Skip to content

Frequently Asked Questions

Note

This FAQ document covers multiple topic areas - please see the contents table on the right for more information.

Product

What is ArcticDB?

ArcticDB is a high performance DataFrame database built for the modern Python Data Science ecosystem. ArcticDB is an embedded database engine - which means that installing ArcticDB is as simple as installing a Python package. This also means that ArcticDB does not require any server infrastructure to function.

ArcticDB is optimised for numerical datasets spanning millions of rows and columns, enabling you to store and retrieve massive datasets within a Pythonic, DataFrame-like API that researchers, data scientists and software engineers will find immediately familiar.

How does ArcticDB differ from the version of Arctic on GitHub?

Please see the history page.

How does ArcticDB differ from Apache Parquet?

Both ArcticDB and Parquet enable the storage of columnar data without requiring additional infrastructure.

ArcticDB however uses a custom storage format that means it offers the following functionality over Parquet:

  • Versioned modifications ("time travel") - ArcticDB is bitemporal.
  • Timeseries indexes. ArcticDB is a timeseries database and as such is optimised for slicing and dicing timeseries data containing billions of rows.
  • Data discovery - ArcticDB is built for teams. Data is structured into libraries and symbols rather than raw filepaths.
  • Support for streaming data. ArcticDB is a fully functional streaming/tick database, enabling the storage of both batch and streaming data.
  • Support for "dynamic schemas" - ArcticDB supports datasets with changing schemas (column sets) over time.
  • Support for automatic data deduplication.

What sort of data is ArcticDB best suited to?

ArcticDB is an OLA(nalytical)P DBMS, rather than an OLT(ransactional)P DBMS.

In practice, this means that ArcticDB is optimised for large numerical datasets and for queries that operate over many rows at a time.

Does ArcticDB require a server?

No. ArcticDB is a fully fledged embedded analytical database system, designed for modern cloud and on-premises object storage that does not require a server for any of the core features.

What languages can I use ArcticDB with?

Bindings are currently only available for Python.

How can I get started using ArcticDB?

Please see our getting started guide!

Technical

Does ArcticDB use SQL?

No. ArcticDB enables data access and modifications with a Python API that speaks in terms of Pandas DataFrames. See the reference documentation for more details.

Does ArcticDB de-duplicate data?

Yes.

On each write, ArcticDB will check the previous version of the symbol that you are writing (and only this version - other symbols will not be scanned!) and skip the write of identical segments. Please keep in mind however that this is most effective when version n is equal to version n-1 plus additional data at the end - and only at the end! If there is additional data inserted into the in the middle, then all segments occuring after that modification will almost certainly differ. ArcticDB segments data at fixed intervals and data is only de-duplicated if the hashes of the data segments are identical - as a result, a one row offset will prevent effective de-duplication.

Note that this is a library configuration option that is off by default, see help(LibraryOptions) for details of how to enable it.

How does ArcticDB enable advanced analytics?

ArcticDB is primarily focused on filtering and transfering data from storage through to memory - at which point Pandas, NumPy, or other standard analytical packages can be utilised for analytics.

That said, ArcticDB does offer a limited set of analytical functions that are executed inside the C++ storage engine offering significant performance benefits over Pandas. For more information, see the documentation for the QueryBuilder class.

What does Pickling mean?

ArcticDB has two means for storing data:

  1. ArcticDB can store your data using the Arctic On-Disk Storage Format.
  2. ArcticDB can Pickle your data, storing it as a giant binary blob.

(1) is vastly more performant (i.e. reads and writes are faster), space efficient and unlocks data slicing as described in the getting started guide. There are no practical advantages to storing your data as a Pickled binary-blob - other than certain data types must be Pickled for ArcticDB to be able to store them at all!

ArcticDB is only able to store the following data types natively:

  1. Pandas DataFrames
  2. NumPy arrays
  3. Integers (including timestamps - though timezone information in timestamps is removed)
  4. Floats
  5. Bools
  6. Strings (written as part of a DataFrame/NumPy array)

Note that ArcticDB cannot efficiently store custom Python objects, even if inserted into a Pandas DataFrames/NumPy array. Pickled data cannot be index or column-sliced, and neither update nor append primitives will function on pickled data.

How does indexing work in ArcticDB?

See the Getting Started page for details of supported index types.

Can I append with additional columns / What is Dynamic Schema?

You can append (or update) with differing column sets to symbols for which the containing library has Dynamic Schema enabled. See the documentation for the create_library method for more information.

You can also change the type of numerical columns - for example, integers will be promoted to floats on read.

How does ArcticDB segment data?

See On Disk Storage Format and the documentation for the rows_per_segment and columns_per_segment library configuration options for more details.

How does ArcticDB handle streaming data?

ArcticDB has full support for streaming data workflows - including high throughput tick data pipelines. Please contact us at [email protected] for more information.

How does ArcticDB handle concurrent writers?

Without a centralised server, ArcticDB does not support transactions. Instead, ArcticDB supports concurrent writers across symbols - but not to a single symbol (unless "staging the writes"). It is up to the writer to ensure that clients do not concurrently modify a single symbol.

In the case of concurrent writers to a single symbol, the behaviour will be last-writer-wins. Data is not lost per se, but only the version of the last-writer will be accessible through the version chain.

To reiterate, ArcticDB supports concurrent writers to multiple symbols, even within a single library.

Note

ArcticDB does support staging multiple single-symbol concurrent writes. See the documentation for staged.

Does ArcticDB cache any data locally?

Yes, please see the Runtime Configuration page for details.

How can I enable detailed logging?

Please see the Runtime Configuration page for details.

How can I tune the performance of ArcticDB?

Please see the Runtime Configuration page for details.

Does ArcticDB support categorical data?

ArcticDB currently offers extremely limited support for categorical data. Series and DataFrames with categorical columns can be provided to the write and write_batch methods, and will then behave as expected on read. However, append and update are not yet supported with categorical data, and will raise an exception if attempted. The QueryBuilder is also not supported with categorical data, and will either raise an exception, or give incorrect results, depending on the exact operations requested.

How does ArcticDB handle NaN?

The handling of NaN in ArcticDB depends on the type of the column under consideration:

  • For string columns, NaN, as well as Python None, are fully supported.
  • For floating-point numeric columns, NaN is also fully supported.
  • For integer numeric columns NaN is not supported. A column that otherwise contains only integers will be treated as a floating point column if a NaN is encountered by ArcticDB, at which point the usual rules around type promotion for libraries configured with or without dynamic schema all apply as usual.