The Data Flowcast: Mastering Apache Airflow ® for Data Engineering and AI

How do you orchestrate AI workflows when LLM outputs are non-deterministic and evaluation costs can quietly exceed compute costs? Shawn Feng, Head of Data at [Firework](firework.com), joins the show to walk through how his team uses Airflow to coordinate ingestion, embeddings, and evaluation pipelines for AI features in a video commerce platform. The conversation covers the technical stack, the challenges of validating LLM output, cost guardrails, and what Shawn wants to see next from the Airflow project.

Key Takeaways:
00:00 Introduction.
00:48 What Firework does: a video commerce platform bringing short-form shoppable video and livestream experiences directly onto brand websites and apps.
02:13 Shawn's team owns the full data and AI platform stack at Firework, from ingestion through BI and applied AI.
02:59 Airflow has been the main orchestration layer for batch transformations since the early open source days, and now powers AI workflows like conversational insights, knowledge-base building, and automated eval pipelines.
06:45 The AI workflow stack: ingestion of user interaction and content data, transformation into personalized profiles and embeddings, separate eval pipelines dispatched from third-party systems, all on Snowflake with Cortex AI and coordinated by Airflow.
08:55 Why Airflow stuck: flexibility, reliability, and clear visibility into dependencies across SQL, Python, and AI tasks.
11:27 The deterministic pipeline problem. Airflow assumes predictable input and output. LLM workflows break that assumption, which forces evaluation layers, traceability, regression testing, and feedback loops into the pipeline itself.
13:07 Evaluation cost can exceed compute cost. Validating a single output can mean running four or five parallel eval jobs across prompts and configurations, so caching, batching, and selective evaluation become essential.
15:58 Building safeguards: alerting and aborting jobs when cost exceeds thresholds, so the platform stays operationally stable while other improvements catch up.
18:30 Wishlist: stronger first-class support for AI workflow patterns (evaluation tracking, prompt experimentation, model observability, event-driven AI orchestration) and continued UI/UX improvement in Airflow 3.

Resources Mentioned:
[Firework](firework.com)
[Apache Airflow](airflow.apache.org)
[Snowflake Cortex AI](snowflake.com/en/data-cloud/cortex)
[Airflow common.ai provider](airflow.apache.org/docs/apache-airflow-providers-common-ai/stable/index.html)
Shawn Feng on LinkedIn (https://www.linkedin.com/in/shawnshifeng/)

Thanks for listening to "The Data Flowcast: Mastering Apache Airflow® for Data Engineering and AI." If you enjoyed this episode, please leave a 5-star review to help get the word out about the show. And be sure to subscribe so you never miss any of the insightful conversations.

#AI #Automation #ApacheAirflow

Saks Global runs one of the largest retail data operations in the US, with around 8 million SKUs flowing across point-of-sale, e-commerce, catalog, and fraud detection systems into Snowflake. In this episode, [Shailesh Kadam](linkedin.com), Architect at [Saks Global](saks.com), joins Kenten to walk through how Airflow acts as the nervous system tying it all together, why they moved from self-managed Kubernetes to Astro, what is driving their Airflow 3 upgrade, and how they are approaching agentic AI, MCP, and credential security.

Key Takeaways:
00:00 Introduction.
01:31 Saks Global today. Shailesh describes the business after separating e-commerce from brick and mortar and acquiring Neiman Marcus, and the modern cloud-native stack on AWS, Snowflake, and Airflow.
02:50 8 million SKUs in motion. Why every name, image, inventory, and price change has to flow in near real time across operational systems.
04:50 What the pipelines look like. Point-of-sale ingestion, fraud signals to third parties like Fiserv, and hourly product catalog feeds out to Meta and Google.
07:30 Moving off self-managed Kubernetes to Astro. Shailesh contrasts past experience with Kubernetes, IBM Tivoli, and Control-M against running on Astro.
09:35 Upgrading to Airflow 3. Event and asset-based scheduling, DAG versioning, task isolation, and using Otto to convert DAGs in a phased rollout.
13:13 Agentic AI and MCP on the roadmap. How Saks plans to use Airflow's MCP for LLM-driven product classification and to feed Snowflake analyses like churn and spend.
18:01 Securing PII and credentials. Secrets backends, cloud secret manager integration, key rotation, and keeping credentials out of DAG code.
21:02 Wishlist for Airflow. Interactive data lineage across DAGs and a UI-based debugging interface for support teams.

Resources Mentioned:
[Apache Airflow](airflow.apache.org)
[Astro](astronomer.io/product)
[Otto, the Astronomer data engineering agent](astronomer.io)
[Snowflake](snowflake.com)
[Saks Global](saks.com)

Thanks for listening to "The Data Flowcast: Mastering Apache Airflow® for Data Engineering and AI." If you enjoyed this episode, please leave a 5-star review to help get the word out about the show. And be sure to subscribe so you never miss any of the insightful conversations.

#AI #Automation #Airflow

Airflow 3.3 is here, with a set of features to help with the messy realities of production pipelines: persisting state across retries, reacting intelligently to different failure types, and partitioning assets by more than just time. In this episode, Marc Lamberti, Education Content Lead at [Astronomer](astronomer.io), joins Kenten Danas to walk through what's new in the release and where each feature actually pays off.

Key Takeaways:
00:00 Introduction.
01:46 The new task state store (AIP-103) lets tasks persist state across retries, so a long-running Spark job can be reattached after a worker failure instead of being duplicated on retry.
03:46 The asset state store enables watermarking patterns: persist the last processed date or offset to an asset and resume from there on the next run.
05:33 Why this matters for agentic workflows: resume an agent from where it left off rather than replaying every action.
06:58 Why XComs don't solve this problem: they get reinitialized on every retry.
09:27 Pluggable retries let you attach a retry policy to a task that branches on the exception type. Retry on transient errors, stop immediately on a 403.
11:42 Subclassing the retry rule for more complex logic, including dynamic retry counts that used to require hacking the metadatabase.
15:52 Updates to asset partitions in 3.3: segment-based partitioning with fan-out and roll-up mappers for downstream DAGs.
21:42 Running tasks in Java and Go, moving Airflow toward a multi-language orchestrator.
23:33 DAG versioning improvement: choose whether a manual rerun uses the most recent DAG version or the original version from that run.
25:03 Advice for teams still on Airflow 2: use the upgrade ebook and Astro's AI migration tooling to handle the undifferentiated heavy lifting.

Resources Mentioned:
Astronomer
Airflow 3.3 Release Notes
The Task State Store
Updates to the asset partitions feature
Retry policies
Multi-language support
3.3 Webinar
Upgrading from Airflow 2 to 3 ebook

Thanks for listening to "The Data Flowcast: Mastering Apache Airflow® for Data Engineering and AI." If you enjoyed this episode, please leave a 5-star review to help get the word out about the show. And be sure to subscribe so you never miss any of the insightful conversations.

#AI #Automation #Airflow

Running Apache Airflow at a major pension fund means balancing strict compliance requirements with the need to move fast on new capabilities. On this episode, Kowsy Narayan, Cloud Data Platform Lead, Data Engineering at [Ontario Teachers' Pension Plan](otpp.com), joins host Kenten Danas to walk through OTPP's cloud migration, their move to Airflow 3, and going fully live on remote execution.

Key Takeaways:
00:00 Introduction.
01:18 Inside the OTPP data platform team and what they're responsible for across cloud migration, standards, and enablement.
02:33 What's driving OTPP's multi-year move off on-prem to a cloud architecture built around scalability and resilience.
02:57 The new stack: Snowflake as the enterprise data platform, dbt for transformation, and Airflow as the orchestrator in the middle.
04:15 Why OTPP chose Astronomer: active contributions to the Airflow OSS project, fast runtime releases, and built-in monitoring, observability, and RBAC.
05:50 Evolving from dbt core with Bash operators to dbt Cosmos for model-level granularity, lineage, and precise failure recovery, plus a performance boost from watcher mode.
08:00 Upgrading from Airflow 2.9 to Airflow 3, using the Astro CLI and linters to catch deprecations quickly.
09:32 The drivers behind adopting remote execution: keeping data inside the security perimeter and scaling workloads on their own Kubernetes cluster.
11:35 How remote execution replaced a complex network architecture of VPN tunnels and firewall rules, removing latency along the way.
12:53 The POV process, success criteria, and a six week timebox to validate remote execution before going to production.
14:14 Going fully live: OTPP's last hosted deployment was sunset just before recording.
15:06 What Kowsy wants next from Airflow: AI orchestration capabilities and continued maturation of remote execution.

Resources Mentioned:
[Ontario Teachers' Pension Plan](otpp.com)
[Apache Airflow](airflow.apache.org)
[Astronomer](astronomer.io)
[Cosmos](astronomer.io/cosmos)

Thanks for listening to "The Data Flowcast: Mastering Apache Airflow® for Data Engineering and AI." If you enjoyed this episode, please leave a 5-star review to help get the word out about the show. And be sure to subscribe so you never miss any of the insightful conversations.

#AI #Automation #Airflow

Skimlinks runs a reporting platform that serves around 2,000 weekly publisher users, and the data infrastructure behind it runs on Airflow. In this episode, Julian Larralde, Director of Data Engineering at Skimlinks, walks through the stack, the migration from external task sensors to event-driven Assets, and a YAML-based DAG factory the team built to onboard new publishers without rewriting Python.

Key Takeaways
00:00 Introduction.
00:45 What Skimlinks does and how it operates as an affiliate marketing network aggregator for publishers.
02:12 Julian's team and the data platform they own: a reporting portal that serves ~2,000 weekly publisher users.
03:07 The stack: real-time ingestion into BigQuery, Airflow as the orchestrator, raw / silver / gold layers, and Apache Druid as the serving database for sub-second BI queries.
04:50 Reusing the same data marts for ~100 internal customers across marketing, finance, operations, and account management.
06:25 Airflow as the single orchestrator: BigQuery operators for SQL business logic, plus raw file exports for the largest publishers.
08:08 Moving from external task sensors to datasets (now Assets) and what the migration actually solved.
09:18 Why sensor polling created scheduler load and worker overload, and how event-driven Assets fixed both.
10:15 The lineage view in the Airflow UI that came as a bonus after the Assets migration.
10:49 The vision for multi-tenant Airflow inside Skimlinks: replacing cron, Rundeck, and team-local Airflow instances with a shared platform.
14:31 Building a custom DAG factory with YAML configuration for onboarding new publishers.
17:33 Breaking a single Python class into single-responsibility components for the DataPipe project.
19:07 Adding a Pydantic layer so misconfigured YAML fails at DAG parse time instead of run time.
20:31 Using AI assistance to guide refactoring decisions and generate tests across the new class structure.
22:34 What Julian wants from Airflow next: asset watchers paired with data contracts.

Resources Mentioned
Skimlinks - skimlinks.com
Apache Airflow - airflow.apache.org
Astronomer - astronomer.io
Google BigQuery - cloud.google.com/bigquery
Apache Druid - druid.apache.org
Pydantic - docs.pydantic.dev
Looker - cloud.google.com/looker

Thanks for listening to "The Data Flowcast: Mastering Apache Airflow® for Data Engineering and AI." If you enjoyed this episode, please leave a 5-star review to help get the word out about the show. And be sure to subscribe so you never miss any of the insightful conversations.

#AI #Automation #Airflow