Teradata vs. Snowflake: Differences in Architecture

Snowflake is not a Shared Nothing architecture; the computing resources access shared data.

Therefore there is no concept like the Primary Index. All computing resources can access the data.

This is the advantage of cloud architectures such as Snowflake: It simply scales.

On a Teradata system, an upgrade always means restructuring. When AMPs come in, tables must be redistributed to ensure optimal parallelism.
A new hashmap is needed, which contains the additional AMPs. However, it must be said that since Teradata 16, the restructuring can be delayed by using several hashmaps.

A significant advantage over a full Teradata system is given if you want to save space, e.g., by MVC. The paradox: In this case, you first need a lot of space to save space. In Snowflake, the data is always compressed, and if no more space is available, new space can be added at short notice.

There is no such thing as a secondary index or join index in Snowflake.

Snowflake relies on the concept of a virtual warehouse that separates the workload. Snowflake says there is no need for workload management, but it makes sense to have both when you look at Teradata.
Teradata offers a genuinely sophisticated Workload Management (TASM) and the system’s ability to partition. A virtual partition can access the full resources (CPU) if they are not needed by another partition—a significant advantage.

Scaling is not as flexible as a cloud database in which compute nodes and data can scale independently.

In Teradata, computing power and data scale together. This ensures that IO (whatever the bottleneck is) is always available in sufficient quantities.

But the disadvantage is that more storage cannot merely be added without distributing the data. Modern database systems run as much as possible in the main memory. On the other hand, Teradata uses spool files in each step of the execution plan, which often has to be stored temporarily on the mass storage.

In contrast to Teradata, Snowflake collects the required statistics on its own, without the user having to do anything.

Tables are stored in so-called micro partitions (max. 500MB in size). Within each micro partition, the data columnar is stored (the individual column values are compressed).

Each micro partition has metadata. In the metadata, the offset of each column and the range of values are stored. Access is gained by looking up the metadata in which a micro partition is needed to query the tables and where the required columns’ data are located.

This is completely different from the hashing used in Teradata. But as you can see, Snowflake is a pure column store. Similar to the row-level partitioning used in Teradata is the sorting of data within columns in Snowflake.

The possibilities of tuning in Snowflake are limited in contrast to Teradata. However, the numerous options Teradata offers are unnecessary due to the architecture.

There is no workload management comparable to Teradata (TASM) in Snowflake.

Workload tuning is done by defining virtual warehouses. In principle, nothing more than independent clusters of computers accessing the same data.

By forming clusters, performance can be optimized at the highest level.

Snowflake uses intensive disk-level caching. The Disk Cache is a combination of SSD and memory. When creating virtual warehouses, it is advisable to create users that execute the same queries in the same virtual warehouse. Then the disk cache can be used optimally.

Snowflake also offers a cache of the result set, which is available to every user independent of the virtual warehouse.

This cache holds the result sets of all queries of the last 24 hours. Each time the cache is used, the counter starts from zero again. This is far more extensive than caching the query plan in Teradata.

Teradata offers a block-level cache using Teradata Virtual Memory. The cache is available to all data warehouse users but is purged frequently and is temperature-based. But it’s not a cache of the result set. It’s caching of data blocks only.

It may be tempting to shut down a Snowflake virtual warehouse as soon as possible to save costs, but the disk cache is also lost when the warehouse is closed. Here, costs must be weighed up against performance. This is not an issue at Teradata. The system is available 24/7.

Snowflake does not use indexes or partitioning that the user can determine. However, it is possible to define a cluster key to locate specific column values within one or a few micro partitions:

ALTER TABLE Customer CLUSTER BY (CustomerId);

As a counterpart to the parsing engine, which handles sessions and delivers result sets to the client, Snowflake has a service layer.

This manages sessions and everything that goes with them (queries, transactions, result sets, etc.) and maintains a cache of the results of executed queries.

Snowflake only offers what Teradata understands by Soft Referential Integrity. There is no forced referential integrity as in Teradata.

When we speak of skew in Teradata, we mean an uneven load on the AMP.
Skewing is also present in Snowflake at the node level. Snowflake uses a method to minimize skew automatically: file stealing.

When a worker process has finished its work (i.e., scanned its input files), it requests additional files from other worker tasks. The other worker tasks check how many input file sets they have left to process and pass them to the requesting worker task. This automatically balances the skew.

This kind of skew optimization is not possible on a shared-nothing system like Teradata.

No. Snowflake runs exclusively in the cloud. In principle, Snowflake has understood how to make the best use of the various cloud services (Amazon, Google, Azure).

Take AWS (Amazon Web Services) as an example:
Data Storage uses Amazon S3.
Computing power in the form of a virtual warehouse is provided via clusters of EC2 instances.

Disk cache is provided by SSD/Main Memory of the EC2 instances.
Meanwhile, there is also Teradata in the cloud. However, it should be remembered that the disadvantages of Shared Nothing Architecture exist here.

Like Teradata, Snowflake uses a cost-based optimizer.

However, all required statistics are automatically maintained while loading or updating the data. In Teradata, these must be defined at least once by the user (there is, however, the option of having Teradata analyze the query logs to suggest statistics).

We know IPE (Incremental Planning and Execution) in Teradata. Here, the execution plan is modular, and the final plan is determined during the query’s execution.

Snowflake also makes certain decisions only at query runtime. This includes, for example, the selection of the join method.
Both optimizers can, therefore, generate dynamic execution plans.

I think not. I hope not.

Teradata is simply a perfectly tuned system, and I don’t think it will quickly disappear from the market.

However, I do see Hadoop’s chances dwindling.

For example, using Hadoop to store JSON and load it after conversion into a database system is trendy. Snowflake can handle this task with flying colors—no need for Hadoop.

I must admit that I am impressed with Snowflake. I have always been skeptical of the latest hypes like Cloud, Data Lake, Hadoop, etc., to name a few. It’s difficult to separate the marketing scream from the actual values. But here, the developers of Snowflake seem to have made a big hit.

The architecture of Teradata limits this. Even if Teradata is offered in the cloud, this does not mean it is a full “cloud database”.

Here one must look very carefully at what the individual providers understand about the cloud. There is a lot of marketing involved.

It’s not necessarily a matter of Snowflake being “better” than Teradata, as both systems have their own advantages and disadvantages. Snowflake offers better elasticity and flexibility due to its cloud-based architecture, while Teradata is a well-tuned, traditional system with more robust workload management features. Choosing between the two depends on your organization’s specific needs and requirements.

Teradata and Snowflake have different architectures and designs. Teradata uses a shared-nothing architecture, while Snowflake employs a shared-data architecture. Snowflake is designed for the cloud and offers more flexibility in scaling, whereas Teradata scaling requires time-consuming restructuring. Snowflake collects statistics automatically, while Teradata requires manual definitions and tools to achieve the same result. Additionally, Teradata has a more advanced workload management system than Snowflake’s virtual warehouse concept.

Organizations might consider moving from Teradata to Snowflake for a number of reasons:

Elasticity and scalability:

Snowflake allows for more flexibility in scaling resources, as it’s designed for the cloud and can independently scale compute and storage.
Automatic statistics collection:

Snowflake automatically collects required statistics without any user intervention, simplifying database maintenance.
Lower maintenance:

Snowflake doesn’t require (or offer) secondary indexes or join indexes, reducing the overall maintenance effort.

Cost savings:

Organizations might sometimes experience cost savings by moving to a cloud-based system like Snowflake due to its pay-as-you-go model.
However, it’s essential to carefully evaluate your organization’s specific needs before deciding.

Yes, people still use Teradata. As mentioned in the article, Teradata is a well-tuned, traditional system that offers robust workload management features and has been widely adopted in various industries. While there are cloud-based alternatives like Snowflake that have gained popularity, Teradata remains a viable choice for many organizations, particularly those that require sophisticated workload management or have significant investments in Teradata infrastructure.