making data warehouse easy conor cunningham – principal architect thomas kejser – principal pm
TRANSCRIPT
Introduction• We build and implement Data Warehouses
(and the engines that run them)• We also fix DWs that others build• This talk covers the key patterns we use• We will also show you how you can make your
life easier with Microsoft’s SQL technologies
What World do you Live in?
Hardware should be bought when I
know the details
I need to know my hardware CAPEX before I
decide to invest
I can’t wait for you to figure all that out
Do it, NOW!
Sketch a Rough Model1. Define Roughly on
Business Problem2. Decide on Dimensions– Dim columns can wait
3. Build Dimension/Fact Matrix
Fact/Dim Sales Inventory Purchases
Customer X
Product X X X
Time X X X
Date X X X
Store X X
Warehouse X X
Estimate Storage
|𝐹𝑎𝑐𝑡𝑛|=|𝐹𝑎𝑐𝑡𝑟𝑜𝑤𝑛|∗ [ 𝑹𝒐𝒘 /𝒅𝒂𝒚 𝒏 ]∗[𝒅𝒂𝒚𝒔 𝒔𝒕𝒐𝒓𝒆𝒅𝒏 ]
|𝐷𝑊|=[𝑪𝒐𝒎𝒑𝒓𝒆𝒔𝒔𝒊𝒐𝒏 ]∗1𝑇𝐵1012𝐵
∗ ∑𝑖∈𝐹𝑎𝑐𝑡𝑠
¿𝐹𝑎𝑐𝑡 𝑖∨¿¿
≈ 4B
≈ 1/3 or sp_estimate_compression
≈ 8B
Why Integer Keys are Cheaper
• Smaller row sizes• More rows/page
= more compression• Faster to join• Faster in column stores
Pick Standard HW Configuration
• Small (GB to low TB) : Business Decision Appliance
• Medium (up to 80TB): Fast Track• Large (100s of TB): PDW– Note: Elastic scale plus for lower sizes too!
• Careful with sizes, some are listed pre-compression
Why does Fast Track/PDW Work?
• Warehouses are I/O hungry– GB/sec– This is high
(in a SAN terms)
• We did the HW testing for you
• Guidance on data layout
Implement Prototype Model• Design schema • Analyse data quality with
DQS/Excel– Probably not what you
expected to find!• Start with small data
samples!
Prototyping Hints• Generate INTEGER keys out
of strings keys with hash• Focus on Type 1
Dimensions• PowerPivot/Excel to show
data fast• Drive conversation with
end users!
Key Name City1 Thomas London
Key Name City From To1 Thomas Malmo 2006 20112 Thomas London 2012 9999
Customer Type 1
Customer Type 2
Prototype: What users will teach you
• They will change/refocus their mind when they see the actual data
• You have probably forgotten some dimension data
• You may have misestimated data sizes
Schema Design Hints• Build Star Schema• Beginners may want to avoid snowflakes (most of our users just use star)• Implement a Date Table (use INT key in YYYYMMDD format)
– Fact.MyDate BETWEEN 20000000 AND 20009999– Fact.MyDate BETWEEN ‘2000-01-01’ and ‘2000-12-31’– YEAR(Fact.MyDate ) = 2000
• Identity, Sequences• Usually you can validate PK/FK Constraints during load and avoid them in
the model• Fact Table – fixed sized columns, declared NOT NULL (if possible)• For ColumnStore, data types need to be the basic ones…
Why Facts/Dimensions?• Optimizers have a tough job• Our QO generates star joins early in search• We look for the star join pattern to do this
– 1 big table, dimensions joined to it…• Following this pattern will help you
– Reduced compilation time– Better plan quality (average)
• You can look at the plans and see whether the optimizer got the “right” shape– Wrong Plan your query is non-standard OR perhaps QO messed up!
Partition/Index the Model• Partition fact by load window• Fact cluster/heap?– Cluster fact on seek key– Cluster fact on date column (if cardinality > partitions)– Leave as heap
• Column Store index on– All columns of fact– Columns of large dim
• Cluster the Dim on Key
If(followedpattern) {expect …}• Star Join Shape• << get plan .bmp>>• Properties:– Usually all Hash Joins– Parallelism– Bitmaps– Join dimensions together, then scan Fact– Indexes on filtered Dim columns helpful if they are
covering
The Approximate Plan
PartialAggregate
Fact CSI Scan
Dim Scan
Dim Seek
BatchBuild
BatchBuild
HashJoin
HashJoin
HashStreamAggregate
Column Store Plan Shapes• For ColumnStore, it’s the same shape • Minor differences
– Batch mode (Not Row Mode)– Parallelism works differently– Converts to row mode above the star join shape
• If you don’t get a batch mode plan, performance is likely to be much slower (usually this implies a schema design issue or a plan costing issue)
• Partitioning Sliding Window works well with ColumnStore (especially since the table must be is readonly)
Data Maintenance• Statistics
– Add manually on Correlated Columns
– Update fact statistics after ETL load– Leave Dim to auto update
• Rebuilding indexes?– Probably not needed– If needed, make part of ETL load
• Switch out old partitions and drop switch target– Automate this
Serve the Data• Self Service– Tabular / Dimensional Cubes– Excel / PowerPivot / PowerView
• Fixed Reports – Reporting Services– PowerView
• Don’t clean data in “serving engines”– Materialise post-cleaned data as
column in relational source