Dimensional Data Modeling Full Course [13 Hours] - Everything in one place

One hundred worked scenarios on the decisions that decide whether your numbers mean anything. A published text-to-SQL benchmark scores around 40% against a raw warehouse schema. Against the same data modelled properly, current models reach the mid-80s. Add a semantic layer and it reaches 98-100% on covered questions. Most of that came from modelling the data, not from the layer on top. This course is about the errors that return a number instead of an error message. Valid SQL, real tables, every key resolving, and a figure that is wrong by a factor nobody notices for a quarter. Those are not coding failures. They are modelling decisions that were never made, or were made by accident. Everything is worked against one running example: a multi-channel retailer with 340 stores, 85,000 products, 1.2 million loyalty members and four source systems that disagree about what a product is. The landscape was frozen before module one and never changed, so the arithmetic is consistent across thirteen hours and you can check it yourself. 13 hours 27 minutes. 100 worked scenarios. 27 modules. No SQL on screen. WHAT YOU WILL BE ABLE TO DO Declare the grain of a fact table, and detect the month it drifted Classify a measure before somebody sums something that does not sum Choose between eight kinds of slowly changing dimension, per attribute Name a fan trap from the multiplier alone, without reading the query Conform four systems that disagree about what a customer is Review a model you did not build, in a week, with a six-step checklist COURSE STRUCTURE Part 1 - Foundations (11%) Part 2 - Fact Tables (11%) Part 3 - Dimensions (24%) Part 4 - Difficult Shapes (13%) Part 5 - Enterprise Integration (7%) Part 6 - Loading and Operating (12%) Part 7 - Performance and Physical (11%) Part 8 - Modern Practice (11%) CHAPTERS 0:00 m00_start_here 10:38 m00b_the_kestrel_model 22:28 m01_why_the_model_still_decides_everything 47:44 m02_the_four-step_design_process 1:13:36 m03_grain_declaring_it,_holding_it,_detecting_drift 1:43:12 m04_transaction,_periodic_snapshot,_accumulating_snapshot 2:19:27 m05_measures_and_additivity 2:51:02 m06_factless_fact_tables_events_and_coverage 3:14:15 m07_dimension_anatomy_keys_and_members 3:49:50 m08_attribute_design 4:17:52 m09_slowly_changing_dimensions_types_0_to_3 4:50:43 m10_slowly_changing_dimensions_types_4_to_7 5:27:03 m11_role-playing,_junk,_outriggers_and_snowflaking 6:02:57 m12_bridge_tables_and_multi-valued_dimensions 6:37:09 m13_ragged_and_variable-depth_hierarchies 7:03:14 m14_fan_traps,_chasm_traps_and_double_counting 7:45:26 m15_conformed_dimensions_and_the_enterprise_bus_matrix 8:20:04 m16_drilling_across 8:43:54 m17_late-arriving_facts_and_dimensions 9:17:20 m18_cdc,_deletes_and_restatements 9:46:32 m19_testing_a_dimensional_model 10:18:02 m20_aggregates_and_summary_design 10:45:25 m21_physical_design 11:15:14 m22_star_schema_or_one_big_table 11:44:12 m23_semantic_layers_and_agents 12:14:06 m24_inmon,_data_vault_and_anchor 12:37:09 m25_confusion_pairs_and_the_design_review 13:06:54 m26_what_holds A FEW OF THE IDEAS The row-type census. A uniqueness test proves rows are distinct. It says nothing about whether they mean the same thing. A correction is not a change. Recording one as the other manufactures history that looks exactly like history, and nothing detects it. Impact against weighted. Revenue by allergen that sums to more than total revenue is not broken. For that question it is the only honest answer. The multiplier names the trap. Stable across groups is a fan trap. Varying by group is a chasm trap. Thirty seconds, and no code read. WHO THIS IS FOR Data engineers who build warehouses and were never taught to model one. Analytics engineers who inherited a model and want a way to tell whether it is any good. Analysts who keep finding two numbers that disagree, from queries that are both correct. Anyone about to point a language model at their warehouse. It assumes you know what a join is and what a foreign key is for. There is no SQL in thirteen hours, no labs, and no vendor certification. Nothing in it depends on which platform you use, and there is nothing to install. A NOTE ON THE EXAMPLE Kestrel Group is invented. The stores, products, source systems and business details are fictional. The failure modes are not. Every drift point, every mismatched key and every wrong number is a pattern from real warehouse work, reconstructed so it can be taught without exposing anyone's data. A NOTE ON CURRENCY Most of this does not date. Grain, additivity and conformance are the same problems they were thirty years ago. Module 23 covers semantic layers and agents, its figures are as at late 2026, and it says so on screen and tells you to check them. Module 26 names the three places I was more confident than the evidence supports. #DataEngineering #DataModelling #Kimball #DataWarehouse #Analytics #SQL #AI