Show HN: DataProfiler – What's in your data? Extract schema, stats and entities | Hacker News Reader

Show HN: DataProfiler – What's in your data? Extract schema, stats and entities

github.com

86 pointslettergram11 comments

Deequ is the big data / Spark alternative for similar functionality in case anyone is interested: https://github.com/awslabs/deequ

Looks like a great project and wanted to highlight why projects like these are much more useful for file formats that don't have metadata (CSV) vs file formats with metadata (Parquet).

Parquet file footers contain metadata that provides column-level metadata, file-level metadata, and schema information.

Cluster computing technologies can use Parquet metadata to skip entire files (Parquet predicate pushdown filtering). If there is an age column in the data, then the min/max values will be stored in the Parquet footer. If you run a Spark query for df = spark.read.parquet(/some_folder).where(age > 90), it'll only read in the files that have a max_age greater than 90. Data skipping is one of the best performance benefits.

I wrote a blog post on analyzing Parquet file metadata with PyArrow if you're interested in learning more: https://mungingdata.com/pyarrow/parquet-metadata-min-max-sta...

File formats that make you infer the schema (CSV) are on the way out. Enjoy Parquet and the other benefits it provides, like column pruning!

> File formats that make you infer the schema (CSV) are on the way out.

Lol, good luck. People will continue using delimited text files for as long as they have eyeballs.

Yea, haha, that was worded too strongly. Should have said "more folks are using Parquet for data workflows and adoption seems to be increasing". CSV is better when you need something mutable / human readable, so they'll always be around.

Main issue with Parquet is actually the metadata. CSV is nice because I can send it to a friend and they can instantly understand it; you can print it and understand it.

In contrast, Parquet is good for systems and I definitely recommend it in programming, but sharing.. not so much.

> it'll only read in the files that have a max_age greater than 90. Data skipping is one of the best performance benefits.

Actually Parquet files have a chunked structure, with maybe 200 row groups in a large Parquet file. The metadata is stored per chunk in the file footer.

So if that suggests you only need a few chunks with a file, your data reads can be orders of magnitude faster.

The data is stored column wise too. So if you only need a few of the columns, you can get another 10x increase in performance.

It's like magic, if your data is ordered in a way that can exploit this.

Good to know, thanks for clarifying!

Parquet predicate pushdown filtering is even more powerful than I thought!

For the HN folks outside the USA, CapitalOne is a bank in the USA.

From Wikipedia: Capital One Financial Corporation is an American bank holding company specializing in credit cards, auto loans, banking, and savings accounts, headquartered in McLean, Virginia with operations primarily in the United States. It is on the list of largest banks in the United States and has developed a reputation for being a technology-focused bank.

(italisation is mine)

This is really cool! I'm glad there's more work going into the area - visions[1] is a similar tool which embeds the type system (user defined schema) into a traversable graph rather than encoding all type information in the trained network. We originally wrote it as the backend for pandas-profiling[2].

I'm not sure if any of the authors are on here but y'all might look into Sherlock[3] as well. They've got pre-trained models for many other semantic types than you've currently implemented.

1. https://github.com/dylan-profiler/visions

2. https://github.com/pandas-profiling/pandas-profiling

3. https://sherlock.media.mit.edu/

Great Expectations is another tool that offers data quality and data profiling functionality but for a whole suit of different data sources including csv, jdbc and parquet.

https://github.com/great-expectations/great_expectations

They use a machine learning model to determine categories that have concrete membership rules?

"Note: The Data Profiler comes with a pre-trained deep learning model, used to efficiently identify sensitive data (PII / NPI). If desired, it's easy to add new entities to the existing pre-trained model or insert an entire new pipeline for entity recognition."

No, they seem to use the ML model for identifying data that typically may be considered private, sensitive, or PII, which can be very difficult to define depending on your data and your organization. For example, identifying phone numbers, social security numbers, addresses, etc. All of those things can take a variety of formats and writing a regex or something to identify them feels a bit silly when you can easily train a decent classification model to do this mostly automatically and then catch this kind of data in new use cases where the recognition pattern you may have written before doesn't necessarily work.

Reply on news.ycombinator.com