool.

Related SE Radio episodes

If this topic is your thing, these pair well:

  • SE Radio 556 — Alex Boten on OpenTelemetry: telemetry interoperability, collectors, and the OpenTelemetry project
  • SE Radio 591 — Yechezkel Rabinovich on Kubernetes Observability: three pillars, eBPF, and observability costs
  • SE Radio 455 — Jamie Riedesel on Software Telemetry: foundational concepts of tracing, logging, and monitoring infrastructure
  • SE Radio 534 — Andy Dang on AI/ML Observability: data drift and production failures in ML applications
  • SE Radio 610 — Phillip Carter on Observability for LLMs: observability-driven development and debugging LLMs
Fear has killed more dreams than failure ever will.
© 2026 Amey Ambade
Houston, TX
currently obsessing over: the long catalog of bossa nova standards