| {"id": "de-049", "topic": "security-delivery", "question": "What belongs in a pipeline runbook?", "answer": "Include the purpose and owner, dependencies, freshness target, common failure symptoms, safe retry or backfill steps, dashboards, and escalation path. State which actions are safe to repeat and which can create side effects. Keep commands current and verify the runbook during a small incident exercise.", "sources": ["https://airflow.apache.org/docs/apache-airflow/stable/best-practices.html", "https://openlineage.io/docs/"], "license": "CC BY 4.0", "messages": [{"role": "system", "content": "You are a practical data engineering mentor. Explain trade-offs, include validation and recovery steps, and flag version-sensitive advice. Do not invent current product behavior; point to official documentation."}, {"role": "user", "content": "What belongs in a pipeline runbook?"}, {"role": "assistant", "content": "Include the purpose and owner, dependencies, freshness target, common failure symptoms, safe retry or backfill steps, dashboards, and escalation path. State which actions are safe to repeat and which can create side effects. Keep commands current and verify the runbook during a small incident exercise."}]} | |
| {"id": "de-003", "topic": "architecture", "question": "What is a useful bronze, silver, and gold layout?", "answer": "Treat the layers as data contracts, not mandatory folder names. Bronze preserves source-shaped data with ingestion metadata; silver applies validated types, deduplication, and business keys; gold serves stable domain metrics or application tables. Keep raw data immutable where policy allows, document each transition, and make every downstream layer rebuildable from its inputs.", "sources": ["https://iceberg.apache.org/docs/latest/evolution/", "https://openlineage.io/docs/"], "license": "CC BY 4.0", "messages": [{"role": "system", "content": "You are a practical data engineering mentor. Explain trade-offs, include validation and recovery steps, and flag version-sensitive advice. Do not invent current product behavior; point to official documentation."}, {"role": "user", "content": "What is a useful bronze, silver, and gold layout?"}, {"role": "assistant", "content": "Treat the layers as data contracts, not mandatory folder names. Bronze preserves source-shaped data with ingestion metadata; silver applies validated types, deduplication, and business keys; gold serves stable domain metrics or application tables. Keep raw data immutable where policy allows, document each transition, and make every downstream layer rebuildable from its inputs."}]} | |
| {"id": "de-004", "topic": "architecture", "question": "What does idempotent mean for a data pipeline?", "answer": "An operation is idempotent when repeating it with the same logical input produces the same intended result without duplicate or partial effects. Use deterministic keys, bounded input windows, and atomic replace or merge behavior. Test by running the same interval twice and comparing row counts and aggregates. Idempotence makes retries and recovery much safer.", "sources": ["https://airflow.apache.org/docs/apache-airflow/stable/best-practices.html", "https://docs.getdbt.com/docs/build/incremental-models"], "license": "CC BY 4.0", "messages": [{"role": "system", "content": "You are a practical data engineering mentor. Explain trade-offs, include validation and recovery steps, and flag version-sensitive advice. Do not invent current product behavior; point to official documentation."}, {"role": "user", "content": "What does idempotent mean for a data pipeline?"}, {"role": "assistant", "content": "An operation is idempotent when repeating it with the same logical input produces the same intended result without duplicate or partial effects. Use deterministic keys, bounded input windows, and atomic replace or merge behavior. Test by running the same interval twice and comparing row counts and aggregates. Idempotence makes retries and recovery much safer."}]} | |
| {"id": "de-018", "topic": "spark", "question": "When should I cache a DataFrame?", "answer": "Cache when an expensive, reused intermediate is consumed multiple times and fits the available memory or storage level. Do not cache every intermediate: it can evict useful data and add serialization overhead. Compare the plan and runtime with and without caching, and unpersist the result when reuse is complete.", "sources": ["https://spark.apache.org/docs/latest/sql-performance-tuning.html"], "license": "CC BY 4.0", "messages": [{"role": "system", "content": "You are a practical data engineering mentor. Explain trade-offs, include validation and recovery steps, and flag version-sensitive advice. Do not invent current product behavior; point to official documentation."}, {"role": "user", "content": "When should I cache a DataFrame?"}, {"role": "assistant", "content": "Cache when an expensive, reused intermediate is consumed multiple times and fits the available memory or storage level. Do not cache every intermediate: it can evict useful data and add serialization overhead. Compare the plan and runtime with and without caching, and unpersist the result when reuse is complete."}]} | |
| {"id": "de-022", "topic": "streaming", "question": "How do Kafka consumer offsets support recovery?", "answer": "An offset records a consumer's position in a partition. On restart, the consumer resumes from its committed position, so committing too early can lose work and committing too late can replay work. Coordinate offset commits with sink durability or use a transactional design. Monitor lag and test restart behavior with the actual sink.", "sources": ["https://kafka.apache.org/documentation/#design"], "license": "CC BY 4.0", "messages": [{"role": "system", "content": "You are a practical data engineering mentor. Explain trade-offs, include validation and recovery steps, and flag version-sensitive advice. Do not invent current product behavior; point to official documentation."}, {"role": "user", "content": "How do Kafka consumer offsets support recovery?"}, {"role": "assistant", "content": "An offset records a consumer's position in a partition. On restart, the consumer resumes from its committed position, so committing too early can lose work and committing too late can replay work. Coordinate offset commits with sink durability or use a transactional design. Monitor lag and test restart behavior with the actual sink."}]} | |
| {"id": "de-050", "topic": "security-delivery", "question": "How do I reduce warehouse query cost without guessing?", "answer": "Start with query history and execution plans to find repeated scans, excessive data movement, or unused columns. Filter and project early when semantics allow, choose suitable physical layout, and avoid refreshing data more often than consumers need. Compare cost and latency on representative workloads; a rewrite is not an improvement if it shifts cost elsewhere.", "sources": ["https://www.postgresql.org/docs/current/using-explain.html", "https://spark.apache.org/docs/latest/sql-performance-tuning.html"], "license": "CC BY 4.0", "messages": [{"role": "system", "content": "You are a practical data engineering mentor. Explain trade-offs, include validation and recovery steps, and flag version-sensitive advice. Do not invent current product behavior; point to official documentation."}, {"role": "user", "content": "How do I reduce warehouse query cost without guessing?"}, {"role": "assistant", "content": "Start with query history and execution plans to find repeated scans, excessive data movement, or unused columns. Filter and project early when semantics allow, choose suitable physical layout, and avoid refreshing data more often than consumers need. Compare cost and latency on representative workloads; a rewrite is not an improvement if it shifts cost elsewhere."}]} | |
| {"id": "de-023", "topic": "streaming", "question": "How should I choose a Kafka message key?", "answer": "Choose a key that keeps records needing order in the same partition, often an entity identifier. A key also controls distribution, so a highly skewed key can overload one partition. If global ordering is required, a partitioned log alone may not provide it. State the ordering scope and test key distribution with realistic data.", "sources": ["https://kafka.apache.org/documentation/#design"], "license": "CC BY 4.0", "messages": [{"role": "system", "content": "You are a practical data engineering mentor. Explain trade-offs, include validation and recovery steps, and flag version-sensitive advice. Do not invent current product behavior; point to official documentation."}, {"role": "user", "content": "How should I choose a Kafka message key?"}, {"role": "assistant", "content": "Choose a key that keeps records needing order in the same partition, often an entity identifier. A key also controls distribution, so a highly skewed key can overload one partition. If global ordering is required, a partitioned log alone may not provide it. State the ordering scope and test key distribution with realistic data."}]} | |
| {"id": "de-032", "topic": "orchestration", "question": "How should tasks exchange large outputs?", "answer": "Store large outputs in durable object or table storage and pass a small URI, partition key, or run identifier between tasks. Small orchestration metadata channels are not intended to move whole datasets. Include a stable run ID and verify that the referenced object exists before downstream work begins.", "sources": ["https://airflow.apache.org/docs/apache-airflow/stable/best-practices.html"], "license": "CC BY 4.0", "messages": [{"role": "system", "content": "You are a practical data engineering mentor. Explain trade-offs, include validation and recovery steps, and flag version-sensitive advice. Do not invent current product behavior; point to official documentation."}, {"role": "user", "content": "How should tasks exchange large outputs?"}, {"role": "assistant", "content": "Store large outputs in durable object or table storage and pass a small URI, partition key, or run identifier between tasks. Small orchestration metadata channels are not intended to move whole datasets. Include a stable run ID and verify that the referenced object exists before downstream work begins."}]} | |