SQL Saturday Atlanta 2024

Hey there happy coders! Last weekend I had the pleasure of speaking at the SQL Saturday Atlanta event in Georgia! It was an awesome time of seeing data friends and getting to make some new friends. If you live near a SQL Saturday event and are looking for a great way to learn new things, I can’t recommend SQL Saturday’s enough. They are free and an excellent way to meet people who can help you face challenges in a new way. Below are my notes from various sessions attended as well as the materials from my own session. Enjoy!

Link to the event – https://sqlsaturday.com/2024-04-20-sqlsaturday1072/#schedule

My session – Real-Time Analytics in Fabric

Thank you so much to everyone who came out to see my first ever session on Real-Time Analytics in Fabric! We had a couple of glitches in the Logic App creation, but had an excellent time troubleshooting together as a group. Please check out my GitHub for the slide deck as well as all the code used in the demonstration. Happy coding!

Link to my GitHub – https://github.com/Anytsirk12/DataOnWheels/tree/main/Real-Time%20Analytics%20in%20Fabric

Introduction to SQL Server Essential Concepts by Bradley Ball

Goal is to understand the core things about databases to enable deeper understanding.

ACID = atomicity, consistency, isolation, and durability.
Atomicity = either all of it commits or none of it commits. Consistency = my data must be in a consistent state before a transaction completes. Isolation = transaction must operate independently from other transactions. Durability = complex logging that is the transaction log. The log has all the commit history and ensures accurate data.

Transaction isolation levels – serializable, read committed (SQL server default), read uncommitted, repeatable read, snapshot isolation. You can set some of these at the db level. Serializable = blocks anything trying to get data on the same page, you can set this at the transaction level. Read committed = I can’t read data if a transaction is currently occurring. Read uncommitted = a dirty read, grabs data that isn’t committed. Repeatable read = nobody uses this lol. It’s a shared lock that holds for a longer period of time than the typical micro second. Shared lock means everyone can read the data. Snapshot isolation = Oracle and postgres use this. Every thing is an append and update. If I have 4 transactions, 1 update and 3 read, usually update would block reads, but this would redirect readers to a copy of the data in the TempDB at the point in time they requested to read it (aka before the update commits).

DMV = dynamic management view. Bradley used one that allows you to see active sessions in your database. We can see that a read query is blocked by an uncommitted transaction. We can see the wait_type = LCK_M_S and the blocking_session_id which is our uncommitted transaction. To get around this, he can run the read script and get a dirty read by setting the isolation level to read uncommitted. To unblock the original request, he can use ROLLBACK TRANSACTION to allow it to unlock that page of data.

How does SQL Server work on the inside? We have a relational engine and a storage engine. User interacts with a SNI which translates the request to the relational engine. User > SNI > Relational Engine [command parser > optimizer (if not in planned cache otherwise goes straight to storage engine) > query executer] > Storage Engine [access methods (knows where all data is) > buffer manager (checks the data cache but if not found then goes to the disk and pulls that into the buffer pool data cache). This gets extremely complicated for other processes like in-memory OLTP. The SQL OS is what orchestrates all these items.

SQL OS – pre-emptive scheduling (operating system) & cooperative pre-emptive scheduling (accumulates wait stats to identify why something is running slower).

Locks, latches, waits. Locks are like a stop light (row, page, and table escalation). If you lock a row, it will lock a page. Latches are who watches the locks/watchmen. It’s a lock for locks. Waits are cooperative scheduling. If a query takes too long, it will give up it’s place in line willingly. That creates a signal wait which signals there’s too much lined up.

SQL data hierarchy. Records are a row. Records are on a data page (8 k). Extents are 8 pages (64 k). It’s faster to read extents than pages. Allocation bit maps are 1s and 0s that signify data on a data page that enables even faster data reads – allows governing on 400 GB of data on 1 8KB page. IAM chains and allocation units allows quick navigation of pages. Every table is divided into in row data, row overflow data (larger than 8064 k), and lob data (large object like VARCHAR max and images).

Allocation units are made of 3 types:
1. IN_ROW_DATA (also known as HoBTs or Heap or B-Trees)
2. LOB_DATA (also known as LOBs or large object data)
3. ROW_OVERFLOW_DATA (also known as SLOBs, small large object data)

Heaps vs Tables. Oracle stores data as a heap which is super fast to insert. In SQL, these have bad performance due to clustered indexes and inserting new data. This is very situational. A table is either heap or clustered index, cannot be both. But heaps can have non-clustered indexes.

B-Tree allows you to get to the record with less reads by following a logic tree (think h is before j so we don’t need to read records after j). Heaps create a 100% table scan without a clustered index. Adding the clustered index dropped that significantly to only 1 read instead of the 8000 reads.

Recovery models – full, bulk logged, simple (on-prem). In the cloud everything is full by default. Full means everything is backed up. Bulk means you can’t recover the data but you can rerun the input process. Simple means you can get a snapshot but you can’t do any point in time restore. This will largely be determined by any SLAs you have.

Transaction log. This will constantly be overwritten. Your log should be at least 2.5 as large as your largest cluster. DBCC SQLPERF(logspace) will get you all the space available for logs in the various dbs. Selecting from the log is always not recommended since it creates a lock and logs are always running, so don’t do this in prod lol. Rebuilding indexes will grow your transaction log massively. To free up space in the transaction log, you have to a backup log operation which is why those are super important.

Fun tip, when creating a table you can put DEFAULT ‘some value’ at the end of a column name to provide it a default value if one is not provided. Pretty cool.

You can use file group or piecemeal restores to restore hot data much faster then go back and restore older, cold data afterward. To restore, you must have zero locks on the db. While restoring, the database is not online. Note, if you do a file group restore, you cannot query data that is in a unrestored file group so queries like SELECT * will not work.

Tales from the field has a ton of YouTube videos on these subjects as well.

Lessons Learned using Fabric Data Factory dataflow by Belinda Allen

What are dataflows? Dataflows are a low-code interface tool for ingesting data from hundreds of data sources, transforming your data using 300+ data transformations. The goal is to allow for more people to manipulate and use data within your organization. At the heart, it’s Power Query.

Why start with dataflows as a citizen developer? It’s power query and you know that. It’s low-code data transformation. Excellent for migrating Power BI reports to Fabric.

Lots of great discussion about when it makes sense to use a dataflow gen2.

You can copy and paste power query from Power BI by going into the advanced editor OR you can hold shift and select all the queries you want then ctrl c then go to power query for a dataflow gen2 in the online service and hit ctrl v and it will populate with all your tables! Pretty neat. You can also make your relationships within the online portal.

DBA Back to Basics: Getting Started with Performance Tuning by John Sterrett

For the code visit: https://johnsterrett.com/blog-series/sql-server-performance-root-cause-analysis/

Goal of today’s session – arm anyone who is new to performance tuning with processes and sills to solve common problems.

Basic query runtime. SQL has something called wait stats that tells you what caused the query to be slow. When you run a massive query, it will go into a suspended state which will require reading from disc instead of from memory cache (T1). After that, you’re in a runable state (T2). Finally, you get to run it (T3).

Basic bottlenecks = memory, disk, CPU, network, locking blocking & deadlocks. Adding memory is typically the fastest way to improve performance.

Identify performance problems happening right now:

EXEC sp_whoisactive. This is an open source script that gives you insight into who’s running what right now. You can get this from https://whoisactive.com. The cool thing about this is there are more ways to run it than just EXEC sp_whoisactive. Identify what’s consuming the most CPU from the column. There’s also some parameters you can use like @sort_order. EXEC sp_whoIsActive @sort_order = ‘[CPU] DESC’, @get_task_info = 2. The task info parameter will give more information in a wait_info column. The best command is exec sp_whoIsActive @help = 1. This provides ALL the documentation on what it does. Adam (the creator) also has a 30 day blog series on everything it can do for you! One option to make things run faster is to kill the process causing the issue lol.

How to handle blocking. You can do explicit transactions with BEGIN TRANSACTION which will lock the table. At the end, you need to either COMMIT or ROLLBACK or else that lock holds. SQL uses pessimistic locking as default so it won’t let you read data that’s locked – it will simply wait and spin until that lock is removed. You can use exec sp_whoisactive @get_plans = 1 to get the execution plan. Be careful, the wait_info can be deceptive since the thing that takes the most time may not be the problem. It may be blocked by something else, check the blocking_session_id to ve sure. Also check the status and open_tran_count to see if something is sleeping and not committed. Keep in mind that the sql_text will only show you the last thing that ran in that session. SO if you run a select in the same session (query window) as the original update script, it won’t be blocked and can run and THAT query will show up in the who is active which can be super confusing. To resolve this issue, you can use ROLLBACK in that session to drop that UPDATE statement.

To find blocking queries use EXEC sp_whoIsActive @find_block_leaders = 1, @sort_order = ‘[blocked_session_count] DESC’.

Identifying top offenders over the long term:

There’s a feature in SQL 2016 forward called Query Store which persists data for you even after you restart data. It’s essentially a black box for SQL. Query Store is on by default in SQL 2022 and online servers. It’s available for express edition as well. Be sure to triple check this is on, because if you migrated servers it will keep the original settings from the old server. If you right click on the DB, you can navigate to query store and turn it on via Operation Mode (Requested) to Read write. Out of the box is pretty good, but you can adjust how often it refreshes and for how much history. To see if it’s enabled, you should see Query Store as a folder under the db in SSMS.

Under query store, you can select Top Resource Consuming queries. There’s lots of configuration options including time interval and what metric. SQL Server 2017 and newer have a Query Wait Statistics report as well to see what was causing pain. It’ll show you what queries were running in the blocking session. You won’t get who ran the query from query store, but you can write sp_whoisactive to a table that automatically loops (very cool). This will have overhead on top of your db, so be mindful of that.

Intro to execution plans:

Keep in mind, SQL’s goal is to get you a “good enough” plan, not necessarily the best plan. Follow the thick lines. That’s where things are happening. Cost will tell you the percentage of the total time taken.

Key lookups. It’s a fancy way to say you have an index, so we can skip the table and go straight to the data you have indexed. BUT if there’s a nest loop, then there’s an additional columns in the select statement so it’s doing that key lookup for every value. More indexes can make your select statements worse if it’s using the wrong index that isn’t best for your query.

Index tuning process.
1. Identify tables in query
2. Identify columns being selected
3. Identify filters (JOIN and WHERE)
4. Find total rows for each table in the query
5. Find selectivity (rows with filter/table rows)
6. Enable statistics io, time, and the actual execution plan
7. Run the query and document your findings
8. Review existed indexes for filters and columns selected
9. Add index for lowest selectivity adding the selected columns as included columns
10. Run the query again and document findings
11. Compare findings with baseline (step 7)
12. Repeat last 5 steps as needed

To see existing indexes, you can run sp_help ‘tableName’. In the example, there’s an index key on OnlineSalesKey but that field is not used in our filter context (joins and where statements) in the query. Order of fields in indexes do matter because it looks in that order.

Brent Ozar made a SP you can use called sp_blitzIndex that will give you a ton of info on an index for a table including stats, usage, and compression. It also includes Create TSQL and Drop TSQL for that index to alter the table.

To turn on stats, use SET STATISTICS IO, TIME ON at the beginning of the query. Be sure to also include the actual execution plan (estimated doesn’t always match what actually happened). Now we can benchmark. Use SET STATISTICS IO OFF and SET STATISTICS TIME OFF. Create an non clustered index with our filter context columns.

Derby City Data Days

It was awesome to see the Kentucky data community come out for the first Derby City Data Days in Louisville, KY! Bringing together communities from Ohio, Tennessee, and Kentucky, the Derby City Data Days event was an excellent follow-up to Data Tune in March and deepened relationships and networks made at the Nashville event. In case you missed it, below are my notes from the sessions I attended as well as the resources for my session. Be sure to check out these speakers if you see them speaking at a conference near you!

Schedule

Building Self-Service Data Models in Power BI by John Ecken

What and why: we need to get out of the way of business insights. If you try to build one size fits all, it fits none. Make sure you keep your data models simple and streamlined.

Security is paramount to a self-service data model. RLS is a great option so folks only see their own data. You can provide access to the underlying data model for read and build which enables them to create their own reports off the data they have access to. If you give your user contributor access, then RLS will go away for that user. Keep in mind, business users need pro license OR need to be in a premium workspace.

One really great option is for people to use the analyze in Excel option to interact with the most popular BI tool – Excel. This allows them to build pivot tables that can refresh whenever needed. You can also directly connect to Power BI datasets from their organization! You can set up the display field option as well to get information about the record you connect to. Pretty slick! With this, security still applies from RLS which is awesome.

Data modeling basics – clean up your model by hiding or removing unnecessary columns (ie sorting columns). Relationships matter. Configure your data types intentionally. Appropriate naming is vital to business user success. Keep in mind where to do your transformations – SQL vs DAX (think Roche’s Maxum). Be sure to default your aggregations logically as well (year shouldn’t be summed up).

Power BI Measures – creations, quick create, measure context, time-based functions. Whenever possible, make explicit measures (using DAX) and hide the column that it was created off of so people utilize the measure you intended. Make sure you add descriptions, synonyms (for Copilot and QA), featured tables, and folders of measures. The functionality of featured tables makes it wise to use folders of measures within your fact tables.

John likes to use LOOKUP to pull dims back into the fact table so he ends up with as few tables as possible. There are drawbacks to this such as slower performance and model bloat, but the goal is for end users who don’t have data modeling experience or understanding. Not sure I agree with this method since it’s not scalable at all and destroys the purpose of a data model. Make sure you hide columns you don’t want end users to interact with.

To turn on feature table, go to the model view then go to Properties pane and toggle that is featured table button. It will require a description, the label that will populate, and the key column (cannot be hidden) that the user will put in excel as a reference for the business user to call records off of.

The PIVOT() TSQL Operations by Jeff Foushee

GitHub: https://github.com/jbfoushee/MyPresentations/tree/main/TSQL_Pivot_Operators

Be sure to look at his GitHub for the awesome source code!

Come to Lousiville on May 9th to see a presentation on JSON and TSQL.

The goal of this is to avoid FULL OUTER JOINs. This is extremely unscalable since maintenance would be terrible. We will avoid this by using pivot. Pivot means less rows, more columns. PIVOT promotes data to column headers.

You get to decide how the tuple that’s created on the pivot is aggregated (count, min, max, sum, avg, etc.). Exactly one aggregate can be applied, one column can be aggregates, and one columns values can be promoted into the column header.

PIVOT ( SUM(Col1) FOR [ID] IN ([ID_value_1], [ID_value_2], etc.)
SUM = the aggregate, ID = the column that will become more columns, the IN values = the column values from ID that will be promoted into the column header.

Time for a 3 column pivot. For this, we are doing a two column pivot and ignoring one of the fields. You can even pivot on computed fields but make sure you include the values in that inclusion clause. Be careful about adding unnecessary data.

How do you manage the VTCs (the column values that end up as column headers)? Option 1 – don’t. Option 2 – explicitly request the ones of interest and provision for future extras. Option 3 – use dynamic SQL! You can use cursor, XML, etc. Check out his ppt deck from github for code samples!

An n-column PIVOT works by essentially creating a 2-column pivot (at the end of the day it’s only two columns that ever get pivoted) and knowing which you want split into new columns.

Ugly side of PIVOT = lookups. The more fields you need to add in from additional tables, the worse performance will be. Your best option there would be to do a group by, the pivot. Another limitation is you can’t use a function in your pivot aggregation (SUM() vs SUM() *10). Get your raw data clean then pivot.

Time for UNPIVOT!

Unpivot = convert horizontal data to vertical. Less columns, more rows. Unpivot demotes column headers back into data.

Be very very careful with your data type your about to create. Remember lowest common denominator, all the values must be able to fit in one common datatype without overflow, truncation, or collation.

UNPIVOT( newColPropertyValue FOR newColPropertyName IN ([originalCol1], [originalCol2],etc.)

You need to make sure all your original columns have the same datatype. NULLs get automatically dropped out. If they are needed, you can convert them using an ISNULL function to a string value or int value depending on your need.

There’s also an option for XML-based unpivot.

Cross Apply = acquires a subset of data for each row found by the outer query. Behaves like an inner join, if no subset is found then the outer row disappears from the result set. Outer Apply is similar but it’s more like a left join. Cross Apply does keep your NULL values. You can also use a STRING_SPLIT with cross apply.

Multi-Unpivot normalizes hard-core denormalized data. You just add more UNPIVOT lines after your initial FROM statement! Make sure you have a WHERE statement to drop any records that don’t align at a column level. Something like WHERE LEFT(element1,8) = LEFT(element2, 8).

My Session – Time for Power BI To Git CI/CD

Thanks to everyone that attended my session! Had some great questions and conversations! Here’s the link to my github with the slide deck from today: https://github.com/Anytsirk12/DataOnWheels/tree/main/Power%20BI%20CICD

Medallion Architecture for Fabric by Steve Hughes

This session was awesome! We were able to watch a series of Fabric 5 minute videos and had an amazing discussion between them about options for building out a Fabric infrastructure using medallion techniques. Check out Steve’s YouTube channel for his Fabric 5 playlist and to learn more about his experience working with ALS.

SQLBits 2024 Recap

Had an absolutely amazing time at SQLBits this year! It was lovely to see all my data friends again and had the opportunity to introduce my husband to everyone as well! In case you missed a session, or are curious about what I learn at these conferences, below are my notes from the various sessions I attended.

Here is a link to my GitHub that contains the slides and code from my presentation on creating a M-agical Date Table. Thank you to everyone who attended! https://github.com/Anytsirk12/DataOnWheels/tree/main/SQLBits%202024

What’s new in Power Query in Power BI, Fabric and Excel? by Chris Webb

Power query online is coming to PBI desktop and will include diagram & schema views, the query plan, and step folding indicators! Currently in private preview, coming to public preview later this year.

Power query in Excel. Nested data types now work (very cool, check this out if you don’t know what data types are in Excel). Python in excel can reference power query queries now which means we don’t have to load data into Excel! That will allow analysis of even more data because you can transform it upstream of the worksheet itself. Very cool.

Example: import seaborn as sns

mysales = xl(“Sales”)

sns.barplot(mysales, x=”Product”, y = “sales”, ci=None)

Limitations – can’t use data in the current worksheet.

Power query is in Excel online! Later this year, you will be able to connect to data sources that require authentication (like SQL Server!).

Power query can be exported as a template! This will create a .pqt (power query template). You can’t import a template file in Excel, but you can in Dataflows Gen2 in Fabric.

Fabric Dataflow Gen2 – data is not stored in the dataflow itself. We now have Power Query copilot! It’s in preview, but it’s coming. You will need a F64 to use it.

Paginated reports are getting power query!! It will be released to public preview very soon. All data sources will be available that are in power bi power query. This will allow snowflake and big query connections annnddd excel! You can even pass parameters to the power query get data by creating parameters within the power query editor. Then you go to the paginated report parameters and create the parameter there. Finally go to dataset properties and add a parameter statement that sets those two equal to each other.

Rapid PBI Dev with ChatGPT, Tips & Tricks by Pedro Reis

Scenario 1: DAX with SVG & CSS.
SVG – Problem is to create a SVG that can have conditional formatting that doesn’t fill the entire cell. ChatGPT does best with some working code, so give it a sample of the SVG you want and as ChatGPT to make it dynamic. Demo uses GPT4 so we can include images if needed. He uses Kerry’s repo to get the code sample that we put into GPT (KerryKolosko.com/portfolio/progress-callout/).
CSS – uses HTML Content (lite) visual (it’s a certified visual). Prompt asks gpt to update the measure to increase and decrease the size of the visual every 2 seconds. Pretty awesome looking. He’s also asking it to move certain directions for certain soccer players (Ronaldo right and Neymar zigzag).

Scenario 2: Data Modeling and analysis. He uses Vertabelo to visualize a datamodel and it creates a SQL script to generate the model. He passes the screenshot of the datamodel schema to ChatGPT and it was able to identify that a PK was missing and a data type was incorrect (int for company name instead of varchar). Super neat. The image does need to be a good resolution. He also put data from excel into GPT to have it identify issues like blanks and data quality issues (mismatched data types in the same column). It could grab invalid entries and some outliers but it’s not very reliable now.

Scenario 3: Report Design Assessment. He grabbed four screenshots from a Power BI challenge and asked GPT to judge the submissions with a 1-10 rating based on the screenshots and create the criterion to evaluate the reports. Very neat! Then he asked it to suggest improvements. It wasn’t very specific at first so he asked for specfic items like alternatives to donut charts and other issues that lowered the score.

Scenario 4: troubleshooting. He asks it to act as a power bi expert. It was able to take questions he found on the forum, but sometimes it hallucinate some answers and it’s not trained on visual calculations yet.

Scenario 5: Knowledge improvement. He asked GPT to create questions to test his knowledge of Power BI models. It even confirmed what the correct answers. It even created a word doc for a shareable document that can be used to test others in the future.

Streamlining Power BI with Powershell by Sander Stad and Linda Torrang

Data-Masterminds/PSPowerBITools is an open source code that is in the powershell gallery as well as within GitHub. Github link: https://github.com/Data-Masterminds/PSPowerBITools.

There’s a module called MIcrosoftPowerBIMgmt that is used with a service account to wrap around the rest api and allow easier calls of those APIs through powershell. When running PowerShell, to see the things in a grid use the code ” | Out-GridView” at the end of your command and it creates a very nice view of the data. All demos are available at github. The first demo is to grab all the orphaned workspaces then assign an admin. Their module includes more information like the number of users, reports, datasets, etc. They also have a much easier module that will grab all the activity events from a given number of days (it will handle the loop for you).

Another great use case will be moving workspaces from a P SKU capacity to a F SKU capacity automatically. The admin management module doesn’t work for this, but the PSPowerBITools module can! To see help – use “get-help” in front of the command to get details and examples. For more details, add “-Detailed” at the end to see that. Adding “-ShowWindow” at the end of a script will pop open a separate window to see the response (note- doesn’t work with -Detailed but does contain the details). To export use this script: $workspaces | Export-Csv -NoVolbber -NoTypeInformation -Path C:\Temp\export_workspaces.csv.

The goal is to use the module for more ad hoc admin tasks like remove unused reports, personal workspaces, data sources, and licenses.

New to Powershell like me? Check out this article on how to automate Powershell scripts to run on your laptop: https://www.technewstoday.com/how-to-automate-powershell-scripts/.

Using ALLEXCEPT or ALL-VALUES by Marco Russo

ALL: removes all the filters from expanded table/columns specified.
REMOVEFILTERS: like ALL (it’s the same function). We use this for code readability.
ALLEXCEPT: removes filters from expanded table in the first argument, keeping the filter in the following table/columns arguments.

The following produce the same result:
ALLEXCEPT(customer, customer[state], customer[country])
REMOVEFILTERS(customer[customerkey], customer[name], customer[city])
The nice thing about ALLEXCEPT is that it’s future proof against additional fields being added to the table.

But is it the correct tool?

Demo – goal is to get % of continent sales by getting country or state or city sales / total continent sales

To get that total continent sales, we need sales amount in a different filter context. The simplest way to do this is using ALLEXCEPT since we always only want the continent field from the customer table.
CALCULATE([Sales Amount], ALLEXCEPT( customer, customer[continent]))

This works if we include fields from the same table – customer. If we pull in fields from another table, it won’t work as we want because it will be filtered by a different table’s filter context. To get rid of that, you can add REMOVEFILTERS(‘table’) for each OR we can change the ALLEXPECT to use the fact table.
CALCULATE([Sales Amount], ALLEXCEPT( Sales, customer[continent]))

One problem is if a user drops the continent from the visual, the filter context no longer includes continent so the ALLEXCEPT will no longer work properly as it will function as an ALL.

To fix this, use remove filters and values. Using values adds in the continent value in the current filter context, then removes the filter context on customer. Lastly, it applies the value back into the filter context so we end up with what we want which is just continent is filtered. This works quickly because we know each country will only have one continent.

CALCULATE( [sales amount], REMOVEFILTERS( Customer, VALUES (customer[continent]))

Performance difference? Yes – ALLEXCEPT will be faster, but the REMOVEFILTERS and VALUES should not be too much slower.

Analytics at the speed of direct lake by Phillip Seamark and Patrick LeBlanc

Phil did a 100 minute version of this session on Wednesday.

What is direct lake? We will evaluate over three categories – time to import data, model size, and query speed. Direct lake mode takes care of the bad things from both import and direct query modes. Direct lake only works with one data source. Data is paged into the semantic model on demand triggered by query. Tables can have resident and non-resident columns. Column data can get evicted for many reasons. Direct Lake fallback to SQL server for suitable sub-queries. “Framing” for data consistency in Power BI reports. Direct lake model starts life with no data in memory. Data is only pulled in as triggered by the DAX query. Data stays in that model once pulled for hours. We only grab the columns we need which saves a ton of memory.

Limitations – no calculated columns or tables. No composite models. Calc groups & field parameters are allowed. Can’t be used with views, can only be used with security defined in the semantic model (you can have RLS on the semantic model itself). Not all data types are supported. If any of these are true, Power BI will fall back to Direct Query mode.

SKU requirements: only PBI premium P and F skus. No PPU nor pro license.

Why parquet? column-oriented format is optimised for data storage and retrieval. Efficient data compression and encoding for data in bulk. Parquet is also available in languages including rust, java, c++, python, etc. Is lingua franca for data storage format. Enhanced with v-order inside of Fabric which gives you extra compression. Analysis services can read and even prefers that form of compression so we don’t need to burn compute on the compression. V-Order is still within open source standards so this still doesn’t lock you into Fabric only. Microsoft has been working on V-Order since 2009 as part of Analysis services. Row group within parquet files corresponds directly to the partitions/segments in a semantic model. Power BI only keeps one dictionary, but parquet has a different dictionary id for each row group/partition. When calling direct query, low cardinality is key because most of the compute will go to the dictionary values that need to be called across the various partitions. This is why it’s VITAL to only keep data you need for example, drop or separate times from date times and limit decimals to 2 places whenever possible.

You can only use a warehouse OR a data lake. So if you need tables from both, make a shortcut from the data lake to the warehouse and just use the warehouse. You can only create a direct lake model in the online service model viewer or in Tabular editor. You cannot create it in PBI desktop today.

Microsoft data demo. One file was 880 GB in CSV > 268 GB in Parquet > 84 GB with Parquet with V-ORDER.

Sempty is a python library that allows you to interact with semantic models in python.

In the demo, Phil is using sempty and semantic-link libraries. Lots of advantages to using the online service for building the semantic model – no loading of data onto our machines which makes authoring much faster. After building this, you can run a DMV in DAX studio. Next, create a measure/visual that has a simple max of a column. That will create a query plane to reach out to parquet and get the entire column of data needed for that action then filter as needed. After that, the column will now be in memory. Unlike direct query, direct lake will load data into model and keep it for a bit until eviction. The direct lake will also evict pages of columns as needed when running other fields. You can run DAX queries against the model using python notebook and you can see via DMV what happened.

Framing. Framing is a point in time way of tracking what data can be queried by direct lake. Why is this important? data consistency for PBI reports, delta-lake data is transient for many reasons. ETL Process – ingest data to delta lake tables. What this does is creates one version of truth of the data at that point of time. You can do time traveling because of this, but keep in mind that for default models it frames pretty regularly. If you have a big important report to manage, you need to make sure to include reframing in the ETL process so it reloads the data properly. This can be triggered via notebook, service, api, etc. You can also do this in TMSL via SSMS which gives you more granular table control.

DAX to SQL Fallback. Each capacity has guard rails around it. if you have more than a certain number of rows in a table, then it will fall back to direct query automatically. Optimization of parquet files will be key to stop it from falling back.

Deep (Sky)diving into Data Factory in Microsoft Fabric by Jeroen Luitwieler

Data Pipeline Performance – How to copy scales.

Pipeline processing – less memory & less total duration. No need to load everything into memory then right. Read and Write are done in parallel.

Producer and consumer design. Data is partitioned for multiple concurrent connections (even for single large files). Full node utlization. Data can be partitioned differently between source and sink to avoid any starving/idle connections.

Partitions come from a variety of sources including physical partitions on the DB side, dynamic partitions from different queries, multiple files, and multiple parts from a single file.

Copy Parallelism concepts: ITO (Intelligent Throughput Optimization), Parallel copy, max concurrent connections. ITO – A measure that represents the power used which is a combo of CPU, memory, and network resource allocation used for a single copy activity. Parallel copy – max number of threads within the copy activity that read from source and write to sink in parallel. Max concurrent connections – the upper limit of concurrent connections established to the data store during the activity run (helps avoid throttling).

Metadata-driven copy activity. Build large-scale copy pipelines with metadata-driven approach in copy inside Fabric Data pipeline. You can’t parameterize pipelines yet, but it is coming. Use case is to copy data from multiple SQL tables and land as CSV files in Fabric. Control Table will have the name of the objects (database, schema, table name, destination file name, etc). Design has a lookup to get this control table and pass that through a foreach loop and iterate through these to copy the tables into Fabric.

Dataflows Gen2 – performance optimization and AI infused experiences. 4 performance principles – delegate to the most capable resource, sometimes you have to do the most expensive thing first, divide and conquer, be lazy do as little work as possible.

Query folding – known as query delegation, push down, remote/distributed query evaluation. Wherever possible, the script in Power Query Editor should be translated to a native query. The native query is then executed by the underlying data source.

Staging – load data into Fabric storage (staging lakehouse) as a first step. The staged data can be referenced by downstream queries that benefit from SQL compute over the staged data. The staging data is referenced using the SQL endpoint on the lakehouse. Staging tips – data sources without query folding like files are great candidates for staging. For faster authoring, have a different dataflow for staging and another for a transformations.

Partitioning – a way to break down a big query into smaller pieces and run in parallel. Fast copy does this in the background.

Lazy evaluation – power query only evaluates and executes the necessary steps to get the final output. It checks step dependencies then applies query optimization to make the query as efficient as possible.

AI infused experiences – column by example, table by example, fuzzy merge, data profiling, column pair suggestions. Table by example web is pretty awesome. You can extract any data from any HTML page and have it figure out how to turn that HTML into the table you want to see. Basically allows you to screenscrape as part of your ETL process and it will refresh if any changes are done on the website. Table by example for a text/csv is also amazing, it can try to figure out how to take a human readable file and split it into multiple columns. Game changers!

Website Analytics in my pocket using Microsoft Fabric by Catherine Wilhelmsen

Catherine built her own website and blog hosted on Cloudflare. Cloudflare has an API that pulls all the stats. To build this, she worked with Cloudflare GraphQL API. They have one single endpoint for all API calls and queries are passed using a JSON object.

Data is returned as a json object. She built the orchestration inside of Fabric Data Factory.

First portion is to set the date, she has this check if she wants to use today’s date or the custom date.

After that, she accesses the data using a copy data activity. Last step is to copy that data into a lakehouse. The file name is dynamic so she can capture the date that the data was pulled from.

The destination is a Lakehouse where she can build a report off of it. The best part of the copy data is the ability to map that nested JSON response into destination columns within the lakehouse. As a bonus for herself, she made a mobile layout so she can open it on her phone and keep track of it on the go.

Next steps for her project – see country statistics, rank content by popularity, compare statistics across time periods.

Calculation Groups & C# the perfect couple by Paulina Jedrzejewska

Love story between calculation groups and C#.

All slides are available on her github: https://github.com/pjedrzejewska

What makes it a perfect couple? We can add C# to our calc group dev to make it much more flexible.

Calc groups don’t allow you to create a table with customizable sorting of the measures and have measures not using the calc group within the same visual. For example, if I have previous year and current year in my calc group, it would display all of my current year measures together and all the prior year measures together, but I may want to see current year sales right next to prior year sales instead of having current year sales, revenue, cost, etc. then have the prior year much further away. To solve this, we will create separate measures for each time intelligence option. But why do that by hand when we can automate?

What can we automate? Creating & deleting measures, renaming objects, moving objects to folders, formatting objects.

Why automate? Works through a bunch of objects in seconds, reusable, homogeneous structure and naming, and no more tedious tasks for you!

For this demo, we will use Tabular Editor 2, PBI Desktop, Code Editor (visual studio code). She uses a .cs file to script out the code. All the columns in the fact table that will be turned into a measure have a prefix of “KPI”.

column.DaxObjectFullName will give you the full table, column reference which allows you to reference that dynamically within C# code.

Calculation groups are essentially a template for your measures. It’s a feature in data modeling that allows you to take a calculation item and use it across multiple measures. These are very useful for time intelligence, currency conversion, dynamic measure formatting, and dynamic aggregations.

When creating a calc group, always be sure to include one for the actual measure value by using SELECTEDMEASURE(). You can also refer to other calculation items within others as a filter.

Now that we have a calc group, the C# can loop through every base measure and through all the calculation group items and generate our new measures off of that. Remember all the code is in her github 🙂

Another great use case for C# is to add descriptions to all the measures.

This allows for self-documented code.

SQL Saturday ATL – BI Recap

First of all, wow! What an incredible event! Loved seeing everyone and meeting so many wonderful new people. Thank you to everyone who organized, spoke, and attended this event! Below are my notes from the sessions I attended as well as a link to my github with all the resources from my talk. Thank you again to all those who came to my session, you were all an amazing crowd and I really enjoyed learning with you all!

Link to the event and speakers: https://sqlsaturday.com/2024-02-10-sqlsaturday1071/#schedule

Introduction to Fabric Lakehouse by Shabnam Watson

For SQL Server users, a lakehouse will be used for landing zone and the data warehouse will be used for silver and gold layers. Currently, no demonstrated difference in price/performance between lakehouse and warehouse. Raw/Bronze = likely a lot of small files. Use with a lakehouse. Lots of ways to do (or not do) medallion architecture.

Cost is now capacity based. If you go over your capacity, there’s some smoothing but if you’re still over then they will throttle performance until smoothing catches up with your existing capacity. There is no per user license for Fabric items. F SKUs can be purchased on Azure portal, but P SKUs are through M365 licenses. F2 jobs will not be killed in the short term and the experience will match a F64. Allows you to play around for a cheaper price.

Parquet file formats are not deletable, so Delta tables are key. They build on top of parquet to soft delete (it toggles that parquet file as inactive). You can time travel because of that, but there is a cost to that since it requires a lot of storage. Fabric engines write Delta with V-order (Vertipaq). Data is automatically optimized for PBI VertiPaq engine. For example, Databricks writing to a datalake will run slower in PBI than a Fabric notebook writing to that datalake since it’ll be optimized.

Lakehouse works directly with files in the lake (workspace). Data ingestion options: Low Code/No Code (pipelines, dataflows gen2, shortcuts) and code first (notebooks, pyspark, sparkSQL). Data transformation uses the same tools. Data orchestration is through pipelines or notebooks.

Data load considerations
Use Case (Recommendation)
Small file upload from local machine (use local file upload)
Small data (dataflows)
Large data (Copy tool in pipelines)
Complex data transformations (Notebook code)

Use pipelines to orchestrate, not to transform. Use dataflows or notebooks to transform data. Notebooks can optimize your data files into larger files for faster runs.

Power BI import mode will still be fastest and direct query will still be more real time, but direct lake will allow for near real time and nearly as fast as import since there are no more scan operations. Limitations with direct lake – web modeling only (desktop is coming), it will fall back to direct query when capacity runs low which will slow your reports.

You can tie down security for the sql objects that are query-able via the SQL endpoint. You can query cross-workspace, so you can lock down the editing of data sources in a different workspace from where people consume it. Spark runs much faster in Fabric than in Synapse because the clusters are always up in Fabric so it doesn’t need to spin up a cluster to run your workload. Lakehouse land requires notebooks to interact with it. You can toggle to the SQL analytics endpoint within the service to build relationships, create measures, stored procedures, views, etc. Visual query allows you to drag and drop and right tables to generate sql behind the scenes and allows you to save that sql as a view. You can also manipulate the data using M queries. You can use CICD currently with semantic models, report, and notebooks.

How to Weave DataOps into Microsoft Fabric by John Kerski

Goal is to make production easier by combining principals from DevOps, Agile, and Lean manufacturing (every transformation steps needs to be tested for quality). If you check early, you’ll beat the customer. John has a video on this on YouTube as well.

GitHub: https://github.com/kerski
GitHub for this subject: https://github.com/kerski/fabric-dataops-patterns

Principals:

1. Make it reproducible using version control
2. Quality is paramount (TEST

Pattern 1 – DAX Query View Testing Pattern

Prerequisites: dax query view, PBIP format, Azure DevOps/Git integration (Visual Studio Code)

Use a naming convention for the tabs because we can get this out from the pbip file later. Recommended to use Measure.Tests and Column.Tests. You’ll also want to apply a schema like – TestName, ExpectedValue, ActualValue, Passed. John has a template for these tests, it’s essentially a union of rows. When you build the test cases, use the performance analyzer to grab the dax from a KPI card visual that has your testing values.

Good examples of this are making sure your date table has the correct number of days or your yearly sales are within a threshold. This only really works if you have expected values. This can be achieved by having a static dataset that eliminates variables of error to ensure your DAX is or is not the issue.

You can also test if your table still has unique values, or the % of null values in a table. DAX query view has a quick query feature to get the column features. AMAZING. You can now use the INFO.TABLES() DAX function that came out in December to grab the schema and see if it’s changing (which can break visuals). This is also vital if a data type changes and it breaks the relationship. You do need to hard code and define the schema of the table (easily create this using another script).

If you save this in the PBIP file, within the .dataset folder all your DAX queries are stored in a folder called DAX Queries! That means you can easily see if a test has changed during the PR and also easily add these tests to other models. May be wise to have sample set of tests with instructions for set up that you can easily copy/paste into other existing dataset folders.

You’ll need your workplace governance set up (dev, test, prod). Standarize your schema and naming conventions for the tests (name.environment.test). Build tests – test calcs, content, and schema.

Pattern 2 – Gen2, Notebook Automated Testing Pattern

Devops stores DAX queries > Dataflow Gen 2 can call those queries > store results in Lakehouse. Leverage notebook to orchestrate and output back into Lakehouse for PBI report to consume.

Push the pbip into DevOps git repo so DevOps has the DAX queries. You do need to have environment variables (workspace guid, branch) called out in a config file (John has instructions in his github). You can use Azure DevOps Services REST APIs to grab the items in a project/repo from a dataflow. Dataflow is called API – Azure DevOps DAX Queries – Gen 2. Lakehouse consumes the data from dataflow then we can have the notebook use Semantic Link to query a dataset using the DAX queries. You do need to have the public library for semantic link preloaded instead of using pip install. John’s sample contains easy variables to set for the branches.

Not using data factory in Fabric to call the APIs since there are a number of APIs missing for that product at the moment, but you can use a pipeline to run the dataflow and notebook in a scheduled fashion. You can work with DevOps pipeline to call this as part of a PR requirement.

Data activator could be used to trigger alerts. Be sure someone is actually monitoring the report tool since email is not super reliable.

Modernizing ETL: Leveraging Azure Databricks for Enhanced Data Integration by Joshua Higginbotham

GitHub: https://github.com/Jhiggin/Common_ETL_Patterns_Using_DataBricks/tree/main

Not everything here is transferable to Fabric, but a lot of it is.

What is Azure Databricks? Unified data analytics platform. Contains data engineering, analytics, ML, AI.

Common Architecture: ingest using ADF > store in ADLS > prep and train in Databricks > Azure Synapse and AAS to model and serve.
Microsoft recommends ADF does extraction (E) and databricks for transformation (T), but there are reasons to do more of the E within databricks.

Delta Lake Archicture: bronze/raw > silver/filtered > gold/business-level aggregates. Delta lake allows you to incrementally improve quality of data until it’s ready for consumption. Silver could contain a datalake, lots of ways to do this medallion process.

New to Databricks? Check out DataCamp and MS learning pathways. Josh can provide some resources for his favorite people to follow.

You can deploy databricks as a set environment which can make it easy to avoid egress charges or latency that would be caused by having the tenant in a region different from the data.

You have to set up access to the workspace and the items within it. Recommended using security groups for this. You can link this to source control (git hub and Azure DevOps) and do everything including PRs within the databricks UI.

Compute has lots of options. All-purpose can do everything, but may not be fastest nor cheapest option. Summary tells you how many workers, run time, memory, cores, and features that are enabled. Photon enabled is much faster. Tags are super useful for assigning accounting details so you can assign spend to teams. Don’t mess with your configuration unless you know what you’re doing. One useful config is azure data lake storage credential passthrough to pass creds through, but now it’s recommended to use unity catalog. SQL warehouses now have a serverless option that can allow you to have quick spin ups and auto shut downs when not in use to save money. Policies set up limits for scale or who can use what compute.

A mount point can tell databricks where the data is stored within ADLS. You can mount it once and can make your code much more readable and secure without the full ADLS path. dbutils.secrets.get allows you to grab information from keyvault without any access to the actual values.

Recommendation – explicitly set your schema, don’t allow databricks to guess.

Recommendation – don’t worry about partitioning delta lake until you reach 7 million records since the gains are minimal.

When you use delta tables, you can look at the delta log to see changes and time travel.
You can use %sql DESCRIBE DETAIL table to see meta data for a table.

You can use Partner Connet to open data in Power BI desktop. Keep in mind that if it’s too large it will connect with direct query which means your cluster will always be up and costing you a lot of money.

%sql DESCRIBE HISTORY will let you see what the data looked like as of a certain date. Very dependant on how much data you’re storing/pruning.

You can create CLONE s off of a delta table, so you can clone the table as of a certain time. This can also create snapshots.

Does Clone persist with vacuum? Shallow clone is a metadata copy, deep is default which also copies the full data. Unsure if either are impacted by vacuum, but the idea with clone is that it’s not persisted typically.

Configuration_Driven_Development_Demo is another notebook he walked us through (see github). He has a config file in yaml that can config for a certain client (file path and schema for the table and transformations). It does a lot of things including mapping SQL data types to spark data types. This pipeline is for more repeatable patterns for ETL.

Power BI and External Tools: This is the way! by Jason Romans

Goal is to make you a jedi of Power BI.

History of tools – back in the day you would start with a project in visual studio. Behind the scenes it was mostly editing the JSON file. Tabular Editor came out as a glorified json editor.

SSMS (SQL Server Management System): dependable and very useful for SQL. No DAX formatting, no column names displayed. It can process (full refresh, recalculate, etc) premium datasets and you can schedule that using sql jobs. Can’t connect to PBI desktop. Great for a dba.

Azure Data Studio (ADS): shiny but not the tool you’re looking for. Cannot connect to analysis services nor pbi desktop.

DAX studio: an analyst. Has intellisense, links to dax guide, dax formatter, performance tuning. You can see hidden date tables. Select “Server Timings” in DAX studio then run a query to get all the server timings to understand where it’s spending it’s time. You can also use DAX studio to compare two versions of a measure to see which runs faster.

Tabular Editor: the builder or tinkerer. There is a free and paid version. Create and edit measures & calculation groups. You can also do C# scripting and automation (my favorite!) and command line integration! Can help with source control. Tabular Editor github has tons of easy to use scripts you can use for common BI tasks (Tabular Editor GitHub and Useful Scripts Blog). The best practice analyzer can easily apply fixes and script out a fix.

PBI Explorer – navigator. Able to explore the changes. Pbip format is new and compares projects using vscode. Gitignore file is in pbip to tell git to ignore the data itself and only bring in metadata. PBI explorer can visualize the report differences. You don’t need the project file format to use PBI explorer. To run it, you have to run the executable – not located in the external tools tab in power bi desktop. It can actually visualize the changes. There’s additional functionality to even explorer and analyze the visuals.

Bravo – https://bravo.bi/. Jack of all trades. This one seems to fill in the gaps. It can help optimize data model by looking at unused columns, size, cardinality. Allows the adding of custom date tables. Exports data. First screen analyzes the model for you. Also allows you to format dax. You’ll get a warning for two things on date tables – 1 is if you still have auto date/time enabled, 2 is if you have a Date or Holiday table already in the measure. If you give the bravo date table a new name, it will resolve issue number two. The bravo date table is a DAX calculated table.

Semantic Link – A Magic Cauldron to Unlock Power BI with Python by Stephanie Bruno & Stacey Rudoy

Blog – https://data-witches.com/

Content at – bit.ly/ILDpres

Check out fabric.guru for the inspiration behind this session. Here’s the specific blog post relating to semantic link and a data catalog (https://fabric.guru/fabric-semantic-link-and-use-cases#heading-what-about-power-bi-developers).

Semantic link is a Fabric feature announced in Dec 2023. https://learn.microsoft.com/en-us/fabric/data-science/semantic-link-overview

Features – we can evaluate DAX queries, get metadata, and take actions like kicking off a refresh.

Two ways to install – pip install semantic-link or create an environment. When you create an environment, you can add libraries and configure Spark properties. Once created, you can set a notebook to use it.

Why use it? Document the tenant (semantic model catalog), save snapshot of semantic model, download usage metrics.

Data Quality – there’s a find dependency script that allows you to see what columns rely on others. You can also use the Great Expectations library.

For REST APIs, use the FabricRestClient to call the Fabric REST endpoints.

Usage metrics! With semantic link, you can loop through all the workspaces and query the model to dump the data directly into a Datalake! You cannot connect to the Admin monitoring workspace using semantic link (not supported currently). So excited to be able to loop through all the workspaces one has access to!

Before running a notebook in Fabric, you need to create an environment. Add a python library for semantic-library. Then save and publish and you can use this within the notebook. If you don’t use environment, you’ll have to pip install semantic link. There is a learn module about the semantic link library (sempy fabric read table). If you call datatsets() without anything in the parenthesis it will grab datasets within the workspaces that the notebook lives in now. You can also plot relationships.

You can also query the DMVs if needed.

Ooo there’s an option for read_table so you don’t need to evaluate a dax expression to return a Fabric dataframe. To send to lakehouse all you need to use is to_lakehouse_table function and it creates or appends data to a table in lakehouse. Keep in mind, you’ll want to remove spaces for lakehouse functionality. You can only use to_lakehouse_table if your dataframe is a Fabric dataframe. Otherwise, you’ll needto use write.format(“delta”).mode(“overwrite”).

Phil Seamark has a great blog about this as well: https://dax.tips/2023/12/05/visualize-power-bi-refresh-using-sempy/.

Power BI – Performing Data Quality Checks Using Python & SQL by me 🙂

Abstract:

Picture this, you have a report in Power BI that someone passes off to you for data quality checks. Simple enough until they clarify they need every measure checked against the source database.
There are a few ways to make sure the measures match what is in the source data system, but you need something easily repeatable and self documenting. In this presentation, I will walk through how to use python and excel to perform our data quality checks in one batch and document any differences. In order to do that, we are going to build a python script that can run Power BI REST APIs, connect to a SQL Server, and read/write to Excel.
By the end of this session, attendees will have a plug and play tool to check data from SQL against Power BI.

Link to GitHub with Power Point and final scripts: https://github.com/Anytsirk12/DataOnWheels/tree/main/SQL%20Saturday%20ATL%20-%20BI%202024

Power BI VertiPaq Engine: Optimizing Column Encoding

Power BI encoding is a powerful optimizing option that is often overlooked because it’s not visible in neither the Power BI Desktop tool nor in Power BI Service. Natively, the VertiPaq engine in Power BI investigates all columns in the data model and determines how it can store that data most efficiently. To achieve maximum compression, the VertiPaq engine starts by encoding each column which determines the method of compression applied to that column. There are a couple types of encoding – value and hash.

Value and hash encoding work differently and can drastically impact the size and performance of your data model. Koen Verbeeck defines these two types of encoding very well (blog):

Value encoding – Here the value of the cell itself is used. It’s possible a mathematical operation is included. For example, if a column contains the values 10,000 and 10,001, SSAS can subtract 10,000 from the values and keep the results 0 and 1, which can be stored using fewer bits. Whenever the data is needed, SSAS only needs to add 10,000 again to retrieve the original values. Typically columns used in aggregations benefit from value encoding.

Hash encoding – With this type of encoding, the values are transformed into meaningless integers. This transformation step is then kept in a dictionary, so SSAS can translate the integers back to the original value. Hash encoding is used for string columns and for every column where value encoding is less efficient (for example if the column is very sparsely distributed). Typically columns used as group-by columns or foreign keys benefit from hash encoding.

In general, your data model will consume less storage and populate visuals faster if you have more columns that are compressed using value encoding. Value encoding is only available for numerical columns, which is part of the reason star schema is so powerful in Power BI. If you have a standard star schema, the table with the greatest number of rows/records will be a fact table which will have entirely numerical column types. Ideally, all of these numerical columns will be value encoded by default, but occasionally the VertiPaq engine requires a little extra guidance. While going through the steps below, you can monitor your dataset size and column encoding using the VertiPaq Analyzer.

Unfortunately, there’s not currently a way to check or even change the encoding for a column within the Power BI desktop. Thankfully, the guru’s over at Tabular Editor have exposed this for report developers. If you’re unfamiliar with Tabular Editor, I highly recommend downloading the free version (Tabular Editor 2) and checking out the getting started information on their website. To enable changing the Encoding Hint Type, you will need to go to File > Preferences and ensure that “Allow unsupported Power BI features (experimental)” box is checked. Without this box checked, it will appear that your code is running but no changes will be made within the model.

Once inside Tabular Editor, navigate to the C# script window and input the following code to see the current encoding type. If the encoding type is default (like the example below), this means the engine will determine the encoding on the fly.

Templated code snippet - all caps values are variables
var table_TABLENAME = Model.Tables["TABLENAME"];
var column_nameid = table_TABLENAME.Columns["COLUMNNAME"];
Output(column_nameid.EncodingHint);

Example in screenshot
var table_allocatedcharges = Model.Tables["Product"];
var column_acudfid = table_allocatedcharges.Columns["ProductKey"];
Output(column_acudfid.EncodingHint);

Unfortunately, you cannot set the encoding type itself, but you can set the encoding hint to encourage that column to be encoded one way or another. To change this encoding hint type, use the code below within that same C# editor window and hit the green run arrow. One of the major benefits of using a tool like Tabular Editor is programmability. In the script below, we are using a foreach statement to loop through all of the Int64 columns within our data model and setting the encoding hint to Value.

// for all Int64 columns set EncodingHint
foreach(var column in Model.Tables.SelectMany(t => t.Columns))
{
if(column.DataType == DataType.Int64)
column.EncodingHint = EncodingHintType.Value;
}

Let’s run the first script again and see if that EncodingHintType changed. And boom! Now it says Value! Please note, if yours still says Default, navigate to File > Preferences and ensure that “Allow unsupported Power BI features (experimental)” box is checked. Unfortunately, changing the encoding hint type falls under the unsupported Power BI feature category.

If you are using the VertiPaq Analyzer, you may notice that not all Int64 columns may be encoded as Values. Unfortunately, there are limitations to which columns will be encoded even after running the script. Keep in mind, the encoding hint type is just a hint. This community post covers some of the exceptions that fellow developers have run into: VertiPaq Decimal Columns Encoding Type Help. Marco Russo, Microsoft MVP & co-author of SQLBI, summarizes this phenomenon by saying “The hint is not enforcement. If the column cannot be represented as a value encoded column, it’s compressed as a hash encoded column. For example, the value encoded is 32-bit, if you have values outside the 32-bit range, it will be always hash-encoded. Also keep in mind that encoding hints are applied to the following refresh.”

In the sample script, we only looped through Int64 whole number columns. You can adjust this script to also look for decimal type columns for further optimization. Don’t forget to use VertiPaq Analyzer and enjoy watching that dataset shrink!

Additional Resources

• https://community.powerbi.com/t5/Desktop/Vertipaq-Engine-VALUE-vs-HASH/m-p/690874
• https://docs.tabulareditor.com/te2/Advanced-Scripting.html
• https://www.mssqltips.com/sqlservertip/5150/improve-analysis-services-tabular-2017-processing-with-encoding-hints/
• https://learn.microsoft.com/en-us/power-bi/guidance/import-modeling-data-reduction#optimize-column-data-types
• https://blog.enterprisedna.co/vertipaq-power-bi-how-columns-are-encoded/