Staging tables in DB vs storage area - sql

Question

Typically on an on-premise SQL server ETL workflow via SSIS, we load data from anywhere into staging tables and then apply validation and transformations to load/merge them into downstream data warehouse tables.
My question is if we should do something similar on Azure, where we have set of staging tables and downstream tables in azure SQL database or use azure storage area as staging and move data from there into final downstream tables via ADF.
As wild is it may seem, we also have a proposal to have separate staging database and downstream database, between which we move using ADF.

Answer 1

There are different models for doing data movement pipelines and no single one is perfect. I'll make a few comments on the common patterns I see in case that will help you make decisions on your application.
For many data warehouses where you are trying to stage in data and create dimensions, there is often a process where you load the raw source data into some other database/tables as raw data and then process it into the format you want to insert into your fact and dimension tables. That process is complicated by the fact that you may have data arrive late or data that is corrected on a later day, so often these systems are designed using partitioned tables on the target fact tables to allow re-processing of a partition worth of data (e.g. a day) without having to reprocess the whole fact table. Furthermore, the transformation process on that staging table may be intensive if the data itself is coming in a form far away from how you want to represent it in your DW. Often in on-premises systems, these are handled in a separate database (potentially on the same SQL Server) to isolate it from the production system. Furthermore, it is sometimes the case that these staging tables are re-creatable from original source data (CSV files or similar), so it is not the store of record for that source material. This allows you to consider using simple recovery mode on that database (which reduces the Log IO requirements and recovery time compared to full recovery). While not every DW uses full recovery mode for the processed DW data (some do dual load to a second machine instead since the pipeline is there), the ability to use full recovery plus physical log replication (AlwaysOn Availability Groups) in SQL Server gives you the flexibility to create a disaster recovery copy of the database in a different region of the world. (You can also do query read scale-out on that server if you would like). There are variations on this basic model, but a lot of on-premises systems have something like this.
When you look at SQL Azure, there are some similarities and some differences that matter when considering how to set up an equivalent model:
You have full recovery on all user databases (but tempdb is in simple recovery). You also have quorum-commit of your changes to N replicas (like in Availability Groups) when using v-core or premium dbs which matters a fair amount because you often have a more generic network topology in public cloud systems vs. a custom system you build yourself. In other words, log commit times may be slower than your current system. For batch systems it does not necessarily matter too much, but you need to be careful to use large enough batch sizes so that you are not waiting on the network all the time in your application. Given that your staging table may also be a SQL Azure database, you need to be aware that it also has quorum commit so you may want to consider which data is going to stay around day-over-day (stays in SQL Azure DB) vs. which can go into tempdb for lower latencies and be re-created if lost.
There is no intra-db resource governance model today in SQL Azure (other than elastic pools which is partial and is targeting a different use case than DW). So, having a separate staging database is a good idea since it isolates your production workload from the processing in the staging database. You avoid noisy neighbor issues with your primary production workload being impacted by the processing of the day's data you want to load.
When you provision machines for on-premises DW, you often buy a sufficiently large storage array/SAN that you can host your workload and potentially many others (consolidation scenarios). The premium/v-core DBs in SQL Azure are set up with local SSDs (with Hyperscale being the new addition where it gives you some cross-machine scale-out model that is a bit like a SAN in some regards). So, you would want to think through the IOPS required for your production system and your staging/loading process. You have the ability to choose to scale up/down each of these to better manage your workload and costs (unlike a CAPEX purchase of a large storage array which is made up front and then you tune workloads to fit into it).
Finally, there is also a SQL DW offering that works a bit differently than SQL Azure - it is optimized for larger DW workloads and has scale-out compute with the ability to scale that up/down as well. Depending on your workload needs, you may want to consider that as your eventual DW target if that is a better fit.
To get to your original question - can you run a data load pipeline on SQL Azure? Yes you can. There are a few caveats compared to your existing experiences on-premises, but it will work. To be fair, there are also people who just load from CSV files or similar directly without using a staging table. Often they don't do as many transformations, so YMMV based on your needs.
Hope that helps.