TileDB 2.0 and the Future of Data Science

https://medium.com/tiledb/tiledb-2-0-and-the-future-of-data-science-929cdcfe95ed

by dalke • 6 years ago

81 33 6 years ago

dalke2• 6 years ago

The highlighted new features for version 2.0 are: 1) Support for heterogeneous dimensions in sparse arrays (that is, full dataframes), 2) Support for string dimensions in sparse arrays, 3) Google Cloud Storage and Azure Blob Storage support, and 4) Completely revamped R API.

They write: "These features mark an important step in our journey to deliver a universal data storage engine for all applications and data science tools."

goodSyntax• 6 years ago

i agree, the title and article is mostly hyperbole

stavrospap6• 6 years ago

Stavros form TileDB here. Here is a more verbose explanation. Up until before 2.0, TileDB was already powerful for the main applications we targeted at: geospatial and genomics. The support for both dense and sparse arrays and the way it handles data versioning made it quite unique vs. HDF5 and Zarr. But we noticed that most of the data scientists we were working with had a lot of data beyond genomic variants, LiDAR points and rasters. They had tons of dataframes. And they were using at least two storage engines, TileDB for arrays, and Parquet or a relational database for dataframes. If you are in a large organization, this a big pain.

In TileDB 2.0 we made a huge refactoring to support something seemingly simple: dimensions in sparse arrays that can have different types and that could even be strings. This allowed us to model any dataframe as a sparse array, effectively making TileDB act as a primary multi-dimensional index. In relational databases, this means that your data is sorted in an order on disk that favors your multi-column slicing enormously, so range search becomes rapid.

Therefore, what we are telling the community with this release is that you can have dense arrays, sparse arrays, and dataframes in a single embeddable library being integrated with pretty much every data science tool out there, so that data scientists never have to worry about backends, files, updates, or anything other than their scientific analysis. In other words, we believe the future of data science is more science.

xyzzy_plugh3• 6 years ago

This seems similar to the direction Apache Arrow is going in, but I don't see the use case for this once Arrow dataframes are prevalent.

tpetry• 6 years ago

One use case is that is running within postgres, so you dont need to install another (complex?) software and backup it.

Shelnutt26• 6 years ago

Disclosure: I am a member of the TileDB team.

TileDB is designed primarily for persistent storage, similar to HDF5, Zarr and Parquet. TileDB is a cloud-optimized, dense and sparse multi-dimensional array storage engine, which is broader than Arrow's Dataframe. In addition TileDB also handles updates, time traveling and partitioning at the library level, which are not possible with the Arrow project's current choice of persistent format, Parquet. TileDB removes the need for using extra services like Delta Lake for updates or Hive for cataloging and partitioning, as you get all this functionality in a single, embeddable library.

missosoup• 6 years ago

I feel like you guys need to create some additional engineering-level documentation to convey all this. The landing page and few links I navigated to were all too fluffy.

xiaodai• 6 years ago

After reading it, it's not clear it is the future of data science.

ipsum2• 6 years ago

I don't know if the creators of TileDB visit HN, but if they do - it would be nice to see a code example that showcases what makes TileDB unique. Right now, I don't see any benefit versus serializing a data frame in hdf5 and storing it in Redis or cloud blob storage.

(Reading https://tiledb.com/developer, the Github, or the blog - none of it shows me anything amazing)

Edit: browsing around the site, I finally found out what tileDB offers, and I'm still confused (too much fluff language) To save people time:

https://docs.tiledb.com/main/handling-dataframes

Multi-column slicing: By defining any subset of the columns as dimensions and due to TileDB's tiling flexibility, you can increase the pruning effectiveness of multi-column slicing, thus leading to better overall read performance. Essentially, a TileDB array acts as a primary multi-dimensional index on the columns selected as the array dimensions.

Data updates and versioning: TileDB offers rapid, parallel, cloud-optimized updates. All the update logic is pushed into the storage engine and is completely transparent to the user. TileDB also exposes useful time traveling functionality, such as reading arrays at time snapshots, effectively implementing data versioning built into a single embeddable library.

Partitioning: TileDB enables balanced partitioning, without limiting each partition to single column values. Moreover, we will soon expose API functions for dynamically selecting different partitioning schemes (e.g., on different subsets of columns with different orders), without the need for reorganizing/rewriting the array.

Sorting: All sorting is taken care of by TileDB internally with multi-threading.

dalke2• 6 years ago

I heard about TileDB because one of the authors commented at https://news.ycombinator.com/item?id=22423632 . That discussion includes what you mentioned, plus a mention that it integrates with MariaDB, PrestoDB and Spark, including the ability to use an embedded MariaDB.

stavrospap6• 6 years ago

Thanks for the post!

stavrospap6• 6 years ago

Stavros from TileDB here. TL;DR, it's all about fast slicing on multiple columns while supporting updates, locally or in the cloud.

Suppose you serialize your dataframe in HDF5 or Redis. Let your dataframe have schema (Date, Stock, Price). Assume this dataframe is 1 TB long and stored on S3, GCS or Azure (as they are cheap). How would you be able to efficiently perform an average query on Price for a specific Date range and Stock symbol? With HDF5 you would have to download 1 TB (no notion of "fast slicing on variable predicates") and apply the predicates locally. If you stored the dataframe in Parquet (a better choice for this use case), then you would be able to build some logic in your code that uses the Parquet metadata/indexes and prune a lot of unnecessary information (as Spark does). However, Parquet is "one-dimensional", i.e., your pruning would be efficient on Date, but not on Stock (you'd have to "partition" your Parquet files with Spark or Hive and things could get quite complicated). Most importantly, you wouldn't be able to update the Parquet files; you would have to generate new files and build a catalog on top (or use services like Delta Lake) to manage your Parquet files. And this is an extremely cumbersome task.

TileDB abstracts everything for you, while allowing you to slice fast on any number of columns. You just define Date and Stock as "dimensions", and slicing on both those columns becomes uber efficient locally or in the cloud. Effectively, you turn this dataframe into a sparse 2D array. Updates and time traveling are handled by TileDB. You get to use Spark, Dask, MariaDB and PrestoDB as you did before, but there is no need for Hive, Delta Lake or any other cataloging service. Thank you for pointing out the confusion though. We just launched and we have tons of examples coming up.

hxzhao• 6 years ago

Stavros thanks for the explanation, how does TileDB avoid downloading the entire matrix and do the slicing (locally)? Are we achieving this by breaking down a big matrix to a set of smaller ones? so that you only down the subset of that the current query need? If this is the case, what is the current measure we have to avoid mismatch on metadata (e.g. some error while uploading them to S3) that links them together? thanks

ihnorton3• 6 years ago

Here is a quick demo of dataframe i/o: reading to/from a pandas dataframe, and then creating and slicing a sparse array with heterogeneous dimensions (int and string):

https://gist.github.com/ihnorton/1544bb169b4ec8cd385e9c3e756...

korzel• 6 years ago

This demo doesn't run properly. Eg in the 2nd code block I got: "AttributeError: module 'tiledb' has no attribute 'from_csv'". I received this message in the latest TileDB Docker image.

willvarfar3• 6 years ago

So, basically, tiledb is a file format like paraquet, avro and orc are formats?

And it comes with a nice library that makes accessing the data fairly storage-transparent?

Shelnutt26• 6 years ago

Disclosure: I am a member of the TileDB team.

Yes you are correct, TileDB introduces a storage format, but is also a full-fledged storage engine that implements it. Having the complete storage engine is what allows us to make accessing the data transparent regardless of the API or application you use to access the data. We are able to integrate into different computational tools without having to reimplement or sacrifice features.

willvarfar3• 6 years ago

How do you compare to Iceberg?

jiangplus• 6 years ago

Looking at the TileDB docs, I like the idea that you can store and manipulate array efficiently. It works like a columnar storage, but organized in tiles. I am not clear how it could scale to multiple nodes and how to do queries. Maybe it should be integrated into postgres.

Shelnutt26• 6 years ago

Disclosure: I am a member of the TileDB team.

TileDB fully supports multi-writer/multi-reader, so scaling compute to multiple nodes is mostly about choosing the computation layer. We have integrations with Spark, Dask and PrestoDB for distributed compute. In our cloud product we also offer serverless UDFs and SQL (via MariaDB) to allow for scaling out computations elastically without managing your own clusters.

On the postgres side, we are looking at eventually adding a postgres storage engine, now that the new storage support is in PostgresSQL 11. To start with we have built a storage engine for MariaDB, as MariaDB has an excellent storage engine API, which I have past experience with. We hope to upstream the storage engine to MariaDB after the MariaDB 10.5 release.

MariaDB integration: https://docs.tiledb.com/mariadb/

Spark integration: https://docs.tiledb.com/spark/

Dask integration: https://docs.tiledb.com/dask/

PrestoDB integration: https://docs.tiledb.com/prestodb/

Cloud Docs: https://docs.tiledb.com/cloud/

TileDB 2.0 and the Future of Data Science

Related Stories