Data Observability

Introducing the First Semi-Structured Data Observability

Introducing the First Semi-Structured Data Observability
July 22, 2022
Max Lukichev
Max Lukichev

Why Semi-Structured

As databases were evolving and migrating to the cloud, the data models were changing too. Most modern cloud data warehouses or data lakes now support semi-structured schemas. Today pretty much any big data platform supports working with NDJSON, i.e., New Line Delimited File, where each row is a proper JSON representing a record. And even native support of JSON schema type: []

This brought a great opportunity for data architects to design the most efficient data model for storage and querying. However, at the same time, it created a challenge for Data Observability. And the reason is multi-value fields, aka arrays. Such data requires special logic to properly calculate basic data quality KPIs like completeness, uniqueness etc. as there can be some records with no values vs multiple values per record.

Example of a single record of such data:

  "first_name": "Jon",
  "last_name": "Doe",
  "address": [{
    "street":"123 Main St",
    "city": "Neverland",
    "zip": "94404"

Normally the data model is designed to make the most common queries to be efficient. Typically for analytical/reporting purposes a flatten operation (sometimes called unnest or explode) is used and is sufficient because such queries only require a single or a couple attributes to be flattened at a time. Effectively, it means that if a record has an attribute with two values, it will get converted into two records with the corresponding value for the attribute.

In the example above that would be:

{"first_name" : "Jon", … , "phone" : "123-124-23-45"}
{"first_name" : "Jon", … , "phone" : "124-234-3467"}

However when monitoring data metrics and their drifts, such an approach becomes very inefficient, as it would require flattening and calculating aggregations one attribute at a time. It is not uncommon to see hundreds and even thousand attributes in real-life cases. The accumulated overhead on each query with such an approach quickly becomes unacceptable.


Telmai is designed from the ground up to support very large semi-structured schemas. It can run with a constant number of aggregation operations (or queries) not depending on number of attributes, avoiding the overheads, and can process virtually unlimited complexity of the schema. 

In the example above Telmai extracts following attributes and automatically predicts trends for dozens of metrics per each of them to detects unexpected drifts:

These metrics include:

  • Total record count
  • % of unique values
  • % of incomplete/null values
  • Distribution of categorical values and patterns
  • Mean value frequency, length, …
  • % of date, number, … 
  • And many more

Telmai can process semi-structured data in any form: JSON (ndjson), parquet, data warehouse tables with JSON files. For example, one of the customers expressed a need to monitor product JSON data stored in Snowflake. The challenge was that this data was of a very complex schema - a hierarchy of over 8000 attributes. Such schemas are not unique for enterprises and specifically for data describing products, as each type of a product can bring a set of attributes specific to its type. Yet, the Telmai engine was able to handle such data efficiently and give peace of mind about the quality of the data.

In summary, Telmai architecture allows the analysis of data of any volume, located anywhere (files, streams, SQL engines), of flat and semi-structured schemas of any complexity. On top of that, it can scale horizontally well, so aggressive sampling is not required either.

On this page

Start your data observibility today

Connect your data and start generating a baseline in less than 10 minutes. 

No sales call needed

Start your data observability today

Connect your data and start generating a baseline in less than 10 minutes. 

Telmai is a platform for the Data Teams to proactively detect and investigate anomalies in real-time.
© 2023 All right reserved.