Glossary
Key terms and concepts
title: “Glossary” subtitle: “Key terms and concepts” —
A
- ADLS2 (Azure Data Lake Storage Gen2)
-
Azure’s scalable cloud storage service optimized for big data analytics. Supports hierarchical namespace and
abfss://protocol. Workshop source:mhpdeworkshopsa/nyc-taxi-data. - ADLS2 SAS token
-
Shared Access Signature for the workshop ADLS2 container (
nyc-taxi-data). Used for Snowflake external stage access in trial accounts. - ai_query() (Databricks)
-
Databricks SQL function for LLM inference over table data (Module 6). Complements notebook Assistant and Genie — different from Snowflake
AI_COMPLETE(). - Aiven Kafka
-
Managed Apache Kafka broker used in Module 8 (optional). Facilitator runs a shared free-tier cluster; attendees consume the
user-activitytopic (Avro + Karapace Schema Registry). Separate dataset from NYC Taxi batch bronze — story uses taxi GPS; lab uses user-activity for a live topic per attendee. - Asset Bundles (DABs)
-
Databricks’ infrastructure-as-code tool for deploying pipelines, jobs, and notebooks via CLI. Pairs with LSDP and Lakeflow Jobs in production patterns. See
databricks bundle deploy.
B
- Bronze Layer
- First layer of medallion architecture. Raw data ingested as-is from source, with metadata added. No business transformations. Metadata column names vary by ingest path — see Bronze schema.
C
- Catalog (Unity Catalog)
- Top-level namespace in Databricks: Catalog > Schema > Table. Provides centralized governance, auditing, and lineage.
- Checkpoint
- A file on cloud storage that records the exact offset a streaming query last processed. Enables exactly-once semantics and fault recovery in Structured Streaming.
- COPY INTO
- Snowflake SQL command for bulk loading data from external stages into tables. Efficient for Parquet, CSV, JSON.
- Cortex AI (LLM)
-
Snowflake LLM and assistant features for Module 6:
AI_COMPLETE()(replaces legacySNOWFLAKE.CORTEX.COMPLETE()),AI_EXTRACT(), Snowflake CoCo (formerly Cortex Code / Copilot in Workspaces), Cortex Analyst, Cortex Search. Not the same as Cortex ML (Module 9). - Snowflake CoCo
-
Snowflake-native AI coding agent (rebrand of Cortex Code, 2026). In Snowsight Workspaces, use
Cmd+I/Ctrl+Ifor NL → SQL/Python assistance with warehouse context and RBAC. Same workshop role as the old “Cortex Code in Snowsight” exercise — see Module 6 Exercise 4. CLI/Desktop are optional; not required for the lab. - Cortex Analyst
- Snowflake service that translates natural language questions into SQL against governed semantic models (YAML definitions of tables and metrics).
- Cortex ML
-
Snowflake managed ML SQL functions for Module 9:
ML.FORECAST,ML.ANOMALY_DETECTION, and related time-series / classification APIs. RequiresSNOWFLAKE.CORTEX_USERon your role. No custom model artifacts — contrast with Databricks sklearn + MLflow. - Cortex Search
- Snowflake hybrid search service (semantic + keyword) over unstructured and structured content. Enables RAG patterns within Snowflake.
- Cross-validation
- ML evaluation technique that splits data into k folds, training on k-1 and testing on 1, rotating through all folds. More robust than a single train/test split.
D
- Data Leakage
-
Including information in ML features that wouldn’t be available at prediction time or mathematically encodes the target. The #1 ML mistake (e.g.,
total_amountincludestip_amount). - dbt (data build tool)
-
Transformation framework that turns SQL SELECT statements into managed tables/views with testing, documentation, and lineage. Does not ingest — reads existing Bronze via
source(). - Delta Lake
- Open-source storage format providing ACID transactions, time travel, and schema evolution on top of Parquet files.
- dynamic_table (dbt)
-
dbt materialization type on the Snowflake adapter only (
materialized: dynamic_table). Module 8 stretch — declarative refresh similar to native Dynamic Tables; lands in{attendee}_DBT_STREAMING, isolated from Snowflake-native_STREAMINGobjects. - Dynamic Table (Snowflake)
-
Snowflake object defined by a SQL SELECT that refreshes automatically based on
TARGET_LAG. Declarative streaming — you write SQL, Snowflake manages incremental refresh. Can also be created via dbtdynamic_tablematerialization.
E
- ELT (Extract, Load, Transform)
- Pattern where raw data is loaded first, then transformed in-place within the data warehouse. Used by all three pipelines in this training.
- Exactly-once Semantics
- Guarantee that each event is processed exactly once, even during failures. Achieved in Structured Streaming via checkpoint files + Delta Lake transaction log.
- External Stage
- Snowflake object pointing to a cloud storage location (ADLS2, S3, GCS) for reading external files.
F
- Feature Engineering
-
Process of creating input variables (features) for ML models from raw data. In this workshop each platform has its own path: Databricks sklearn on
silver_nyc_taxi_enriched, Snowflake Cortex ML on Gold/time-series views, optional Snowpark ML, and optional dbtml_features_tip_predictionin{attendee}_DBT_GOLD— dbt is a contract/testing layer, not the single source all trainers consume.
G
- Genie Space
- Databricks AI/BI feature. Natural language interface over data tables — ask questions in plain English, get SQL + results (Module 6).
- Gold Layer
- Third layer of medallion architecture. Pre-aggregated business KPIs and metrics optimized for reporting and dashboards. This training materialises 12 KPI tables.
I
- Incremental Materialization
-
dbt strategy that appends or merges only new rows instead of rebuilding the entire table. Requires
is_incremental()logic and aunique_key. Advanced pattern — base workshop models usetable/viewmaterializations only.
J
- Jinja
-
Templating engine embedded in dbt SQL files. Enables conditional compilation (
target.type), reusable macros,ref()/source()dependency management, and configuration blocks.
K
- KPI (Key Performance Indicator)
- Pre-computed business metric. This training computes 12 KPIs from taxi trip data (trips by hour, top zones, revenue, etc.).
L
- Lakeflow Jobs
- Databricks job scheduler (formerly Workflows). Runs notebooks, LSDP pipelines, Python scripts, dbt tasks, and SQL as multi-task DAGs with dependencies, retries, and notifications.
- Lakeflow Spark Declarative Pipelines (LSDP)
-
Databricks declarative ETL framework for production (formerly Delta Live Tables / DLT, renamed 2025). Workshop code uses
from pyspark import pipelines as dpwith@dp.materialized_view,@dp.expect_or_drop, andspark.read.table(). Legacyimport dlt/@dlt.tablestill works on many workspaces.
M
- Macro (dbt)
-
Reusable SQL function defined in
.sqlfiles undermacros/. Similar to functions in Python — accepts parameters and returns SQL snippets. Example:{ time_of_day('pickup_hour') }. - Materialization
-
Strategy for how a dbt model becomes a database object. Core types:
view,table,incremental,ephemeral,materialized_view(Databricks adapter). Snowflake adapter addsdynamic_tablefor Module 8 streaming. - Medallion Architecture
- Multi-layer data design pattern: Bronze (raw) → Silver (cleaned) → Gold (aggregated). Also called “multi-hop architecture.”
- Metric Views (Unity Catalog)
-
Databricks UC object (public preview, 2025) for defining reusable business metrics — measure, dimensions, and filters — once in catalog metadata. Consumers query the same definition from dashboards, Genie, and alerts instead of re-implementing KPI SQL. Workshop uses physical Gold
kpi_*tables, not Metric Views; see Module 2 callout and Gold Layer. - MLflow
- Open-source platform for ML lifecycle management. Tracks experiments (hyperparameters, metrics, model artifacts). Databricks includes MLflow autolog for automatic experiment capture.
- Model (dbt)
-
A SQL SELECT statement that dbt materializes as a table or view. Each
.sqlfile inmodels/is one model.
P
- Pipeline expectations (
@dp.expect_or_drop) -
LSDP decorators that declare data quality rules on pipeline tables. Violations can be dropped or flagged with metrics in the pipeline event log — declarative alternative to manual
WHEREfilters (Module 5). - Power BI
- Microsoft BI tool for the five-page KPI dashboard (Module 7 lab). Connects in Import mode to Gold tables from any completed batch path (Databricks, Snowflake, or dbt).
- PySpark
- Python API for Apache Spark. Used in Databricks notebooks for distributed data processing.
R
- ref() (dbt)
-
dbt function that creates a dependency between models.
{ ref('my_model') }references another model and builds the DAG. - Relay consumer
-
Facilitator-managed Python script (
streaming/snowflake/00_relay_consumer.py) that reads Aiven Kafka and lands NDJSON files on ADLS2 for Snowflake Snowpipe. Used because Aiven free tier has no Kafka Connect (Module 8).
S
- Schema tracks (workshop)
-
Per-attendee isolated schemas so pipelines do not overwrite each other: Databricks
{attendee_id}_bronze|silver|gold; Snowflake_SQL_*,_SP_*,_DBT,_DBT_GOLD,_STREAMING,_DBT_STREAMING. See recommended workshop tracks. - Silver Layer
-
Second layer of medallion architecture. Cleaned, validated, enriched data with derived metrics. Trustworthy data for analysis. Outputs:
silver_nyc_taxi_cleaned+silver_nyc_taxi_enriched. - Slim CI
-
dbt testing pattern that only tests changed models + their downstream dependencies instead of the entire project. Uses
--select state:modified+. - Snowpark
- Snowflake’s Python DataFrame API. Runs Python code natively on Snowflake compute. API is similar to PySpark. Used for batch ingest/transform and optional Snowpark ML (Module 9).
- Snowflake Intelligence
- Umbrella brand (announced 2025) bundling Cortex LLM, Analyst, Search, and ML functions into a unified AI layer.
- Snowpipe
-
Snowflake continuous loader that ingests new files from a stage into a table. Module 8 uses it on the relay-consumer NDJSON path under
streaming/user-activity/. - source() (dbt)
-
dbt function referencing a raw table not managed by dbt. Example:
{ source('bronze', 'nyc_taxi_trips') }. dbt does not read ADLS2 directly — your notebook or Snowpark job must create Bronze first. - Stream (Snowflake)
- Snowflake object that tracks changes (inserts, updates, deletes) on a table. Used for change data capture (CDC).
- Structured Streaming
-
Databricks / Spark API for processing unbounded data in micro-batches. Module 8 reads Aiven Kafka with
readStreamand writes append-only Delta bronze tables.
T
- Task (Snowflake)
- Snowflake object that runs SQL on a schedule or when triggered by a condition. Used with Streams for event-driven pipelines.
- Three Constraints
- YellowLine NYC decision framework (Module 7): Cost, Performance, and Compliance — used to recommend a platform mix for Marcus, not a single “winner.” See Module 7 — Comparison & Wrap-up.
U
- Unity Catalog
- Databricks’ centralized governance layer. Manages catalogs, schemas, tables, permissions, lineage, and data sharing.
V
- Virtual Warehouse (Snowflake)
- Snowflake compute resource. Auto-suspends when idle, auto-resumes on query. Sizes from X-Small to 6X-Large.
W
- Watermark
- In streaming, a threshold that tells the engine to ignore events older than X minutes. Enables window finalization and state memory cleanup. Tradeoff: shorter = lower latency but more dropped late events.
Y
- YellowLine NYC
- Fictional NYC yellow-taxi operator used as the training client story (inspired by TLC public trip data, not a real company). Marcus Chen (Operations Manager) works for YellowLine NYC; MHP is the consulting vendor engaged to build analytics. On-screen dispatch UI and Marcus’s badge use YellowLine branding — not MHP.