I'm trying to point two different copy activity pipelines into single output dataset. All pipelines and dataset have frequency/availability set to Day. I've tried configuring pipeline1 as "style": "StartOfInterval" and pipeline2 as "style": "EndOfInterval". But with that setup I'm getting error on publish:
The Activity schedule does not match the schedule of the output
Dataset. Activity: 'MyCopyActivity'. Dataset:
'MyDataset'.","code":"ActivityDataSetSchedulerMismatch"
As a workaround I could create two different datasets, and point them to the same resource.
Is it possible to achieve this with single output dataset?
If the reason is to merge multiple inputs into one output.
You could instead have a single copy activity pipeline have two separate inputs.
The data set inputs could have different availability windows and then the copy activity could combine them into one output dataset.
Both pipeline and output datasets availability/scheduling properties should be same in all the cases.
In you case, you have different "style" for Pipelines, but you are referring single output dataset which has only one style(default is Endofinterval).
For One pipeline it will match, but for other pipeline it will throw error.
To overcome this, you have to create two output datasets with same linked service. Don't forget to match the "style" of OutputDatasets with corresponding pipelines
No it is not possible to use a single output dataset in two copy activities. You need to create two datasets and point them to the same resource.
Related
In Azure Data Factory I have a pipeline, created from the built-in copy data task, that copies data from 12 entities (campaign, lead, contact etc.) from Dynamics CRM (using a linked service) and outputs the contents as parquet files in account storage. This is run every day, into a folder structure based on the date. The output structure in the container looks something like this:
Raw/CRM/2022/05/28/campaign.parquet
Raw/CRM/2022/05/28/lead.parquet
Raw/CRM/2022/05/29/campaign.parquet
Raw/CRM/2022/05/29/lead.parquet
That's just an example, but there is a folder structure for every year/month/day that the pipeline runs, and a parquet file for each of the 12 entities I'm retrieving.
This involved creating a pipeline, dataset for the source and dataset for the target. I modified the pipeline to add the pipeline's run date/time as a column in the parquet files, called RowStartDate (which I'll need in the next stage of processing)
My next step is to process the data into a staging area, which I'd like to output to a different folder in my container. My plan was to create 12 scripts (one for campaigns, one for leads, one for contact etc.) that essentially does the following:
accesses all of the correct files, using a wildcard path along the lines of: Raw/CRM/ * / * / * /campaign.parquet
selects the columns that I need
Rename column headings
in some cases, just take the most recent data (using the RowStartDate)
in some cases, create a slowly changing dimension, ensuring every row has a RowEndDate
I made some progress figuring out how to do this in SQL, by running a query using OPENROWSET with wildcards in the path as per above - but I don't think I can use my SQL script in ADF to move/process the data into a separate folder in my container.
My question is, how can I do this (preferably in ADF pipelines):
for each of my 12 entities, access each occurrence in the container with some sort of Raw/CRM///*/campaign.parquet statement
Process it as per the logic I've described above - a script of some sort
Output the contents back to a different folder in my container (each script would produce 1 output)
I've tried:
Using Azure Data Factory, but when I tell it which dataset to use, I point it to the dataset I created in my original pipeline - but this dataset has all 12 entities in the dataset and the data flow activity produces the error: "No value provided for Parameter 'cw_fileName" - but I don't see any place when configuring the data flow to specify a parameter (its not under source settings, source options, projection, optimize or inspect)
using Azure Data Factory, tried to add a script - but in trying to connect to my SQL script in Synapse - I don't know my Service Principal Key for the synapse workspace
using a notebook Databricks, I tried to mount my container but got an error along the lines that "adding secret to Databricks scope doesn't work in Standard Tier" so couldn't proceed
using Synapse, but as expected, it wants things in SQL whereas I'm trying to keep things in a container for now.
Could anybody point me in the right direction. What's the best approach that I should take? And if its one that I've described above, how do I go about getting past the issue I've described?
Pass the data flow dataset parameter values from the pipeline data flow activity settings.
I have and ADF pipeline which reads data from an on-prem source and copies it to a dataset in azure.
I want to perform some datachecks:
If the data contains the features I need
If there is null in some features
If the feature is all nulls
It should fail if the conditions above dnt meet
Is there a way to do this in data factory without using a batch service and just activities in data factory or maybe a dataflow.
Many approaches to this you could do a traditional batch process running function/code in a process. You could weave together ADF activities into multiple steps combination of 'Lookup Activity' possibly followed by a 'Validation Activity' and 'Delete Activity' with your criteria and rules defined.
Azure Data Factory 'Data Flows' - https://learn.microsoft.com/en-us/azure/data-factory/concepts-data-flow-overview - Allows you map out data transformation as data moves through the pipeline in a codeless fashion.
A pattern with ADF Data Flows is 'Wrangling Data Flows' to work with data and prepare it for consumption. Ref Article - https://learn.microsoft.com/en-us/azure/data-factory/wrangling-overview
The Copy activity in Azure Data Factory (ADF) or Synapse Pipelines provides some basic validation checks called 'data consistency'. This can do things like: fail the activity if the number of rows read from the source is different from the number of rows in the sink, or identify the number of incompatible rows which were not copied depending on the type of copy you are doing.
This is probably not quite at the level you want so you could look at writing something custom, eg using the Stored Proc activity, or looking at Mapping Data Flows and its Assert task which could do something like this. There's a useful video in the link which shows the feature.
I tried using Assert activities but for the scope of my work this wasn't enough!
Therefore, I ended up using python code for data checks.
However, assert activity servers better if your datacheck criteria is not hard as mine.
You can try to create data flows and apply conditional split activity. This will help you to achieve your scenario.
There is no such coding for this. This is diagrammatically you can do this in ADF or Azure Synapse Data Flow.
Find my attached data flow diagram that checks a few conditions like when the year is less than the specified year or if data in a column is null, date format, etc.
I wanted to create a data curation framework using Data Flow that uses generic data flow pipelines.
I have multiple data feeds (raw tables) to validate (between 10-100) and write to sink as curated tables:
For each raw data feed, need to validate the expected schema (based on a parameterized file name)
For each raw data feed, need to provide the Data Flow Script with validation logic (some columns should not be null, some columns should have specifici data types and value ranges, etc.)
Using Python SDK, create Data Factory and mapping data flows pipelines using the Data Flow Script prepared with the parameters provided (for schema validation)
Trigger the python code that creates the pipelines for each feed, does validation, write the issues into Log Analytics workspace and tear off the resources at specific schedules.
Has anyone done something like this? What is the best approach for the above please?
My overall goal is to reduce the time to validate/curate the data feeds, thus I wanted to prepare the validation logic quickly for each feed and create python classes or Powershell scripts scheduled to run them on generic data pipelines at specific times of the day.
many thanks
CK
To validate the schema, you can have a reference dataset which will be having the same schema (first row) as of your main dataset. Then you need to use “Get Metadata” activity for each dataset and get the structure of each dataset. Your Get Metadata activity will look like this:
You can then use “If Condition” activity to matches the structure of both datasets using equal Logical Function. Your equal expression will look something like this:
If both datasets’ structure matches, your next required activity(like copy the dataset to another container) will be performed.
Your complete pipeline will look like this:
The script which you want to run on your inserted dataset could be performed using “Custom” activity. You again need to create the linked service and it’s corresponding dataset for your script which you will run to validate the raw data. Please refer: https://learn.microsoft.com/en-us/azure/batch/tutorial-run-python-batch-azure-data-factory
To schedule the pipeline as per your specific pipeline will be take care by Triggers in Azure Data Factory. A schedule trigger will take care of your requirement of auto trigger your pipeline at any specific time.
I want to take data from 2 databases and copy(coalesce) it into 1 using Data factory.
The issue is: It seems that multiple inputs is not allowed for copy activities.
So i resorted to having 2 different datasets which are exact copies but with a different name... and then putting 2 different activities into the 1 pipeline which use their specific output dataset.
It just seems odd and wrong to do it this way.
Can i have some help.
This is what my diagram currently looks like:
Is there no way of just copying data from 2 seperate databases (which have the same structure but different data) to the 1 database?
The short answer is yes. But you need to work within the constraints of how ADF handles this.
A couple of things to help...
You'll always need at least 2 activities to do this when using the copy type activity. Microsoft of course charges per activity execution in ADF, so they aren't going to allow you to take shortcuts having many inputs and output per single copy activity (single charge).
The approach you show above is ok and to pass the ADF validation as you've found you simply need to have the output datasets created separately and called different things. Even if they still refer to the same underlying target table etc. This is really only a problem for the copy activity. What you could do is land the data firstly into separate staging tables in the Azure target database just for the copy (1:1). Then have a third downstream activity that executes a stored procedure that does the union of tables. In this case you could have 2 inputs to 1 output in the activity if you want to have that level of control in ADF.
Like this:
Final point, if you don't want the activities to execute in parallel you could chain the datasets to enforce a fake dependency or add a simple 'delay' clause to one of the copy operations. A delay on an activity would be simpler than provisioning a time slice offset.
Hope this helps
I would like to check if I'm missing any important points here.
My pipeline is only for Featurization. I understand that once a pipeline that includes an Estimator is fitted; then saving the pipeline will persist the params the Estimator has learned. So loading a saved pipeline in this case means not having to re-train the Estimator; which is a huge point.
However; for the case of a pipeline which only consists of a number of Transform stages; would I always get the same result on feature extraction from a input dataset using either of the below two approaches?
1)
Creating a pipeline with a certain set of stages; and configuration per stage.
Saving and reloading the pipeline.
Transforming an input dataset
versus
2)
Each time just instantiating a new pipeline (of course with the exact same set of stages; and configuration per stage)
Transforming the input dataset
So; alternative phrasing would be; as long as the exact set of stages; and configuration per stage is known; a Featurization pipeline can be efficiently (because there is no 'training an estimator' phase) recreated without using save or load?
Thanks,
Brent